<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:atom="http://www.w3.org/2005/Atom" xmlns:media="http://search.yahoo.com/mrss/" xmlns:feedpress="https://feed.press/xmlns" xmlns:podcast="https://podcastindex.org/namespace/1.0" version="2.0">
  <channel>
    <feedpress:locale>en</feedpress:locale>
    <atom:link rel="hub" href="https://feedpress.superfeedr.com/"/>
    <atom:link href="https://feedpress.me/wx-girlswhocode" rel="self" type="application/rss+xml"/>
    <atom:icon>https://localhost:3000/favicon.ico</atom:icon>
    <title>程序媛的日常</title>
    <description>公众号“程序媛的日常”更新 - 使用 Huginn 制作</description>
    <link>https://localhost:3000</link>
    <lastBuildDate>Sun, 03 May 2026 20:14:32 -0700</lastBuildDate>
    <pubDate>Sun, 03 May 2026 20:14:32 -0700</pubDate>
    <ttl>60</ttl>
    <item>
      <title>悟道·文澜：北京智源研究院、人大、中科院联手打造的超大规模多模态预训练模型！</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content" style="visibility: hidden;">
                    

                    
                    
                    
                    <section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: left;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;display: inline-block;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span title="" opera-tn-ra-cell="_$.pages:0.layers:0.comps:0.title1" style="padding: 0.3em 0.5em;max-width: 100%;display: block;border-radius: 0.8em 0.8em 0px 0px;background-color: rgb(174, 11, 42);font-size: 20px;color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style='font-size: 18px;font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;'>导读</strong><br></p></span></section><section style="padding: 10px;max-width: 100%;border-width: 1px;border-style: solid;border-color: rgb(174, 11, 42);border-radius: 0px 0px 0.8em 0.8em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;text-align: justify;font-family: PingFangSC-light;letter-spacing: 2px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);font-size: 15px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>近年来，BERT等预训练语言模型在多类自然语言处理任务上取得了显著的性能提升，从而极大地改写了研究范式。随着OpenAI超大规模语言模型GPT-3的发布，预训练语言模型在自然语言理解能力上再次被推至新的高峰。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);font-size: 15px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>AI江湖里早已风起云涌，年初发布的CLIP和DALL·E模型，即多模态版本的GPT-3，更是展现出了惊人的语言理解和图像生成能力，刷爆了研究人员们的朋友圈。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 15px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>处于AI浪潮中的你，一次次地见证着这些划时代的工作诞生。你是否期待着，</span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">当视觉和语言碰撞时，预训练模型还会为我们带来哪些新的惊喜？</strong></span></span></p></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;font-size: 20px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;letter-spacing: 1px;font-size: 18px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">赶快来围观这个最新发布的多模态预训练模型吧！</strong></span></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>为了更进一步地推动相关领域的研究，</span><span style="max-width: 100%;color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>北京智</span></strong><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>源人工智</span></strong><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>能研究院、中国人民大学</span></strong></span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>和</span><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'>中科院计算所</span></strong><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>的研究团队在</span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">中国人民大学高瓴人工智能学院执行院长文继荣教授</strong></span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>的带领下合作开展了大规模中文多模态预训练模型的研究，并</span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">发布了第一代悟道·文澜</strong></span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>，旨在发掘预训练模型在中文通用多模态数据上的理解能力。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>现阶段的“文澜”已初具规模，具备强大的视觉-语言检索能力和一定的常识理解能力。在“文澜”多模态模型的基础上，</span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">团队还开发了应用《AI心情电台》</strong></span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>，可以为图像搭配符合意境的歌曲。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.3851276359600444" data-type="png" data-w="1802" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiawtCvcGdo7hefH6EZYfwPyTBicIWjQuON8kylHqMk5ULkJ0icxwN9BJoQ/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;overflow-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><blockquote data-type="2" data-url="" data-author-name="" data-content-utf8-length="32" data-source-title="" style='color: rgba(0, 0, 0, 0.498);max-width: 100%;letter-spacing: 0.544px;white-space: normal;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'><section style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p line="VnTF" style="max-width: 100%;min-height: 1em;line-height: 2em;letter-spacing: 0.544px;text-align: left;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);box-sizing: border-box !important;overflow-wrap: break-word !important;">论文地址：</span></p>
<p line="VnTF" style="max-width: 100%;min-height: 1em;line-height: 2em;letter-spacing: 0.544px;text-align: left;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);box-sizing: border-box !important;overflow-wrap: break-word !important;"></span><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 0.544px;text-decoration: underline;box-sizing: border-box !important;overflow-wrap: break-word !important;">https://arxiv.org/abs/2103.06561</span></p>
<p line="VnTF" style="max-width: 100%;min-height: 1em;line-height: 2em;letter-spacing: 0.544px;text-align: left;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);box-sizing: border-box !important;overflow-wrap: break-word !important;">多模态首页：</span></p>
<p style="max-width: 100%;min-height: 1em;line-height: 2em;letter-spacing: 0.544px;text-align: left;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-size: 15px;letter-spacing: 0.544px;text-decoration: underline;box-sizing: border-box !important;overflow-wrap: break-word !important;">https://model.baai.ac.cn/model/wl</span></p></section></section></blockquote>
<section data-mpa-template="t" mpa-from-tpl="t" style='max-width: 100%;font-family: -apple-system, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.544px;white-space: normal;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></section></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;justify-content: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="padding-right: 10px;padding-left: 10px;max-width: 100%;display: inline-block;width: auto;vertical-align: top;background-color: rgb(245, 246, 250);border-top: 4px solid rgb(174, 11, 42);border-top-left-radius: 0px;min-width: 10%;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;color: rgb(174, 11, 42);font-size: 20px;letter-spacing: 0.7px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">模型特点及优势</strong></p></section></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>多模态预训练模型通常可以归类为以下两种网络结构：如UNITER的单塔结构，和如CBT、VilBERT和ERNIE-VIL的双塔结构。</span></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>从2020年9月开始，通过一系列的实验和探索，文澜的研发团队独立提出了基于多模态对比学习的双塔结构，跟2021年1月发布的OpenAI CLIP在模型结构上不谋而合。</span></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>文澜团队推出的第一代图文互检模型在论文中叫做</span><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'>BriVL (Bridging Vision and Language)</span></strong><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>，BriVL具有如下四个特点和优势。</span></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">1. 基于视觉-语言弱相关的假设</strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>现有的大量多模态预训练模型，特别是单塔结构，往往会采用如下强假设：<strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">对于输入的数据，图像与文本之间存在较强的语义相关性。</strong>例如，对于下面这张蛋糕的图片，会假设在多模态数据集中对应“水果蛋糕上有一些蜡烛在燃烧”这样描述性的文本。</span></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.6806451612903226" data-type="png" data-w="620" data-backw="578" data-backh="393" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiaGAvH0gRiadnNKDcqCic4dibUDphsAWRAyDZzVA5aCUWiaNEqhWtqfcHwcA/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 620px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="margin-top: 5px;max-width: 100%;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></section><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>正是有了语义相关性上的强假设，单塔结构才能在词汇与局部图像特征之间进行模态交互。但遗憾的是，<strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">在实际应用场景中，上述的强假设往往并不成立，比如视觉与语言之间通常只有较抽象的关联。</strong>例如，对于蛋糕的照片，配的文字可以是“生日快乐，许个愿吧”，也可以引申到“哎，我的减肥大计又泡汤了”</span>。</p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>文澜的研发者们进行了一系列的实验和探索，实验结果表明，在开放获取（例如互联网上的公开数据）的图文数据集上，简单的双塔结构要优于单塔结构。因此，<strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">BriVL采用了双塔结构作为多模态预训练模型的基本架构。</strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">2. 将多模态与对比学习算法相结合</strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>基于双塔结构的预训练模型网络结构比较简单，需要弥补神经网络在表达能力上的损失。近两年来，自监督学习有着飞速的进展，对比学习被发现可以用来提高神经网络的表达能力。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>基于这一点，文澜的开发团队将对比学习引入到了BriVL的双塔结构中。不同于CLIP，对于给定的某一对图文数据，BriVL同时使用了视觉模态和语言模态去构建该图文数据的负样本，并且基于MoCo的思想扩大负样本数目，从而进一步提高神经网络的表达能力。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">3. 网络结构灵活，方便实际部署</strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>BriVL首先使用了独立的语言和视觉编码器提取语言和视觉信息的特征向量，然后将这些向量传入到对比学习模块中进行训练。采用这样的双塔结构，可以很方便地把编码器模块替换为最新的单模态预训练模型，从而可以持续地增强模型表达的能力。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>在训练完毕后，<strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">BriVL可以对外提供获取图像和语言的向量特征的API，便于在下游任务中部署。</strong>特别地，与向量检索引擎的结合可以极大地提高图文检索效率。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>此外，BriVL模型也可以再融入其他的预训练任务，例如图像caption任务等。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">4. 目前最大的中文多模态通用预训练模型</strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>文澜团队使用</span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">RUC-CAS-WenLan</strong></span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>多源图文数据集对BriVL进行预训练。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>RUC-CAS-WenLan是文澜团队构建的超大规模预训练数据集，该多源图文数据集来自网页用户产生的图文数据。文澜团队的数据组从互联网获取原始数据后，对数据进行了一系列的清洗工作，从而构建了包含5500万对图文数据的RUC-CAS-WenLan通用多模态数据集。RUC-CAS-WenLan的内容十分丰富，涵盖了<strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">新闻，体育，娱乐，文化，生活</strong>等多个主题。</span></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.5899550224887556" data-type="png" data-w="2668" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiattquFXdAria4poIUu6CDibthD14Mn4o8vdzQicBfK5vwRQicWHczoYQR6Q/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;overflow-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>第一代文澜使用了3000万对图文数据，训练得到的BriVL模型的初始版本有10亿参数量，此阶段的BriVL模型在RUC-CAS-WenLan测试集和AIC-ICC测试集上的性能均超过了UNITER和CLIP。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>在未来，<strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">文澜团队将使用5亿级别的图文数据作为预训练数据集，BriVL的参数量将达到百亿级别。</strong></span></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;justify-content: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="padding-right: 10px;padding-left: 10px;max-width: 100%;display: inline-block;width: auto;vertical-align: top;background-color: rgb(245, 246, 250);border-top: 4px solid rgb(174, 11, 42);border-top-left-radius: 0px;min-width: 10%;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;color: rgb(174, 11, 42);font-size: 18px;letter-spacing: 0.7px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 20px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">模型基本架构</strong></span></p></section></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>BriVL模型是基于图文互检预训练任务设计的，采用了如下图所示的基础架构。</span></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.6337792642140468" data-type="png" data-w="1196" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aia2SUicwYlBlUuFBuTWoUJYnellgXWYVA0iamUibLMBdoxmNrPdicQibyFic6Q/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;overflow-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>图文互检模型BriVL结构示意图</span></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>针对图文互检预训练任务，BriVL模型采用了双塔结构。具体地，先使用图像目标检测器去提取图像数据的Bounding Box，再使用这些Bounding Box对经过图像编码器的特征图进行空间上的池化，得到图像特征序列。文本数据也会经过文本编码器进行特征提取。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>在提取特征后，两个模态的序列特征被分别映射到了不同的语义空间，然后对这两个序列分别进行序列池化和多层感知机映射，从而得到它们在同一空间下的特征表示。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>图文互检预训练任务使用了基于MoCo的多模态对比学习方法，优化处于同一特征空间下的双模态特征，使得对应的图文对在该特征空间下有一致的语义表示。在图像编码器部分，BriVL使用了EfficientNet-B7-NS模型和多层单模态的Transformer模型；在文本编码器部分，BriVL使用了RoBERTa-Large和多层单模态的Transformer模型。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>基于上述图文编码器，文澜开发团队使用16台8卡*A100的GPU集群进行了7天的预训练，得到了最终的BriVL模型。在下一轮迭代中，文澜开发团队还会探索如何构建性能更强的文本编码器，以进一步提高其表达能力。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;justify-content: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="padding-right: 10px;padding-left: 10px;max-width: 100%;display: inline-block;width: auto;vertical-align: top;background-color: rgb(245, 246, 250);border-top: 4px solid rgb(174, 11, 42);border-top-left-radius: 0px;min-width: 10%;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;color: rgb(174, 11, 42);font-size: 20px;letter-spacing: 0.7px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">实验分析</strong></p></section></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">1. AIC-ICC的评测结果</strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>AIC-ICC是2017年由创新工场、搜狗、今日头条联合主办的首届AI Challenger全球AI挑战赛中发布的中文图像描述数据集。文澜团队以AIC-ICC公开数据集为基础，进行了一系列的消融和对比实验，并取得两项突破性的结果：</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>（1）在图文互检的任务上，BriVL的性能优于UNITER和CLIP：图检索文与文检索图的结果均明显超过UNITER和CLIP，且检索速度是UNITER的20倍，与CLIP速度相当。在与OPENAI-CLIP的比较过程中，由于CLIP未放出预训练数据，只给了测试模型与接口，文澜研发团队便采取将测试集用主流翻译软件翻译成英文再调用其接口得到特征向量的做法，以最大程度公平比较。</span></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.3287037037037037" data-type="png" data-w="864" data-backw="578" data-backh="190" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiakJsuZss722QmFteSHIrsMa0fW38gpib853YRNZlhjR3T6nfic5J00oow/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 677px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>（2）在图生成文任务上，以四个指标的综合评价，文澜BriVL在公开评测集AIC-ICC上超过了现有最高结果AI Challenger全球AI挑战赛的冠军，同时也超过了UNITER。</span><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.3062645011600928" data-type="png" data-w="862" data-backw="578" data-backh="177" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiaFlTDjoteLgl7prInY8d9N2t7LN2FewSWTy9icBMmM1eyRFAG7VWIciaw/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 677px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;box-sizing: border-box !important;overflow-wrap: break-word !important;">2. RUC-CAS-WenLan的评测结果</span></strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>文澜团队最初就从RUC-CAS-WenLan中分离出了1.1万个图文对作为测试集，其余的用作训练。在测试集的图文互检任务上，文澜BriVL的性能均超过了UNITER和CLIP。</span></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.3325526932084309" data-type="png" data-w="854" data-backw="578" data-backh="192" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiadVM9TcvYV7I9GBfw5QwgRCYkRTagjic9JwFDjRjDEXOhoK7UvhzRIHw/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 677px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">3. 非公开评测结果</strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>针对图文互检任务，文澜团队还进行了非公开评测。给定一组文本和图像查询样例，为每个查询样例从候选集中挑选最好的30个结果，然后进行人工评分。选用三个人工进行标注，从而得到三组结果，每个查询样例有7个级别的打分，每组有30个打分。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>通过计算NDCG和MAP，可以对该查询样例的30个打分进行评测，最终得到所有查询样例的平均结果。通过表4可以看出，文澜BriVL要优于CLIP。此外，若使用BriVL为UNITER选择候选集，UNITER能够在BriVL的基础上继续提高性能。</span></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 677px;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.5350467289719626" data-type="png" data-w="856" data-backw="578" data-backh="309" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiaRWFicmNHKiazbojfrHxbSLcM2n2TJFzdhCickImliaF8S9fN6lXdhvtcAQ/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 677px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;justify-content: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="padding-right: 10px;padding-left: 10px;max-width: 100%;display: inline-block;width: auto;vertical-align: top;background-color: rgb(245, 246, 250);border-top: 4px solid rgb(174, 11, 42);border-top-left-radius: 0px;min-width: 10%;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;color: rgb(174, 11, 42);font-size: 20px;letter-spacing: 0.7px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">结果展示</strong></p></section></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;">1. 文本检索图像</span></strong></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>用“航拍南京玄武湖秋色美若画卷”进行检索，返回结果如下：</span></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.7067961165048544" data-type="jpeg" data-w="2060" src="https://mmbiz.qpic.cn/mmbiz_jpg/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aia67icEcqv8Wb3z0cq0iaEYROI3AEZqBGxX1E28kDfVpNNdxHEjcR0mhRA/640?wx_fmt=jpeg" style="box-sizing: border-box;vertical-align: middle;overflow-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;">query: 航拍南京玄武湖秋色美若画卷</p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(174, 11, 42);font-size: 18px;box-sizing: border-box !important;overflow-wrap: break-word !important;">2. 图像检索文本</span></strong></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 473.891px;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="1.3292929292929292" data-type="png" data-w="495" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiaibcZ9xZBAM7YibuVoaOVfvmLHSzz8BSib7UDnpL2gmeQeZkEfpXWbia9eA/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 473.891px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>用上面这张图片做query，返回的结果如下：</span></p>
<ol class="list-paddingleft-2" style="max-width: 100%;overflow-wrap: break-word !important;">
<li style="max-width: 100%;font-style: italic;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;">梨花开年复年，相思别处生又生；今日有酒今宵醉，何至年头憎恨春</span></em></p></li>
<li style="max-width: 100%;font-style: italic;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;">像冬季清晨里的一抹阳光像夏季傍晚的凉风，像秋季午后凉爽的汽水，像春季凌晨犹如浩瀚宇宙一般的天空，你是我的心小鹿乱撞的原因。</span></em></p></li>
<li style="max-width: 100%;font-style: italic;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;">嘿嘿！小编的品味还不错吧。接下来是真人女头像。小仙女们请拿去食用！</span></em></p></li>
<li style="max-width: 100%;font-style: italic;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;">殿下一族吧是由小吧主天魔舞梦创建的。主要由天魔舞梦、赤黑烁、幻冥黑白、赤黑之友发言。</span></em></p></li>
<li style="max-width: 100%;font-style: italic;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;">《石楠小札》，收录于古风歌手 CcccEs 个人原创音乐专辑《不贰》，由小鱼萝莉填词，TetraCalyx 编曲，CcccEs 演唱。其发行时间为 2014 年 08 月 21 日。</span></em><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;"></span></p></li>
</ol>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;color: rgb(174, 11, 42);font-size: 18px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">3. 下游任务：基于图像的文本生成</strong></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>BriVL在AIC-ICC+MSCoCo+Flickr30数据集上，以图片生成标题为任务，进行微调。生成的结果如下：</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 609.297px;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.5497912939773405" data-type="png" data-w="1677" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiab9MoB0dPqic2OUDQSp6p3Iktxau5f9LicRKQtniczdia54uicKpjs65Fd6A/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 609.297px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">生成文本：一个穿着戏服的男人和一个穿着戏服的女孩在一起</em></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 609.297px;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.666" data-type="jpeg" data-w="500" src="https://mmbiz.qpic.cn/mmbiz_jpg/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiaSVT63qUI3kdStod2ss4kW8h3yflRAbX5aA5ia08Q4ia9TZo8OBnrjt3w/640?wx_fmt=jpeg" style="box-sizing: border-box;vertical-align: middle;width: 609.297px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">生成文本：一条城市街道，有一个红绿灯，上面有一个标志</em></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;font-size: 18px;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">4. 下游任务：为图像打标签</strong></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>用BriVL模型在动漫数据集danboru上进行微调，为动漫图像自动打标签，得到的结果如下所示：</span></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 609.297px;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.7303183" data-type="png" data-w="597" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiatiaW11AibTysFjPTdjlFkrdFHHibLhASlwcbcmcOG2eUQ95V3HERObM2g/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 609.297px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"> 灰色围巾，屋顶，双手插在口袋里，多云的天空，大衣</em></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">鸟，围巾，云，天空，长袖，微笑，长发</em></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 609.297px;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="0.7755444" data-type="png" data-w="597" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiabhibibJzXBGrXmfAn3p14txMQJZOibdsDnssUCXO2JuJZUNWpjdt3rwxw/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 609.297px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;">木栅栏，城镇，街道，房屋，道路，日落，建筑</span></em></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;">风景，窗口，云，天空</span></em><span style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;"></span></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;justify-content: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="padding-right: 10px;padding-left: 10px;max-width: 100%;display: inline-block;width: auto;vertical-align: top;background-color: rgb(245, 246, 250);border-top: 4px solid rgb(174, 11, 42);border-top-left-radius: 0px;min-width: 10%;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;color: rgb(174, 11, 42);font-size: 20px;letter-spacing: 0.7px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">下游应用</strong></p></section></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>基于BriVL模型，可以开发多款跨模态应用。在BriVL的基础上，文澜团队开发了H5小应用《AI心情电台》。</span></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;box-sizing: border-box !important;overflow-wrap: break-word !important;">布灵—</span></strong></span><span style="max-width: 100%;color: rgb(174, 11, 42);box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 18px;box-sizing: border-box !important;overflow-wrap: break-word !important;">—为图片赋予音乐的灵魂</span></strong></span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>用户随便上传一张图片后，便会有一个可爱的小精灵——</span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">布灵</strong></span><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>，为用户配上一首符合意境的歌。《AI心情电台》是使用BriVL提取图像和文本特征，接着进行图文检索，将图片和歌词特征进行匹配，并将歌词准确定位到最符合图片特征的歌词位置。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>《AI心情电台》期待做用户的灵魂DJ。</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 541.594px;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="1.958937198067633" data-type="jpeg" data-w="828" src="https://mmbiz.qpic.cn/mmbiz_jpg/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiaIAsVoCBzAic0FQ0UGSkTC4kMEK6qO2OpwibOEFbSiam5LLiaPo4WxZwq6w/640?wx_fmt=jpeg" style="box-sizing: border-box;vertical-align: middle;width: 541.594px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">你是否怀疑过自己越想越不太可能<br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></em></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">现在开始你会相信带着梦去旅行</em></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">让我看看你的瞳孔正在排练的电影</em></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">我听见你的心房跳动越靠近越大声</em></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">《you make me free》——张惠妹</em></p>
<p style="max-width: 100%;min-height: 1em;text-align: center;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 541.594px;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="1.9577294685990339" data-type="jpeg" data-w="828" src="https://mmbiz.qpic.cn/mmbiz_jpg/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiaD0UNuNBq3H0RjicjOI98UCTjt8coljBTll1w933bYGJ9Z786z15KHtw/640?wx_fmt=jpeg" style="box-sizing: border-box;vertical-align: middle;width: 541.594px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;text-align: center;font-size: 14px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">整理旧照片泛黄的笑脸<br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></em></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">记忆不断涌现湿热我的脸</em></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">铺木的街道无忧的年少<br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></em></p>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><em style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">《小木马》——张信哲</em></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>随着动听的旋律，感受灵魂的对白，用图与音乐，交织梦与理想的绘卷。</span></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box !important;overflow-wrap: break-word !important;"><img data-ratio="1" data-type="png" data-w="529" data-backw="529" data-backh="529" src="https://mmbiz.qpic.cn/mmbiz_png/G7ia3FZ0o0Oq7CLvq9GKtsBib2jUtyn0aiabRxNlEcvmfLQ13gQYc7nd3vY1ibb9ZvYIcBNW72Jvk7Q6LdNwphzgwA/640?wx_fmt=png" style="box-sizing: border-box;vertical-align: middle;width: 529px;overflow-wrap: break-word !important;visibility: visible !important;"></section></section><section powered-by="xiumi.us" style="max-width: 100%;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>据悉，布灵是诞生于中国人民大学高瓴人工智能学院的第一个AI，擅长视觉理解，是个在努力弄懂生活和内涵的AI小精灵。</span><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);color: rgb(155, 0, 32);box-sizing: border-box !important;overflow-wrap: break-word !important;'>赶快来让小布灵为你选首歌曲吧！</span></strong></p></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;justify-content: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="padding-right: 10px;padding-left: 10px;max-width: 100%;display: inline-block;width: auto;vertical-align: top;background-color: rgb(245, 246, 250);border-top: 4px solid rgb(174, 11, 42);border-top-left-radius: 0px;min-width: 10%;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;color: rgb(174, 11, 42);font-size: 20px;letter-spacing: 0.7px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">文澜API文档</strong></p></section></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>文澜团队对外开放了获取图像/文本embedding的API接口，供开发者调用。可以在以下地址获取API的调用文档：</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);text-decoration: underline;box-sizing: border-box !important;overflow-wrap: break-word !important;'>https://model.baai.ac.cn/api</span></p>
<p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;justify-content: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="padding-right: 10px;padding-left: 10px;max-width: 100%;display: inline-block;width: auto;vertical-align: top;background-color: rgb(245, 246, 250);border-top: 4px solid rgb(174, 11, 42);border-top-left-radius: 0px;min-width: 10%;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;color: rgb(174, 11, 42);font-size: 20px;letter-spacing: 0.7px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">文澜研发团队介绍</strong></p></section></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="max-width: 100%;white-space: normal;font-size: 16px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>文澜的研究团队由中国人民大学高瓴人工智能学院院长文继荣教授带领，主要人员包括：</span></p>
<section style="max-width: 100%;line-height: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br></section><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>指导教师：</span></strong><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'></span></p>
<ul class="list-paddingleft-2" style="max-width: 100%;overflow-wrap: break-word !important;">
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">文</span><span style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">继荣 中国人民大学高瓴人工智能学院执行院长</span></span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">卢</span><span style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;">志武</span><span style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"> 中国人民大学高瓴人工智能学院教授，博士生导师</span></span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>宋睿华 中国人民大学高瓴人工智能学院长聘副教授</span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>窦志成 中国人民大学高瓴人工智能学院副院长，北京智源人工智能研究院“智能信息检索与挖掘”方向项目经理</span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>金   琴 中国人民大学信息学院教授，博士生导师</span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>兰艳艳 现清华大学智能产业研究院研究员，曾中国科学院计算技术研究所研究员，博士生导师，北京智源人工智能研究院青年科学家</span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>赵   鑫 中国人民大学高瓴人工智能学院长聘副教授，博士生导师，北京智源人工智能研究院青年科学家</span></p></li>
</ul>
<p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>参与学生：</span></strong></p>
<ul class="list-paddingleft-2" style="max-width: 100%;overflow-wrap: break-word !important;">
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>数据组：刘沛羽、郑伟昊、侯丹阳、李英彦、杨越千、隰宗正、龚政、李军毅、赵一达、张良、金楚浩、高一钊、杨国兴</span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>模型组：霍宇琦、胡安文、洪鑫、张曼黎、卢浩宇、高一钊、杨国兴、温静远、崔莞清、赵金明、张良、李瑞晨、赵一达、张恒、许宝贵</span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>评测组：孙宇冲、宋宇晴、崔莞清、霍宇琦</span></p></li>
<li style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 2px;text-align: left;background-color: rgb(255, 255, 255);box-sizing: border-box !important;overflow-wrap: break-word !important;'>应用组：刘光镇、郑伟昊、隰宗正、金楚浩</span></p></li>
</ul></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;white-space: normal;box-sizing: border-box !important;overflow-wrap: break-word !important;"><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section><section powered-by="xiumi.us" style="margin-top: 10px;margin-bottom: 10px;max-width: 100%;text-align: center;justify-content: center;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="padding-right: 10px;padding-left: 10px;max-width: 100%;display: inline-block;width: auto;vertical-align: top;background-color: rgb(245, 246, 250);border-top: 4px solid rgb(174, 11, 42);border-top-left-radius: 0px;min-width: 10%;height: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><section style="max-width: 100%;color: rgb(174, 11, 42);font-size: 18px;letter-spacing: 0.7px;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style="max-width: 100%;min-height: 1em;box-sizing: border-box !important;overflow-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><span style="max-width: 100%;font-size: 20px;box-sizing: border-box !important;overflow-wrap: break-word !important;">部分参考文献</span></strong><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p></section></section></section></section><section powered-by="xiumi.us" style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"><p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[1] Wu, Z., Xiong, Y., Yu, S., &amp; Lin, D. (2018). Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 3733–3742).<br style="max-width: 100%;box-sizing: border-box !important;overflow-wrap: break-word !important;"></span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[2] He, K., Fan, H., Wu, Y., Xie, S., &amp; Girshick, R. (2020). Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 9729–9738).</span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[3] Chen, X., Fan, H., Girshick, R., &amp; He, K. (2020). Improved baselines with momentum contrastive learningarXiv preprint arXiv:2003.04297.</span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[4] Morgado, P., Vasconcelos, N., &amp; Misra, I. (2020). Audio-visual instance discrimination with cross-modal agreementarXiv preprint arXiv:2004.12943.</span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[5] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., &amp; others (2020). Language models are few-shot learnersarXiv preprint arXiv:2005.14165.</span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[6] Chen, T., Kornblith, S., Norouzi, M., &amp; Hinton, G. (2020). A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning (pp. 10709–10719).</span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[7] Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., &amp; Liu, J. (2020). Uniter: Universal image-text representation learning. In European Conference on Computer Vision (pp. 104–120).</span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[8] Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., &amp; others (2020). Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision (pp. 121–137).</span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[9] Alayrac, J.B., Recasens, A., Schneider, R., Arandjelovi\'c, R., Ramapuram, J., De Fauw, J., Smaira, L., Dieleman, S., &amp; Zisserman, A. (2020). Self-supervised multimodal versatile networks. In Advances in Neural Information Processing Systems.</span></p>
<p style='max-width: 100%;min-height: 1em;caret-color: rgb(51, 51, 51);font-family: -apple-system, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 16px;letter-spacing: 0.544px;white-space: normal;text-size-adjust: auto;box-sizing: border-box !important;overflow-wrap: break-word !important;'><span style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: left;background-color: rgb(255, 255, 255);letter-spacing: 0.5px;font-size: 13px;box-sizing: border-box !important;overflow-wrap: break-word !important;'>[10] Radford, A., Kim, J., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., &amp; others. Learning Transferable Visual Models From Natural Language SupervisionImage, 2, T2.</span></p>
<p><br></p>
<p><br></p></section><p><br></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693390&amp;idx=2&amp;sn=32ef7668f611cb55de60bc791ad6f84e</link>
      <pubDate>Wed, 14 Apr 2021 02:30:17 -0700</pubDate>
      <guid isPermaLink="false">351762</guid>
    </item>
    <item>
      <title>专访 | 悟道文澜团队：如何让 AI 像孩子一样用多模态的方式认识世界”</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content" style="visibility: hidden;">
                    

                    
                    
                    
                    <p style="text-align: center;"><img class="rich_pages" data-galleryid="" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740" style=""></p>
<p style="text-align: left;"><span style='text-align: left;font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;'>前</span><span style='text-align: left;font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;'>段时间，北京智源人工智能研究院、中国人民大学和中科院计算所的研究团队在中国人民大学高瓴人工智能学院执行院长文继荣教授的带领下合作开展了大规模中文多模态预训练模型的研究，并发布了第一代悟道·文澜，旨在发掘预训练模型在中文通用多模态数据上的理解能力。</span><br></p>
<p style="font-size: 15px;white-space: normal;text-align: left;"><br></p>
<p style="font-size: 15px;white-space: normal;text-align: left;">现阶段的“文澜”已初具规模，具备强大的视觉-语言检索能力和一定的常识理解能力。在“文澜”多模态模型的基础上，团队还开发了应用《AI 心情电台》，可以为图像搭配符合意境的歌曲。</p>
<p style="font-size: 15px;white-space: normal;text-align: left;"><br></p>
<p style="font-size: 15px;white-space: normal;text-align: left;">最近，小S有幸近距离采访了文澜团队的卢志武老师和宋睿华老师，一起来听听他们对于多模态和预训练的看法吧！采访分为5个部分：</p>
<ol class="list-paddingleft-2" style="list-style-type: decimal;">
<li><section style="font-size: 15px;white-space: normal;text-align: left;text-indent: 0em;">文澜模型：<em>问题 0-5</em><br></section></li>
<li><section style="font-size: 15px;white-space: normal;text-align: left;text-indent: 0em;">多模态研究：<em>问题 6-7</em></section></li>
<li><section style="font-size: 15px;white-space: normal;text-align: left;text-indent: 0em;">“布灵”心情电台：<em>问题 8-11</em></section></li>
<li><section style="font-size: 15px;white-space: normal;text-align: left;text-indent: 0em;">超大规模预训练模型：<em>问题 12-14</em></section></li>
<li>
<section style="font-size: 15px;white-space: normal;text-align: left;text-indent: 0em;">开放生态及合作：<em>问题 15-18</em></section><p style="font-size: 15px;white-space: normal;text-align: left;"><br></p>
</li>
</ol>
<p style="white-space: normal;text-align: center;"><span style="color: rgb(178, 178, 178);font-size: 14px;">全文共计：14309字 </span></p>
<p style="white-space: normal;text-align: center;"><span style="color: rgb(178, 178, 178);font-size: 14px;">预计阅读时间：18分钟</span></p>
<p style="text-align: left;"><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;font-size: 16px;"><section powered-by="xiumi.us" style="margin-top: 15px;margin-bottom: 15px;box-sizing: border-box;"><section style="padding-right: 10px;padding-left: 10px;font-size: 15px;line-height: 2;box-sizing: border-box;"><section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">0. “悟道”和“文澜”等名字的起源</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><span style="font-size: 15px;"><strong style="box-sizing: border-box;"><span style="box-sizing: border-box;">宋老师</span></strong><strong style="box-sizing: border-box;"><span style="box-sizing: border-box;">：</span></strong></span>在比较早的时候，清华的刘知远老师和孙茂松老师的团队就已经发布过一个版本的中文预训练模型，命名为“文源”。孙茂松老师讲<strong><span style="text-decoration: none;">“文源”的名字来自圆明园里的文源阁</span></strong>，起“文源”这个名字也希望能让机器去大量地阅读人类留下来的文字，学会其中的规律。因此，刘知远老师坚持要将中文预训练模型叫“文源”。而唐杰老师希望四个模型有一定的关联，像兄弟姐妹的名字。所以后来我就问刘知远老师，圆明园里还有哪些地名，以“文”字开头的，发现还真的有不少而且很美，因为蕴含了咱们古人对文字的掌握还有一些趣味在里面。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">特别巧的是，我们悟道的四个模型刚好各有特色，比如说我们文澜的模型跟图像有关的，<strong>一听到“文澜”就觉得眼前会有波澜，会有浪漫的感觉，是有视觉的这种意向在里面</strong>。所以我很快就挑了这个名字给多模态项目。随后，我们也很快就找到了和其他两个预训练模型相关的名字：“文汇”就像从认知角度把知识都汇集在一起了，而“文溯”很适合蛋白质组，跟人的起源有关。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">定完这四个模型的名字后，就要考虑整个大的项目名字叫什么。有人提到“悟空”，但“悟空”这个名字已经有人用了。<strong>后来也是刘知远老师说到北京智源研究院位于五道口，鉴海防老师就提议说我们不如就叫“悟道”</strong>。唐杰老师也非常赞成，既有那种从数据中体会感悟出规律的感觉，又跟“文”字系列的名字很相关。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p></section></section></section><section style="box-sizing: border-box;font-size: 16px;"><section powered-by="xiumi.us" style="margin-top: 15px;margin-bottom: 15px;box-sizing: border-box;"><section style="padding-right: 10px;padding-left: 10px;font-size: 15px;line-height: 2;box-sizing: border-box;"><section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">1. 多模态数据和多模态模型的优势</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">卢老师</strong>：我们总共有两个世界，一个是我们脑海里的世界，还有一个是脑海外的客观存在的世界。我们认知和描述两个世界的时候，一个是通过语言，一个就是我们的视觉。其实就是我们要建立和两个世界的一个联系。如果从最简单的做起，就是去从视觉去认知现实世界，也就是图像处理。对应精神世界来看，其实我们是用语言去描述精神世界。<strong>所以说多模态数据是一个自然的联系两个世界的方式。</strong>当然后面我们可以去利用音频视频，这都是很自然的事情，接下来也会有顺延合作。<br></p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">宋老师</strong>：以前深度学习出现之前，大家建模问题时，都是用浅层学习把一个任务划分成很多阶段。举例来说，图像识别就是先识别出一张图像中的物体，然后再描述出来，之后才能用一些复杂的模型去推理。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">其实到今天为止，我们可以看到有一个变化。就好像在教育孩子的时候一样，你有时候并不需要把那个东西肢解开了给孩子讲，而是让他看到一个目标是什么。孩子会根据大人的做法，去模仿大人。其实孩子也不知道他自己头脑里发生了什么，但是他就可以认识（认知），他就可以做（模仿，实现，完成）。这样反而要更有效率。<strong>所以当我们把机器和小孩的学习方式进行类比后，我们也希望让数据本身去说话（去发挥价值），而不是一步步去教机器。</strong></p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">我们可以用网上大量存在的数据，比如网上的图片很多配有文字，它们中蕴含着很高层次的推理，有很高层次的情感触发，都以文字的形式放在了图片的周围（配文）。我们就可以利用这样的数据，让我们的文澜模型（包括布灵应用）直接地去学习这样的关系，这就是多模态模型的一个非常大的优势。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">卢老师</strong>：而且我认为我们的这种做法更符合我们中国的传统文化，它更强调从宏观上去做一些事情。当然无论是宏观角度还是微观角度，都各有优缺点。但语言和精神世界太复杂了，想要去完全拆解它，比如去构建知识图谱，再按照各种离散概念组合它，是很繁杂的工作。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>宋老师</strong>：我也同意，虽然知识图谱可能是我们人类在传递一些信息的时候找到的一个比较有效的方式，但不一定是我们教给机器知识的最有效的方式。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p></section></section></section><section style="box-sizing: border-box;font-size: 16px;"><section powered-by="xiumi.us" style="margin-top: 15px;margin-bottom: 15px;box-sizing: border-box;"><section style="padding-right: 10px;padding-left: 10px;font-size: 15px;line-height: 2;box-sizing: border-box;"><section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">2. 文澜模型的弱相关假设</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">卢老师</strong>：<strong>弱相关假设是我们整个项目成功的一个最核心的地方。</strong>虽然 OpenAI 也有做类似的工作 CLIP，还有微软也有做多模态预训练，但他们的论文中并没有意识到弱相关的这个问题。也就是说，他们的模型方法中无意识地使用了强相关假设，但这个假设在现实生活中的真实数据（网络数据）上往往是不成立的。总之，通过强假设的方法去学弱相关的数据，肯定有损模型效果。<br></p>
<p style="text-align: center;"><img class="rich_pages" data-galleryid="" data-ratio="0.6144578313253012" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG9aITupiazjELR6Rx4yaaiaO69EqzOiapg0uGzPDKarSF72U7BlsfPognAA/640?wx_fmt=png" data-type="png" data-w="1328" style=""></p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">宋老师</strong>：因为我们有一些产品的经历，就提了个问题，说我们开发这个东西对真正的用户有没有价值。举一个例子，比如说我们之前有一个团队做过 AI 识书的产品，用户把书拍个照，应用告诉你这本书的书名是什么，这样很奇怪，用户也不买账。这说明直白的演示AI技术不够有趣不够吸引人。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">我们的产品经理最终做了一个无奈的选择，把识书的功能包装了一下，变成根据识别到的书的类别，预测用户的性格。结果惊奇地发现，这样做之后的流量要大得多。这件事情和刚才的例子一样，<strong>只有我们的模型和产品真正能触及到用户的精神世界，关联到他自己的命运，他的想法和情感，这些人类很柔软的部分的时候，用户才会感兴趣。</strong>换句话说，我们的产品不应该是让用户帮我们检查 AI 的效果，而是让用户有一种跟（精神）世界的互动。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">我作为应用评测组的负责人，我主要想的问题就是，不管模型怎么做，我一定要让它做出来的东西对用户是有价值的。于是，我们当时就自己在朋友圈里偷偷下了很多图和朋友圈配文，形成了测试集。这个测试集的目的就是去看看，如果是人类朋友圈这样的图文匹配，模型有没有办法把它检索出来？</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">比如一句话说“我今天回去外婆家看她”，模型能不能检索出一个老房子呢？这个其实差别挺大的，因为模型（更）可能检索出一个老奶奶，就是很直白的检索。反过来说，一句朋友圈回老家的话，配上一个老房子，是很贴切的。配图中的老房子会引起大家很多对童年的想象和回忆。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<p style="white-space: normal;box-sizing: border-box;">所以我们当时并没有在意以前的数据集是怎么设定这个问题的，而是我们能够跳出来，更关注用户会感兴趣什么，人在用文字和图片去互相搭配的时候，他们是怎么做的。后来，随着我们的数据集不断扩充，微博也成了我们收集数据的来源之一，而且微博上的图文和朋友圈的很像。用了这样大量的数据后，模型的效果也提高了很多。所以我觉得从一个项目的角度来看，我们总结出来：<strong>目标很重要，目标设定对了，才能够带着大家去做出一个（对用户）有价值的模型。</strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong><br></strong></p></section></section></section><section style="box-sizing: border-box;font-size: 16px;"><section style="margin: 15px 0%;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 15px;line-height: 2;padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">3. 文澜双塔模型的核心与优势</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：我们发布会时的 PPT 也特别整理了双塔模型与单塔的区别，包括双塔模型的关键设计点。</span></p>
<p style="text-align: center;"><img class="rich_pages" data-galleryid="" data-ratio="1.1359832635983265" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG9LxPduGC9grXA3h1ghibM2a6X6HKXuqshTU4TYakfFsvW02RrKZKHpNg/640?wx_fmt=png" data-type="png" data-w="956" style="width: 424px;height: 482px;"></p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>卢老师</strong>：假设我们有两种模态，比如图、文两种模态。如果用两个像 Transformer 一样的结构去分别表示它们，那就是双塔结构，反过来如果是用一个 Transformer 同时表示两个模态，那就是单塔结构。<strong>双塔除了在假设上有一定优势和特点，在实际部署时也具备它独特的优势</strong>。双塔模型中的一个塔去学习图的表示，另一个塔去学习文的表示。这样（的方式）在工程实践上有个特别大的优点，那就是其特征提取的部分可以离线进行，更容易满足工业落地的需求。相反，如果是单塔的模型话，就不可能提前进行特征提取的计算。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>宋老师</strong>：另一个优点是，像我们论文里提到的，当我们出现比 Transformer 更好更强的特征提取模型后，我们可以把双塔模型中的特征表示部分都替换到更好的新模型。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>卢老师</strong>：<strong>说到这里就得提到我们对于双塔工作的一个最新拓展。</strong>目前我们双塔还是图文两种模态，但我们扩展后设计了一个四塔结构。四塔结构的核心思想就是说，在图文两种模态上都分别变成双塔结构。在单一模态中，使用多个分支，每个分支依然可以利用单塔的优点。比如说我们目前图像表示有两类方法，一种是基于 Transformer 的 ViT 方法，一类是基于 CNN 模型的，这两类方法其实是各有优缺点的，目前暂时是无法替代彼此。所以我们就想说，那要把图像的表示学习变得更强，就可以把这两类方法融合在一起，让它们在图像的单一模态学习中进行交互。这样的结构还有一个优点，就是可以利用别人提前训练好的 Transformer 和 CNN 作为基础。另一方面，我们的文本模态中，也可以变成双塔结构，很适合多语言的学习。这样会非常非常强大，我们对于这个模型的效果非常有信心，<strong>具体细节我们会在六月份进行重磅发布。</strong></p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>宋老师</strong>：其实我们的文澜模型在这第一版发布的时候，我们跟 OpenAI CLIP 是同时开始的，卢老师的模型已经设计好了，只是我们的机器一直没有到位，跑不了大量的数据。然后到1月1日的时候，卢老师已经把双塔模型优化得比单塔模型效果要好得多了，然后结果人家1月6日 CLIP 和 Dall-E 就出来了，可惜我们的机器还是没有到。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">不过我们仔细学习了他们的论文，发现其实有几个点我们的模型是很独特的（是跟 CLIP 不一样的）。所以，当时我就跟他们说，不要着急，咱们慢就是快，还是按照自己的设想去一步一步来，最后效果肯定会超过他们的。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">一方面是 CLIP 当时还是主要基于实体的，它先找的是 wikidata，把 wiki 上的 entity。所以你可以想象到，它们（的数据）还是一个个独立的实体。而我们（文澜）其实是在互联网上，包括咱们大家常见的微博、微信这样有大量热度在的数据上去收集的图文。<strong>那么我们这样得到的数据其实是句子和图片（而不是实体和图片）。</strong>而且这样的数据分布是非常满足我们网民（用户）大家关注点的一个真实分布。这样一来，我觉得我们其实从数据上就可以打败 CLIP，因为 CLIP 太依赖于实体了，这会导致 CLIP 虽然可以把一个物体做得很好，但它很难上升到句子级别整句的理解。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">而我们高瓴人工智能学院的一个观点是，<strong>做自然语言理解其实应该是一个整句的理解</strong>，也不要像以前一样非要把它分成主语、谓语、宾语，然后再慢慢去组合。其实人有时候就是在多模态的学习中，别人说了一句话，发现在这个情景里就会有这么一个动作或者有这么一个图景出现，他就对应上了。就像我们后来说的，其实2005年《Nature》的文章就已经证明了，就是说人对于同一实体的图像和文字的（认知建模）其实是在同一个神经元上去激活的，也就是它会映射到一个地方。就好比我们卢老师的模型，它可能本来是有一个文字的空间和有一个图像的空间，它通过对比学习之后，让两个空间中相关联的图和文对齐到了一个空间上去。</p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'><br></strong></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">4. 文澜与 OpenAI CLIP 的区别</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：当时我们是从2020年9月份开始做，对比学习肯定是我们和 OpenAI CLIP 不谋而合的地方。</span><br></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'><br></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'>卢老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：OpenAI 做 CLIP 的资源肯定是比我们多，所以他们的 mini-batch 的设置非常大，直接升到了10000+，这点上我们做不到。那我们文澜模型的资源是有限的，于是我们就加了一个辅助的队列，将一些重要信息存在另一个地方。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>这个队列其实是我们存储负样本用的。这里有一个假设是说，当前 mini-batch 中的图文都是相关的，它们是匹配的，至少是“弱”匹配的，因为它们是从数据中直接读取到的。那么我们就把这样的数据叫做正样本。而前面 mini-batch 的东西毕竟不是同时出现的，</span><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>当我们把当前 mini-batch 中的图文与前面 mini-batch 的图文重新组合，我们就假设（组合出的新图文对）它们是负样本，是不匹配的</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'><br></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：这样的设计方式相当于我们选出了数据中最难的负例让模型去学习，也缓解了数据噪音的问题。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>卢老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：我们还在研究另一个问题，当我们把参数提高到100亿的量级后，为了工业级落地，如何把这100亿参数存放在一张卡里，比如变成半精度等。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">5. 对比学习与对抗学习</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'>卢老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：</span></strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>对抗学习的话是从生成模型中来的，它有两个目标，一个是要生成一个东西，另一个是要判别这个生成的东西真不真实。这有点像纳什均衡。换成数学的话就非常简单，生成式对抗网络（GAN）的统一目标函数就是一个 minmax/maxmin 函数。</span><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'></span></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>所以说，对抗学习的两个目标是“反”的，它更强调任务上或者目标上的一个对抗，最后达到某一种平衡，就到最优了。但是对比学习的思想虽然有点类似，其实很不一样。对比学习一般是学习两个东西的相似度，比如图文的一个相似度，也就是机器学习中的度量。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'><br></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'>小S</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：那我再延伸一个小问题，因为 GAN 最初阶段的训练会非常不稳定，是不是对比学习在训练上的难度会小一些。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'><br></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'>卢老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：它也会存在这个（训练不稳定的）问题，但因为现在对比学习也发展得很快，它相较于 GAN 要容易训练一点，模型不是那么容易崩掉，稳定性要好一些。或者只要像我们刚才提到的，只要维护的队列足够大，（对比学习的模型）它肯定不会崩。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">6. 多模态模型的机遇</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'></strong><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;box-sizing: border-box;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：我自己还是挺看好多模态的，</span><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>我觉得比较有价值的几个点，一个就是多模态的理解。</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>以前大家做文字理解的时候，是要把文字打上一些标签，比如这是名词，那是动词等等。但我比较相信整句理解这件事。如果你给了机器足够多的数据让机器同时看到文字和图片的话，机器其实就可以整句地进行理解。而且理解未必是要在单一的模态里头去给它打标签，有可能就是说人理解了一个句子或者一件事物，就是在头脑中模拟出来了一个画面，甚至是一看到别人说那是丁香花，我会有闻到花香的感觉。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'></strong><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>这个其实在认知上是有支撑的。</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>我今年上的新生研讨课，主要就讲一本书叫，《Louder Than Words》。</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>这本书它其实就是一个认知科学家在2012左右的时候出了一本书，讲这个</span><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>体验模拟假说</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>，其在认知领域已经有30年的历史了，这个假说（的出现时间）非常巧合的是跟咱们神经网络的崛起是重合的。</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>那么其实它（这个假说）的意思就是说我们怎么算理解一个语言，30年之前的主流大家认为语言是思想语言，是符号系统，理解语言就是从词开始，再去构建语法，再到理解。</span></p></section></section></section><section style="box-sizing: border-box;font-size: 16px;"><section style="margin: 15px 0%;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 15px;line-height: 2;padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="text-align: center;"><img class="rich_pages js_insertlocalimg" data-ratio="1.5221518987341771" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG9mnFcSCZdlOuHVQHpTxk4yGgpSQUXPpsMX6cDczdWUMsiaysaGZoxr8g/640?wx_fmt=jpeg" data-type="jpeg" data-w="948" style="width: 325px;height: 495px;"></p>
<p style="white-space: normal;box-sizing: border-box;">但是从认知的角度，其实大家后来发现总是有些地方解释不通。然后他们就想了一个新的假说，就是说人其实理解语言是在模拟，他（人）是用他过去的经验在想象说，如果是我的经验的话，我身处其中的话，我是一个什么样的体验。这个假说也有很多实验证据，比如说让被试躺在一个 fMRI 磁共振的机器中，当被试听到一些语言的时候，可以看到被试（的大脑）不仅是语言区在活跃，语言涉及到的视觉的部分，其视觉区在活跃，涉及音乐的部分，听觉区在活跃，你涉及到运动的话，甚至运动区都在活跃。而且他们的实验做得很严格，当对被试说握手和握拳，就会用类似心电图的方式检测手是否有震动，结果虽然被试的手没有动，但（大脑中的）运动区是活跃的。<br></p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">那么其实你要做到一个文字的理解，是可以联系其他的图像、联系其他的视觉联系，联系声音、联系其他的区域去理解的，这样才比较像真正的人的（理解）模式。<strong>人是多模态学习的</strong>。这样有一个特别大的好处。虽然基于符号的理解很简洁，但它会失去很多细节，比如女孩和女孩能匹配上，但女孩的衣服发型等细节都没了。当我们能进行模拟的话（体验模拟假说），我们就能够借助图像，把一些常识、细节的东西填充进去，就不是干巴巴的一个文字了，也不是说我拿文字放到一个知识库里头去找到了一些有限的属性（property）了。这个假说举了一个例子，问被试大猩猩有没有鼻子，你不会是说去脑海里翻一本字典。但是你脑海里一定会有一张大猩猩的图片出来，然后你再仔细地用你的 Mind Eye（心智之眼）放大那个图片鼻子的部分，去想它到底有没有鼻子。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">所以其实多模态的基础是这个东西，就是说人其实理解语言也好，或者是看到的图像为什么突然产生了感想（文字），这些其实是通的。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>而且，体验模拟假说可以非常好地解释，为什么人和人是有时候很难沟通的。这正是因为理解其实是个性化的。</strong>你用你的经历去模拟，我用我的经历去模拟。哪怕是说狗，有些会觉得好可爱，但有些人被狗咬过，有过非常沉痛的经历，就会恐惧会觉得痛。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">所以有了多模态学习的话，未来的理解可以是个性化的，而不是基于人类整体的。当我们有一个库，我们可以去模拟它（不同人的经历），比如说我们有不同的用户的信息的话，我们知道它是来自不同年龄段的，然后大家对一个东西的反响，比如说我们拿一个文字去检索，那就感觉是不一样的。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>小S</strong>：可能这也是证明了为什么我们总说文字的交流是冷冰冰的。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>宋老师</strong>：对，每个人的编码器和解码器都是有细微差异的。当另一个人的解码器输出了一句简短的话，传到了你的耳朵里，但你是用自己的解码器去解的，就会有不一样的理解。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>小S</strong>：刚才您提到的《Louder Than Words》，可以再多介绍一些么。</p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;"><strong>宋老师</strong>：其实我也和我的朋友一起翻译了这本书，应该会在5月份左右由湛庐文化出版其中文版，书名叫做《我们赖以生存的意义》。我们未来在高瓴人工智能学院也会开设认知神经科学相关和人工智能伦理相关的课程。</p>
<p style="text-align: center;"><img class="rich_pages js_insertlocalimg" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG9TeHh5PxZQpgZAyxC6oZz7Z5kwakco6kU21HTBOE4wa6ggUONavJRqw/640?wx_fmt=png" data-type="png" data-w="800" style="width: 321px;height: 321px;"></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">7. 多模态模型的挑战</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;'></strong><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;'>：就像刚才说到的一个，多模态理解是一个非常好的方向。我和卢老师现在也在合作，做文字和图像的理解。另一个是，我非常感兴趣（基于）多模态理解的对话场景。</span><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;'>虽然多模态听起来很有道理，但是其实它很挑应用</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;'>。如果应用定义不好的话，你可以多模态用了还不如单模态好，这点上我们也有一些失败的尝试。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<p style="white-space: normal;box-sizing: border-box;">但是未来所向。我们大家还是希望说，有一个 AI 站在你面前，就像科幻片里面的样子，你能看到 AI，AI 也能看到你。<strong>所以，我有一个特别的研究方向，就是多模态的表达</strong>，也就是说如果我们有了 AI 的形象之后，我们能不能让他也能活灵活现，像人那样讲话的时候是有喜怒哀乐，甚至不止喜怒哀乐，他的动作以及更多的富的表情，像我们的表情包一样，有更细致的一些表达。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<p style="white-space: normal;box-sizing: border-box;">像我们（高瓴人工智能学院的）胡迪老师也有一个特别的方向，也是关于多模态，他主要是声音和图像的一个同步。比如说一个 video 里面有声音，他就能知道是哪个乐器发出的声音，然后他甚至可以研究一些东西，可以帮助盲人看到这个世界。它可以把图像中动的部分转成一个声音的信号，然后通过一个设备，去传递到盲人的脑海中。这样的话，盲人就不会像大家以为的那样戴一个样子，看到前面说有一个凳子，而是说盲人能同时得到图像和声音等细节的输入。盲人得到的不是你（机器）咀嚼过后喂给他的东西，而是盲人头怎么转动，朝不同的地方看，声音信号等都有可能变。这样盲人甚至可能慢慢地形成一些新的连接，会慢慢理解，声音这样的时候可能前面是有东西，那样的时候前面可能是有红色的花，他（盲人）就会慢慢开发出一些新的（认知的）部分。我认为这是非常有价值的。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<p style="white-space: normal;box-sizing: border-box;">所以说多模态是我们人大的人工智能学院一个非常新的方向。我们希望也能把这个方向建设成一个品牌，未来大家可能对多模态感兴趣了，能联系我们。对多模态感兴趣的学生能来我们学院学习。我们也可以给企业输出很多模态的应用，输出懂多种模态的人才。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">8. 文澜首个工业级应用“布灵”心情电台</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'></strong><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：当时卢老师的（双塔）模型已经非常好了，我们就成立了一个应用组。应用组是我带的，学生组长是卢老师的学生刘光镇。我们的想法是，不管是文配图还是图配文，我们一定得是高质量的东西，别人（用户）才需要。我们最开始尝试了另外一个应用（这个应用应该在四月底上线），利用很多金句网站，相当于大家把一些非常经典的人类结晶、有深意的话，去总结摘录到这些网站上。在之前，这些金句很难和图像去配上，因为它是独立的，都是文本的，也不曾有人给很多这样的训练数据。但有了我们卢老师这个模型之后，我们就很容易，因为我们只要把这些金句文字全都转成一个 embedding，一个高维向量，我们再把用户上传的图像直接就去检索的话，就能找到很好的（结果）。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>然后（刘）光镇就带着同学们，郑伟昊、隰宗正、金楚浩，大家一起很快在一个星期内搭出了一个 demo。当时我们就被结果震惊了，大家都在群里玩得停不下来。有一个例子就是，大家把 coding 的图像截了个屏，AI 说“这个也是没有办法，还是重写吧”。居然第一次被 AI 嘲笑了。就这样一直玩一直玩，在那时候我们就很有信心了，这个应用就已经可以凸显出我们匹配的作用和能力了。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>但即使是这样，你还是会觉得别人好像也能做到，因为这种评论和图还是可以在互联网上可以收集到的，还不够有差异化。但是我们后来我硬盘整理的时候，刚好就发现了一堆歌词的数据，而歌词其实也是文本，虽然歌词比较长，但我们可以两句两句去做，这样就能找到（图片）最相关的两句歌词了。（刘）光镇可能就花了一两天的时间，就把（歌词版本的）这个东西调了出来，相当于只是换了一个 index。我们看了一下效果，发现很神奇，完全不需要挑例子（cherry pick），只要随便看一个例子，就是好例子。</span></p>
<p style="white-space: normal;box-sizing: border-box;text-align: center;"><img class="rich_pages js_insertlocalimg" data-ratio="1.7786666666666666" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG9SziaSib6QfHzuicoJl9HyyJ0ngia9jjVPR5MPSeMPN71qMN9vo2sZ7XmbA/640?wx_fmt=png" data-type="png" data-w="1125" style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;width: 361px;height: 642px;'><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>这时候的感觉就是（这个应用和模型效果）能到产品的级别了。然后我上了儿子的照片，我们的模型返回的是五月天的《咿呀呀》，很纯真的歌词；我们又传了一张朋友圈找到的沙滩上的脚丫的图片，返回的是《you make me free》的歌词。我们都感觉这非常配，也感觉有商机，企业是需要的，比如 QQ 音乐啊网易云音乐啊可能是需要的。其实有更新的歌词，更多的版权的时候，我们真的可以做到，直接输入图像来听音乐，可以变成一种新的听音乐的方式。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">9. 没有给图片直接打标签的原因</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：这个和我们项目的思维有点关系，就是说看你的受众是内行还是大众。如果你的受众是内行的话，大家我们也讨论过说想让它变成更有价值的产品的话，可以后续做很多事情，比如我们可以识别到动漫的图片里的（卡通）人物的手插在兜里。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><img class="rich_pages" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG9HZZeAQa7iaDuERopP3icE6yovj1ukNiaia8iaAibedQbBAWYRU6yWHjgffCg/640?wx_fmt=png" data-cropx1="0" data-cropx2="1000.0000000000006" data-cropy1="7.168458781362005" data-cropy2="492.8315412186378" data-galleryid="" data-ratio="0.486" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG9yBGltE3ZWh70PChP4w2T7Eiavgf2TOSQHYkL0wkwYEAN9BtCicPPxt4Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1000" style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;width: 558px;height: 271px;'><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋</strong><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：由于时间有限，我们暂时还没发布这个功能（产品）。不过未来我们还是想做的，因为不光是给内行，我们觉得它对企业的应用都是很开阔的。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>小S</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：这个绝对有价值，因为我了解到很多画画的人，他们画画的时候会想去参考（别人的图片的）姿势，也可能想参考一些构图等等。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：但我们可能也是觉得图片的版权更敏感一些，我们不能去把一些商用或者别人的图片展示给用户看，这也是为什么我们做的应用是让用户自己上传图片。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">10. 如何处理带有滤镜的图片</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>卢老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：我怀疑本来互联网上的很多图片都是有滤镜的，其实我们当时并没有专门针对滤镜做什么特殊处理。而且现在的 Transformer 等模型很强大，滤镜对它们来说并不是难题。</span></p>
<p style="white-space: normal;box-sizing: border-box;text-align: center;"><img class="rich_pages js_insertlocalimg" data-ratio="1.7786666666666666" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG9lIgEQMbYlnJvoel8UsbbEyf2szJPX62OJFicrC501RoOAJSKEOoXdDg/640?wx_fmt=png" data-type="png" data-w="1125" style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;width: 334px;height: 594px;'><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></strong></p>
<p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：我们觉得数据应该是很关键，因为我们收了很多互联网上的数据，其五大来源主要包括 index 图像搜索里的，然后还有微博、微信公众号的，还有新闻的，包括我们窦老师保存了从2014年到现在的各大门户网站上的新闻，然后还有百度百科的。所以其实我们（的文澜模型）也涵盖了 entity（实体），但是我们 比 CLIP 多了4种（数据），这样子一来，我感觉刚好用户也喜欢在互联网上发布有滤镜的图片，我们的模型也就自然会学到有滤镜的一些匹配结果。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>这也是我们觉得：其实做研究的人不必要把那个东西分得那么细。就像（对待）孩子一样，你（教）给他的就是输入输出，然后他（孩子）其实会在里面找到这个规律。</span><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>而且当他有了这个学习机制之后，他是很容易更新的</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>。比如说，未来互联网变了，出现了各种新的东西，我们再用卢老师的模型去训练一下，包括新的语言出现，比如以前没有“内卷”这个词，现在有了，这些都没有关系。只要你在网上留下了足迹，比如我就说这个图就代表内卷，然后模型就可以学到这个过程。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">11. 多模态研究的未来</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：做研究的话，未来我们学院会有不少的东西会产出。我们会关注，（如何）让 AI可以打开眼睛可以看，你可以带它出去看这个世界，然后它都可以基于卢老师的模型，有感而发地去跟你攀谈。再比如对话结合上这种图片的评论，会是很自然的。</span><br></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p>以前的 AI 都是被动式的，你必须发给它一句话，它才能回你一句话。<strong>但以后的 AI 它可能是主动式的</strong>，你带它出去，它看到周围了，可能会说，哇这个郁金香好可爱。</p>
<p><br></p>
<p><strong>小S</strong>：这点确实是，尤其现在很多智能音箱都是有屏的。</p>
<p><br></p>
<p><strong>宋老师</strong>：对，所以这（个方向）可能会打开对话一个新的空间。</p>
<p><br></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">12. 超大规模预训练模型是必然的趋势么</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p><strong>卢老师</strong>：当时发布会上有一个访谈，有几位老师参加了，关于这个问题，有人支持，有人是一个保守的支持。对于目前这样一个起步的阶段，我认为先去加大（模型和数据的规模），是没问题的。但是要考虑部署的话，肯定是要做减法的。<strong>其实我们团队已经开始做减法了，比如100亿的参数我要压缩到一个卡。</strong></p>
<p><br></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">13. GPT-3生成大量数据带来的潜在风险</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋老师</strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：我感觉说 GPT-3 生成的东西其实还是在东拼西凑。虽然乍一看觉得它的模型（生成的东西）还挺顺畅的，但是还不足以产生新的东西，这是我对 GPT-3 的一个看法，就是它没有产生新的价值。</span><br></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p>那么在这种情况下，你再把（GPT-3生成的）数据喂回去训练新的模型的话，你的模型会变坏的，不会变好。从这个角度来看，我觉得人一点压力都没有。比如说写小说这件事情，很多人也声称（他们的模型）可以写小说了。其实小说是有一个新的内涵，你的故事得有新意，（写小说的人）对这个世界的人性很了解，（小说的故事）中间有冲突，要不断地去解决这样的冲突。并不是说，你（的模型）写了一段像小说的文字就叫小说。</p>
<p><br></p>
<p>我一直觉得这个还是文字生成里很难的一个问题，因为它（模型）要对世界人性（和其它）有一些了解。而且我以前也在 CNCC 2020上谈过，你（的模型）生成了也不代表你（的模型）创作了。换句话说，<strong>一个 AI 它能生成一段顺溜的话，和它能够创作，这是两回事</strong>。你（模型生成的东西）的新颖度上，你的价值感上，还得要达到一定的程度，才能去跟人的创作做对比。所以类比来说，我感觉其实 GPT-3 大家也不用太担心（它带来的风险）。</p>
<p><br></p>
<p>而且像我们卢老师，他关注和需要的不是大，而是质量好，这也是我们的（文澜）项目组得到的一个比较成功的经验。从一开始，我们就建设评价，我们不满足于自己这个东西生成了一个不错的样子。我们还是要让人终极地告诉我们，它（模型的效果）到底怎么样。我们也可以看到说，它（模型）离人类的理解还有很大的距离。（虽然）我们可以包装一个产品，它好像是可以掩盖住一些弱点，让人感到有一些趣味性，但是仍然它（模型）跟人的理解力还差了很远。</p>
<p><br></p>
<p>我们（当然）会继续探索，但是我们（团队）这边不会觉得一个超大的模型就一定很 sexy，不会觉得一个模型超大就很有意思。我们要实际能落地，像我们卢老师专门要把它（模型）做小。因为最终一个技术能不能有生命力，你还是要看能不能对别人有影响，人家能不能用得上，而不是发一篇论文，发一篇报道（而已）。</p>
<p><br></p>
<p><strong>卢老师</strong>：我想慢慢把100亿的参数压下来，压到一张卡里，这样能节约很多成本，也能加速很多研究。</p>
<p><br></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">14. 超大规模预训练模型的 fine-tuning</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;box-sizing: border-box;"><strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>卢老师</span></strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：学界关注的设计新的 fine-tuning 模式，可能主要存在于语言文字这个模态中。我们（文澜多模态和图像）这边其实没有那么严重的问题。因为图像不管从哪里得到的，数据源是什么样的，它的范式都是比较固定的。所以我目前没有发现图像这边需要设计什么新的 fine-tuning 机制。比如图像那边也有小样本学习，用我们的模型去做小样本分类，这个落地的结果是值得期待的。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p><strong>宋老师</strong>：像我们当时做漫画的任务，我们是用一个单独的动漫数据集去做 fine-tuning 的，才达到了很好的效果。这也是我们一个不一样的地方，它里面会产生一些很奇妙的东西，是非常具体的标签。</p>
<p><br></p>
<p><strong>卢老师</strong>：谈到这里，我就想起最近经常有人问我，让我用一句话讲清楚多模态的意义。我就说，我们目前做多模态就像很多年前互联网刚兴起时的搜索引擎，你可以这么认为。互联网出现的时候，搜索引擎就是刚需。<strong>那么目前，多模态可能也是刚需</strong>。</p>
<p><br></p>
<p><strong>宋老师</strong>：这可能也是我们建高瓴人工智能学院的动机之一。以前我们的计算机（研究）还是很分离的，视觉是视觉，语音是语音，文本是文本，大家“老死不相往来”。<strong>但是人工智能其实可能就像人一样，它就是一个多模态的输入，然后一个多模态的处理，最后到一个多模态的输出</strong>。所以我们人工智能学院不会把这个东西限制得特别严格。像我们卢老师本来是做视频视觉的，但是他一样可以做多模态。而我本来是做文本的，我也一样可以做多模态。大家会就着这个研究问题，去增加自己的知识，去调整。</p>
<p style="white-space: normal;box-sizing: border-box;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">15. 智源、悟道及文澜坚持开放生态的初心</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p><strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>宋老师</span></strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>：</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>其实我觉得智源做这件事情非常有价值，这也是我们为什么投入了这么多老师和学生去做这件事情。</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>我们一共可能有七、八位老师，然后27位同学参加这个项目。</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>我们觉得做这个事情他非常得有影响，做事情的过程一环会影响下一环。</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>如果你自己不定义好你的任务，做出来的东西却不是很有价值。</span></p>
<p><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></span></p>
<p>另外就是说我们要把数据收好，然后使得它比较通用，我们辛苦做一次这样的（收集数据的）事情，可以方便到可能十家百家，对未来的应用有价值。我们把做的事情做仔细，像卢老师把这个模型这么仔细去设计，（也是希望）能把探索出来的结果共享给千家百家。</p>
<p><br></p>
<p>至于我们的应用为什么要花力气去做，而其实在我们的任务书上没有应用这一条，那么我们是想说，你不能只有一个模型一篇论文扔到那就完了，因为这样大家并不能直观地知道你这个（模型）能干嘛。所以，我们要用“布灵”AI 心情电台去展示我们模型的力量，展示它能做什么。<strong>只有大家能随手去试我们的东西的时候，才会让一些本来并不关心模型不关心预训练的人觉得，原来这个东西跟我有联系</strong>。</p>
<p><br></p>
<p>在智源的平台上，包括北京市政府给我们很多支持，陆续地还会有一些支持，包括算力和数据，包括搜狗公司也给我们大量的支持。那大家最后都可以感觉到，这个支持是集中在了一个点上，大家都认为这件事情该做。（所以）在这么快的情况下，我们能把这个东西构建出来，要感谢北京市政府、北京市科协、北京智源研究院。在文继荣老师的领导下，我们能够率先做这么大规模的通用图文多模态模型，真的感觉是生逢其时。能够发布给大家用，是我们的荣幸。我们也希望，未来有更多的人，能够在这上面做出更深入的研究和应用。</p>
<p><br></p>
<p>这里，也特别想感谢北京智源研究院的黄铁军院长，他很形象地跟我们解释为什么要做超大规模的预训练模型，为什么北京智源研究院要做这件事。他提出，<strong>我们正处在一个从“大炼模型”到“炼大模型”的阶段</strong>：“过去五年多，全球研究人工智能研究者最重要的就是训练各种各样的模型，这几年有一点‘千村万户炼模型’的状态。不论大小公司还是学校，都是拿一个开源的框架，收集一批数据，然后拿着这个模型去解决问题。虽然这是可以成功的，但作为一个现代化产业，近年来人工智能的发展应该从家家户户‘大炼模型’的状态逐渐变为把资源汇聚起来，训练超大规模模型的阶段，通过设计先进的算法，整合尽可能多的数据，汇聚大量算力，集约化地训练大模型，供大量企业使用，这是必然趋势。”</p>
<p><br></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">16. 数据开源与知识产权</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p><br></p>
<p><strong>卢老师</strong>：<strong>我们更强调的是模型上的开源，也就是关注是否能提供给别人使用</strong>。对于文澜目前的阶段，当我们发了一些论文后，也会把之前的一些模型和代码开源给大家使用。</p>
<p><br></p>
<p><strong>宋老师</strong>：这也是当智源老师问我们说，你们那个多模态的数据能不能共享出去，我们为什么说很难的原因。因为那个图片（的版权）不属于我们，文字说白了也不属于我们。但是我感觉在做研究的时候，<strong>我们做研究的人要有一个红线，要有责任感，我们用数据是为了做有益的事情，而不是有害的事情。</strong></p>
<p><br></p>
<p>像我们收集这个数据花了很多时间，然后把数据的知识凝结在了这个模型里，也相当于各个图片的创作者和文字创作者都对我们文澜模型有很大的贡献。所以我们非常感谢他们。那么，我们最后共享的时候也不会直接把他们的图片和文字去共享出去，避免去破坏他们的知识产权和侵害他们的知识产权。所以我们一直坚持说只开放 API，也可以提供我们的 inference model。而且一般来说咱们（研究者和企业）感兴趣的人就是这个东西（model）。</p>
<p><br></p>
<p>当然，我也认为，的确 AI 想要做好，就像刚才说的，我们需要给 AI 很多“经历”，需要让 AI 看很多数据，看人类的文字，它才能够形成自己的“经历”。如果我们把数据管得太死，一点都不能用，那 AI 也就没法做了。就相当于你从来不给你的孩子看这个世界，把他关起来，然后逼迫他必须长大，这是不可能的。</p>
<p><br></p>
<p><strong>卢老师</strong>：宋老师说的也提醒了我一点。等我们6月份公布我们的2.0版本模型后，我们也非常希望感兴趣的研究者能用我们的模型去刷榜，将我们的模型用到各种任务中去，展现它的价值。</p>
<p><br></p>
<p><strong>宋老师</strong>：对，其实从文老师到我们人工智能学院的其他老师，我们都认为要做有价值的研究。我们并不一定要刷榜，而是要发现一些新的有价值的任务。我们中国人不要总是跟在别人后面去做一个类似的（任务）。我们更多的时候（可以）去思考，AI 的本质问题是什么，我们能不能突破它，能不能构建一个任务，吸引全球的研究者一起关注。我感觉我们中国的 AI 学术界是和国际持平的，没有任何落后的地方。</p>
<p><br></p>
<p>而且我们中国人特别勤奋和实干，其实（我们）这个（文澜）模型能做成，我觉得正是因为在中国这个大环境底下才能够做成的。在国外，可能只有公司资助的像 Open AI 才能做。而公司又有它自己的利益，未必会让你去用它的东西。<strong>我们（国内）反而有政府非常强大的支持，我们可以构建生态，回馈到整个企业界，这是一个更正向的循环。</strong></p>
<p><strong><br></strong></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">17. 与智源、悟道及文澜的合作机会 </strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p><strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'></span></strong></p>
<p><strong>宋老师</strong>：我们跟高校应该可以合作，其实我们自己不挑合作伙伴，企业和高校都可以。高校与智源（的合作）也会非常顺畅，毕竟我们就是高校，智源就是政府支持的一个研究院。它（智源）是一个新型的模式，比我们去直接申请国家自然基金的时间周期要快，有效率很多。</p>
<p><br></p>
<p><strong>咱们中国不仅在创新技术，也在创新机制</strong>。那么，我们未来跟高校怎么合作，跟企业怎么合作，我觉得是有很多机会去探讨的。因为政府想给大家营造一个比较宽松的环境，让大家去尝试各种合作机会和形式，而不要管得太严苛。黄老师也是说让大家都能够用起来（我们的模型和成果），把生态建起来，这才是首位的。生态的影响是更有价值的。</p>
<p><strong><br></strong></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="height: auto;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 10px 0% 15px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 84.3137 84.3137 0%;height: auto;border-style: solid;border-width: 1px;border-radius: 12px;border-color: rgb(255, 154, 171);overflow: hidden;padding: 5px 10px;background-color: rgb(255, 249, 251);align-self: center;box-sizing: border-box;"><section style="font-size: 15px;color: rgb(255, 114, 138);letter-spacing: 3px;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;box-sizing: border-box;"><strong style="box-sizing: border-box;">18. 下一步的计划</strong></p></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 10 10 0%;height: auto;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin-right: 0%;margin-left: 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section></section><p><strong><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'></span></strong></p>
<p><strong>宋老师</strong>：<strong>下一步我们想做一个更通用的模型，去落地到各种各样的应用场景和问题中去</strong>。<span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>我们</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>自己文澜的目标的话，第一个是在下一阶段6月份发布的时候，要是一个超过 Open AI 的上亿参数的模型。</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>第二就是要<strong>多语言</strong>，因为我们觉得图像是各个国家都能够看懂的东西，但人类有这么多语言，其里面有一个很根本的问题，就是语言的产生，就是涌现。</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>明明物理的世界都一样，为什么有各种各样不同的语言。</span><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>虽然在语言上，目前翻译已经做得很好了，但我们有没有办法让它们都直接地附着在现实世界上，让大家国际的同行也能受益。</span></p>
<p><br></p>
<p>打个比喻来说，我们可能是做了一个锤子，但是大家看得到很多钉子。<strong>我们是希望能让大家一起共建我们的生态，通过更多的合作机会，创造出更多好玩的落地应用</strong>。因为我觉得生态特别重要，当大家能够互相借力的时候，它的效率是最高的。每人都把时间花在新的一个研究点上，而不是说反复造轮子。</p>
<p><br></p>
<p><strong>卢老师</strong>：我们目前正要投一篇论文，当我们把我们的整个模型换成英文环境，没有经过特殊处理，依然效果很好。</p>
<p><br></p>
<p><strong>宋老师</strong>：而且当我们做到了计划中6月份的那一步，我们的下一步也是更开阔的。比如我们还有视频数据没有尝试过，而我们卢老师其实最拿手的是视频领域。现在我们正是短视频的时代，所以会有很多人感兴趣这样一个研究方向。</p>
<p><br></p>
<p>我觉得这一点特别幸运，原本可能高校由于资源等限制，只能跟着企业跑。但现在智源的创新机制，使我们（高校）也能超前一步，希望我们人大也能给企业带来一些价值。</p>
<p><br></p>
<p>当然，专业的人做专业的事。每个企业厂商都会有自己落地的场景，产品经理也都比我们脑瓜灵，能够想到更好的一个落地的方式。但是我们还是想尽可能地接近他们，能够让他们看到我们的价值。</p>
<p><br></p>
<p><strong>卢老师</strong>：是的，我也觉得我们高校研究者，也应该多多培养工程思维，不要只顾着模型和算法，还要多想想落地，这样做出来的研究有时候价值反而更大。</p>
<p style="white-space: normal;box-sizing: border-box;"><br></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;margin: 7px 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;width: auto;flex: 100 100 0%;height: auto;align-self: center;line-height: 0.1;box-sizing: border-box;"><section style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 154, 171);height: 1px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section><section style="display: inline-block;vertical-align: middle;width: auto;align-self: center;flex: 0 0 0%;height: auto;line-height: 0.1;box-sizing: border-box;"><section style="text-align: right;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 22px;height: 22px;vertical-align: top;overflow: hidden;border-width: 1px;border-radius: 50%;border-style: solid;border-color: rgb(255, 154, 171);background-color: rgba(255, 255, 255, 0);line-height: 0;box-sizing: border-box;"><section style="transform: rotateZ(315deg);-webkit-transform: rotateZ(315deg);-moz-transform: rotateZ(315deg);-o-transform: rotateZ(315deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;justify-content: center;margin-top: 9px;margin-right: 0%;margin-left: 0%;transform: translate3d(3px, 0px, 0px);box-sizing: border-box;"><section style="display: inline-block;width: 8px;height: 8px;vertical-align: top;overflow: hidden;border-style: solid none none solid;border-width: 1px;border-color: rgb(255, 154, 171) rgb(255, 255, 255) rgb(255, 255, 255) rgb(255, 154, 171);box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section><p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">最后，再次感谢文澜团队、卢志武老师、宋睿华老师、以及梁莉娜老师提供给我们这次宝贵的交流机会！<strong>我们也将从本篇文章的转发+评论中选出点赞最高的两位同学，送出宋老师最新翻译的书，《Louder Than Words》的中文版《我们赖以生存的意义》</strong>。</span></p>
<p style="white-space: normal;text-align: left;"><br></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">最后的最后，让我们一起期待文澜团队未来更多的成果！</span></p>
<p style="white-space: normal;box-sizing: border-box;text-align: left;"><br></p>
<p style="white-space: normal;box-sizing: border-box;text-align: left;">以上就是本次采访的精选内容。如果想更进一步了解文澜项目，可以：<br>1. 查看项目主页：https://model.baai.ac.cn/model/wl<br>2. 阅读论文《WenLan: Bridging Vision and Language by Large-Scale Multi-Modal Pre-Training》<span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>https://arxiv.org/abs/2103.06561</span><br></p>
<p style="white-space: normal;box-sizing: border-box;text-align: left;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>3. 使用“布灵” AI 心情电台：https://weixin.elensdata.com/<br></span></p>
<p style="white-space: normal;box-sizing: border-box;text-align: left;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>4. 使用文澜 API：https://model.baai.ac.cn/api</span></p>
<p style="white-space: normal;box-sizing: border-box;text-align: left;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>5. 参观《智能时代·女艺术家作品展》https://mp.weixin.qq.com/s/7fBdL1VPt7pcZ4MVX_cI1g</span></p>
<p style="white-space: normal;box-sizing: border-box;text-align: left;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'>6. 扫描下方二维码使用<span style="font-size: 15px;text-align: left;">“布灵” AI 心情电台</span>：</span></p>
<p style="white-space: normal;box-sizing: border-box;text-align: center;"><span style='font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><img data-ratio="1" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqxeMRwLDQmDLlCEAEnoqG96TzcjZWXcIHg7OzOubA9Iv1iazUEawicI2cianGUwHHq5C1pNib9AyJJQg/640?wx_fmt=png" data-type="png" data-w="529" style="width: 351px;height: 351px;"></span></p></section></section></section><p style="text-align: center;"><br></p>
<p style="text-align: center;"><img class="rich_pages" data-galleryid="" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240" style=""></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693390&amp;idx=1&amp;sn=d0660929e069455fc17302e5c9ce9bb0</link>
      <pubDate>Wed, 14 Apr 2021 02:30:17 -0700</pubDate>
      <guid isPermaLink="false">351761</guid>
    </item>
    <item>
      <title>大S的2021初絮叨</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content" style="visibility: hidden;">
                    

                    

                    
                    
                    <p style="text-align: center;"><img class="rich_pages" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740" style=""></p>
<section style="background-image: linear-gradient(326deg, rgb(237, 255, 233) 0%, rgb(223, 227, 226) 100%);box-sizing: border-box;font-style: normal;font-weight: 400;text-align: justify;font-size: 16px;"><section style="text-align: center;justify-content: center;margin: 25px 0% 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 95%;vertical-align: top;background-color: rgba(255, 255, 255, 0.73);padding: 5px 10px;border-style: solid;border-width: 0px;border-radius: 1px;border-color: rgb(62, 62, 62);height: auto;overflow: hidden;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;border-width: 1px;border-radius: 0px;border-style: solid;border-color: rgb(109, 128, 82);padding: 15px 20px;box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section><section style="font-size: 15px;color: rgb(94, 94, 94);line-height: 1.8;letter-spacing: 1.8px;padding: 0px;box-sizing: border-box;" powered-by="xiumi.us"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><img data-ratio="0.8314815" data-w="1080" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNr9f62KEIPM7myNgtzZjaauFzYsCnmQIpnGxck48duiaSIic7Kop3OgB1zPvibD5fqsGxXC6m9kHYm7A/640?wx_fmt=jpeg" width="100%" data-type="jpeg" style='font-size: 16px;color: rgb(51, 51, 51);font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;vertical-align: middle;width: 100%;box-sizing: border-box;'><br></p></section><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 20px;color: rgb(255, 255, 255);letter-spacing: 15px;line-height: 1.8;padding: 0px;text-shadow: rgb(255, 187, 0) 1px -1px, rgb(254, 221, 71) 1px 1px, rgb(254, 221, 71) -1px 1px, rgb(254, 221, 71) -1px -1px, rgb(254, 221, 71) 1px 0px, rgb(254, 221, 71) 0px 1px, rgb(254, 221, 71) -1px 0px, rgb(254, 221, 71) 0px -1px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">2020再见<strong style="color: rgb(255, 255, 255);font-size: 20px;letter-spacing: 15px;text-align: center;white-space: normal;background-color: rgba(255, 255, 255, 0.73);box-sizing: border-box;">！2021你好！</strong></strong></p></section></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 20%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: flex;flex-flow: row nowrap;justify-content: center;margin: 0px 0%;box-sizing: border-box;"><section style="display: inline-block;width: auto;vertical-align: bottom;flex: 0 0 auto;align-self: flex-end;min-width: 10%;max-width: 100%;height: auto;padding: 0px;box-shadow: rgb(185, 185, 173) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: right;font-size: 0px;justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 12px;height: 12px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: center;margin: 2px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 6px;height: 6px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 1px;border-style: solid;border-color: rgb(185, 185, 173);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section><svg viewbox="0 0 1 1" style="float:left;line-height:0;width:0;vertical-align:top;"></svg></section></section></section></section></section></section></section></section></section></section></section></section><section style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;vertical-align: top;width: 80%;height: auto;margin: 0px -10px 0px 0px;padding: 0px;box-sizing: border-box;"><section style="margin: 2px 0%;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;font-size: 14px;color: rgb(110, 173, 92);letter-spacing: 4px;line-height: 1.8;padding: 0px;box-sizing: border-box;"><p style="text-align: center;margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;"><br></strong></p></section></section></section><section style="display: inline-block;vertical-align: bottom;width: 20%;height: auto;align-self: flex-end;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="text-align: center;justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 95%;vertical-align: top;background-color: rgba(255, 255, 255, 0.4);height: auto;box-sizing: border-box;"><section style="justify-content: center;font-size: 6px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 50%;vertical-align: top;height: auto;box-sizing: border-box;"><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="margin: 0px 0% 15px;justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 15%;border-width: 0px;height: auto;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 91.8 70.9" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><path d="M22.4,26.1c-2.7,0-5.2,0.5-7.6,1.3L40.6,1.6c0.4-0.4,0.4-1,0-1.3c-0.4-0.4-1-0.4-1.3,0L8.9,30.6C3.5,34.7,0,41.2,0,48.5   C0,60.9,10,71,22.4,71s22.4-10,22.4-22.4S34.8,26.1,22.4,26.1z" fill="rgb(110,173,92)" style="box-sizing: border-box;"></path><path d="M69.4,26.1c-2.7,0-5.2,0.5-7.6,1.3L87.6,1.6c0.4-0.4,0.4-1,0-1.3c-0.4-0.4-1-0.4-1.3,0L55.9,30.6   C50.4,34.7,47,41.2,47,48.5C47,60.9,57,71,69.4,71c12.4,0,22.4-10,22.4-22.4S81.8,26.1,69.4,26.1z" fill="rgb(110,173,92)" style="box-sizing: border-box;"></path></g></svg></section></section></section></section><section style="font-size: 14px;box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: left;margin: 0px;padding: 0px;box-sizing: border-box;">2020年过去了，你问我会不会想念它，我也不知道。但是，2020真的是看似平淡，去实现了很多小目标的一年鸭！</p>
<p style="text-align: left;margin: 0px;padding: 0px;box-sizing: border-box;"><br></p>
<p style="text-align: left;margin: 0px;padding: 0px;box-sizing: border-box;">作为一个年更博主感觉是时候该更新一下公众号了，想跟大家聊聊过去的2020我自己的生活发生了哪些变化。这里也就不立flag接下来要以什么频率更新了，毕竟去年的flag也没坚持住…但还是希望今年能尝试一下Podcast～</p>
<p style="text-align: left;margin: 0px;padding: 0px;box-sizing: border-box;"><br></p>
<section><qqmusic class="js_editor_qqmusic qqmusic_iframe js_uneditable custom_select_card" musicid="127514565" mid="000n1GmU29xFE2" albumurl="https://y.gtimg.cn/music/photo_new/T002R68x68M000001ZaCQY2OxVMg.jpg" audiourl="http://isure6.stream.qqmusic.qq.com/C200000nc3df0IMhrS.m4a?guid=2000001731&amp;vkey=CDFC7D35F0B8C76F97EFC0E79781013FEF0B144E203FA25E8C2BC5FA1ECB1AD7DA592BA23902C5BF7595B67AE904057908FDF3CB71D6AC42&amp;uin=&amp;fromtag=50" music_name="从不" singer="THE9-刘雨昕" play_length="104" src="https://localhost:3000/cgi-bin/readtemplate?t=tmpl/qqmusic_tmpl&amp;singer=THE9-%E5%88%98%E9%9B%A8%E6%98%95&amp;music_name=%E4%BB%8E%E4%B8%8D&amp;albumurl=https%3A%2F%2Fy.gtimg.cn%2Fmusic%2Fphoto_new%2FT002R68x68M000001ZaCQY2OxVMg.jpg&amp;musictype=1" musictype="1" otherid="000n1GmU29xFE2" albumid="001ZaCQY2OxVMg" jumpurlkey="" data-pluginname="insertaudio"></qqmusic></section><section><br></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 98%;vertical-align: top;padding: 0px 10px;height: auto;background-color: rgba(255, 255, 255, 0.6);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40px;height: 40px;vertical-align: top;overflow: hidden;border-width: 0px;border-radius: 50px;border-style: none;border-color: transparent;background-color: rgb(255, 187, 0);box-shadow: rgb(236, 233, 224) 5px 1px 0px;box-sizing: border-box;"><section style="margin: 5px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;border-width: 2px;border-style: solid;border-color: transparent;padding: 0.1em 0.3em;color: rgb(255, 255, 255);text-align: left;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">01</strong></p></section></section></section></section><section style="margin: 15px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 90%;height: auto;border-width: 0px;border-radius: 0px;border-style: none;border-color: rgb(62, 62, 62);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><img data-ratio="1" data-w="800" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNr9f62KEIPM7myNgtzZjaaubsl7iajdgH1ibvtDPlOQicTic8T4r5xoOA0pEfKnLtmxKdETFVrRdZdgGQ/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="jpeg"></section></section><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(236, 233, 224);padding: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-family: PingFangSC-ultralight;box-sizing: border-box;" powered-by="xiumi.us"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">改变</strong></p></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="text-align: left;font-size: 14px;letter-spacing: 4px;padding: 0px 20px;box-sizing: border-box;" powered-by="xiumi.us"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">改变可能是2020年的第一个关键词吧，20年好像是以疫情开始的，这样一个黑天鹅事件，让大家久违的在家里呆了很久，在那些时间里面，我好像渐渐的有一些想法也发生了变化。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">好像更加懂得珍惜当下，更加享受生活，保护自己，感恩周边的一切。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">我开始学着理财，做一丢丢的长线投资。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">我一个从来都不打算买房子的人，好像突然觉得有个固定资产保障自己也不错。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">终于开始给自己打疫苗、买重疾险，为长久的健康保障。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">学会延迟满足自己的一些小愿望，让自己更加有前进的动力和目标。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">懂得不要一味地执着，适时的放弃，放过自己也放过别人。</p></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 98%;vertical-align: top;padding: 0px 10px;height: auto;background-color: rgba(255, 255, 255, 0.6);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40px;height: 40px;vertical-align: top;overflow: hidden;border-width: 0px;border-radius: 50px;border-style: none;border-color: transparent;background-color: rgb(255, 187, 0);box-shadow: rgb(236, 233, 224) 5px 1px 0px;box-sizing: border-box;"><section style="margin: 5px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;border-width: 2px;border-style: solid;border-color: transparent;padding: 0.1em 0.3em;color: rgb(255, 255, 255);text-align: left;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">02</p></section></section></section></section><section style="margin: 15px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 90%;height: auto;border-width: 0px;border-radius: 0px;border-style: none;border-color: rgb(62, 62, 62);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><img data-ratio="0.8898148" data-w="1080" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNr9f62KEIPM7myNgtzZjaauqPUtcHW3qiahfLiaicDYzjL25u2MpvGkXmkZ2vxbUM4yyyibZjPDdEmhDw/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="jpeg"></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(236, 233, 224);padding: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-family: PingFangSC-ultralight;box-sizing: border-box;" powered-by="xiumi.us"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">标签</strong></p></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="text-align: left;font-size: 14px;letter-spacing: 4px;padding: 0px 20px;box-sizing: border-box;" powered-by="xiumi.us"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">每个人的身上都有许许多多的标签，有些是自己贴上的，有些是别人贴上的。好像这几年，我一直都在思索，怎么撕掉一些我不喜欢的标签。但不可否认的是，有些标签，虽然我想撕掉，但可能却是别人眼中最有价值的东西。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">16年去东京，那里的我抛掉了“爱操心”、“决策者”的标签，做了半年无忧无虑的小小研究员。一直到现在，回想起那段毫无“偶像包袱”的日子，真快乐鸭～</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">17年迈入职场后，我从别人的跟班，变成独立的个体，再变成不可或缺的存在，这一路的起伏和挣扎，也都是在不断地和标签作斗争。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">20年底我终于鼓起勇气离开了Hulu，换了新的赛道。在面试的过程中，我在反复掂量和纠结的，也是到底我希望自己成为什么样的存在，我身上到底有哪些标签，我又到底要不要撕掉一些标签。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">不过时间长了，反而有些释然啦，为自己热爱的事儿努力，快乐就好。标签什么的，无论他人怎么想，或许时间会告诉我们答案。</p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 98%;vertical-align: top;padding: 0px 10px;height: auto;background-color: rgba(255, 255, 255, 0.6);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40px;height: 40px;vertical-align: top;overflow: hidden;border-width: 0px;border-radius: 50px;border-style: none;border-color: transparent;background-color: rgb(255, 187, 0);box-shadow: rgb(236, 233, 224) 5px 1px 0px;box-sizing: border-box;"><section style="margin: 5px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;border-width: 2px;border-style: solid;border-color: transparent;padding: 0.1em 0.3em;color: rgb(255, 255, 255);text-align: left;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">03</p></section></section></section></section><section style="margin: 15px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 90%;height: auto;border-width: 0px;border-radius: 0px;border-style: none;border-color: rgb(62, 62, 62);box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><img data-ratio="0.6666667" data-w="1080" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNr9f62KEIPM7myNgtzZjaauriaXjAMEwvjKQmjd0YMMR1S34Wu2qT7d7dL7QI0e4vlh407FaCnxSicg/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="jpeg"></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(236, 233, 224);padding: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-family: PingFangSC-ultralight;box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">追星</strong></p></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="text-align: left;font-size: 14px;letter-spacing: 4px;padding: 0px 20px;box-sizing: border-box;" powered-by="xiumi.us"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">疫情期间，在家无聊开始研究粉丝经济。看青春有你2的时候，作为一个从05年超女一路见证我国选秀发展史的资深老秀芬，觉得自己一定不会入坑的…没有想到研究研究着把自己也研究到坑里面去了。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">追星生活让成年人的快乐，好像变得简单了一点。好像只要是刷到爱豆的新物料就能开心半天，露出姨母般的微笑，又好像让一些乱花钱突然有了正经的理由。我跟着饭圈的各位大大，也学到了很多时尚、商务知识，甚至包括很多网站的打榜推荐算法…</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">身边熟悉我的朋友大概经常被我刷屏介绍我的宝藏女孩刘雨昕，也被我“强迫地”听了刘雨昕的歌，看了刘雨昕的表演，买了刘雨昕的代言。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">你要问我我最喜欢刘雨昕什么，我觉得可能是她身上那股坚持为梦想，努力不服输的精神吧。16年认识她的时候那个有实力又善良的姑娘，到20年，作为一个在娱乐圈起起伏伏了四年的小糊糊，还能够不断地去坚守初心，打磨自己的业务能力，很不容易。她的有些故事，让我想起自己在现实的摸爬滚打，还有放弃和坚持。</p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">经常有人跟我说，觉得我这些年实在过得太顺利了，很多痛苦没有经历过。我每次都一笑置之，因为我总希望自己传达出来的，是快乐的能量。但都不能代表，我没有经历过苦难。有些故事说出来，可能别人都不信，比如校园霸凌，又比如…我也不知道在经历过那些事情之后的我，是怎样成长到今天的，但是我知道这些苦难的背后，我放弃和失去的东西挺多的，所以希望雨昕能够一直坚持下去，不要放弃，保持清醒，一起走向更大的舞台吧。</p>
<p style="text-align: justify;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br></p></section></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="margin: 10px 0%;text-align: center;justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 95%;vertical-align: top;background-color: rgba(255, 255, 255, 0.73);border-width: 1px;border-radius: 0px;border-style: solid;border-color: rgb(109, 128, 82);height: auto;box-sizing: border-box;"><section style="margin: 20px 0% 10px;opacity: 0.99;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;color: rgb(88, 88, 88);font-size: 14px;line-height: 1.8;letter-spacing: 1.8px;padding: 0px 20px;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">对了，我领养的长颈鹿也回归大自然啦！本来希望有机会亲眼在大象孤儿院看到它，可惜疫情让人失去了旅行的机会。希望Kiko能够平安顺利地长大。</p></section></section><section style="margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box;"><img data-ratio="0.8009259" data-w="1080" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNr9f62KEIPM7myNgtzZjaauSuMkfjcCPTib1g2kOSiaoC8ssQalibo1ibayVQbbucAMjHibXCFuQJwZ6fA/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section><section style="margin: 20px 0% 10px;opacity: 0.99;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;color: rgb(88, 88, 88);font-size: 14px;line-height: 1.8;letter-spacing: 1.8px;padding: 0px 20px;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="margin: 10px 0%;text-align: center;justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 95%;vertical-align: top;background-color: rgba(255, 255, 255, 0.73);border-width: 1px;border-radius: 0px;border-style: solid;border-color: rgb(109, 128, 82);height: auto;box-sizing: border-box;"><section style="margin: 20px 0% 10px;opacity: 0.99;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;color: rgb(88, 88, 88);font-size: 14px;line-height: 1.8;letter-spacing: 1.8px;padding: 0px 20px;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">2021年真的希望世界和平，大家都平安喜乐。</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">希望自己能够保持钝感力，拥有情绪自由，享受失控之美，体会未来和未知的吸引力，放下包袱，成为自己。实现人生心愿清单上的多两条，向自己的职业梦想再近一点点，在忙碌工作的同时还是保有自己的生活，追求简单快乐。</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">小S帮我算命说21年会恋爱（她是多期待才会去算命…）那就期待吧，也祝福大家新年都可以找到自己的Mr.Right！</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">跟小S认识第11年啦，不知不觉公众号也6年啦，感谢小S和大家一直以来的陪伴，希望接下来的日子里，也有你们～<span style='text-align: center;font-size: 16px;color: rgb(51, 51, 51);font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'></span></p></section></section></section></section><section style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;vertical-align: bottom;width: 20%;height: auto;align-self: flex-end;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><p><br></p></section></section></section></section><p style="text-align: center;"><img class="rich_pages" data-ratio="0.4514877102199224" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNpKic4DiblDA75KduXD0EeqoZ8McBkcR5DUfib3SQ5IiaNYpzLW98Valvic1XIQmFEIbibgFwZUPNb69vcg/640?wx_fmt=png" data-type="png" data-w="773" style=""></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693331&amp;idx=1&amp;sn=ed5bcfe14210e4253c37b50dddff2db1</link>
      <pubDate>Sun, 10 Jan 2021 06:11:45 -0800</pubDate>
      <guid isPermaLink="false">335779</guid>
    </item>
    <item>
      <title>明天下午，3.8女生节线上特别活动，GirlsWhoCode 与 GDG Beijing 、Women Techmaker 联合呈现，轻松愉快又有料的直播，与你一起讨论特殊时期的特殊成长！</title>
      <description><![CDATA[<div id="js_content">
                                <p class="share_notice" lang="en" id="js_image_desc">
                    <script nonce="1988921409">
                    if (!window.__second_open__) {
                        document.getElementById('js_image_desc').innerHTML = "明天下午，3.8女生节线上特别活动，GirlsWhoCode 与 GDG Beijing 、Women Techmaker 联合呈现，轻松愉快又有料的直播，与你一起讨论特殊时期的特殊成长！".replace(/\r/g,"").replace(/\n/g,"<br>").replace(/\s/g,"&nbsp;");
                    }
                    </script>
                </p>
                                <div id="img_list" class="share_media">
                                        <img src="https://mmbiz.qlogo.cn/mmbiz_jpg/qrpuZA2scNqzAyibFhnruJyoibXgeYyh0WHd0w7rBuqZQyEXT11KL2FLXdovUfCd16wicpos9poOibxNia2sg4RBGNw/0?wx_fmt=jpeg" alt="">
                                    </div>
            </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693309&amp;idx=1&amp;sn=5176b7ec60617fe7f67bee8f51ff49af</link>
      <pubDate>Fri, 06 Mar 2020 16:25:25 -0800</pubDate>
      <guid isPermaLink="false">257796</guid>
    </item>
    <item>
      <title>回归 | 与过去的你一起拥抱未来！</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content" style="visibility: hidden;">
                    

                    

                    
                    
                    <section style="background-color: rgb(255, 254, 250);box-sizing: border-box;font-style: normal;font-weight: 400;text-align: justify;font-size: 16px;"><section style="display: block;box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 0px 0%;text-align: center;justify-content: center;box-sizing: border-box;"><section style='display: inline-block;width: 100%;vertical-align: top;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/SIDknOTgwKmFrKiaBZzaclNPjiccAadiclacrCcKrQGkKwuCQKF6YFic0hIo4RG9ib8xTrI4RJic3VeLovm5QiafJq6TQ/640?wx_fmt=png");background-position: 0% 0%;background-repeat: repeat;background-size: 3.2538%;background-attachment: scroll;padding: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;height: auto;box-sizing: border-box;'><section style="margin: 10px 0% -20px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 18px;color: rgb(107, 160, 172);box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br></p></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="rich_pages" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740" style=""></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br></p></section><section style="margin: 0px 0%;transform: translate3d(-4px, 0px, 0px);-webkit-transform: translate3d(-4px, 0px, 0px);-moz-transform: translate3d(-4px, 0px, 0px);-o-transform: translate3d(-4px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 77%;vertical-align: top;height: auto;background-color: rgb(253, 230, 162);box-sizing: border-box;"><section style="justify-content: center;margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 90%;vertical-align: top;height: auto;background-color: rgba(0, 0, 0, 0);box-sizing: border-box;"><section style="justify-content: center;margin: 8px 0% -53px;transform: translate3d(1px, 0px, 0px);-webkit-transform: translate3d(1px, 0px, 0px);-moz-transform: translate3d(1px, 0px, 0px);-o-transform: translate3d(1px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 25px;color: rgba(252, 215, 215, 0.97);letter-spacing: 2px;text-shadow: rgb(255, 255, 255) 2px -2px;line-height: 1.7;box-sizing: border-box;"><p style="text-align: center;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">MARCH</strong></p></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="margin: 3px 0% 5px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 80%;border-style: solid;border-width: 1px;border-radius: 0px;border-color: rgb(255, 249, 206);height: auto;box-sizing: border-box;"><img data-ratio="0.6666667" data-w="1080" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNowugtYSYhFx0TKQFNwu2GDvHib7vuZQiaGlO4zdJfGMjTaHsJAhPtI2AX2zSxfGWURMB9pVQsSibd3g/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="jpeg"></section></section><section style="text-align: right;margin: 3px 0% 7px;transform: translate3d(-3px, 0px, 0px);-webkit-transform: translate3d(-3px, 0px, 0px);-moz-transform: translate3d(-3px, 0px, 0px);-o-transform: translate3d(-3px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 15%;box-shadow: rgb(252, 247, 241) 0px 0px 0px;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 28.1 15.7" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><circle style="box-sizing: border-box;" cx="8.2" cy="8.5" r="2.2" fill="rgb(255,255,255)"></circle><circle style="box-sizing: border-box;" cx="19.5" cy="8.5" r="2.2" fill="rgb(255,255,255)"></circle><circle style="box-sizing: border-box;" cx="13.9" cy="8.5" r="2.2" fill="rgb(255,255,255)"></circle></g></svg></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">2020年3月的第一个工作日，沉（tiao）寂（piao）已久的我们复出啦！大家都复学复工了嘛？</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">在消失的这一年里，我们经常想起大家：会想生活中有哪些值得分享的呢，会想该给大家呈现什么样的内容呢，会想该如何运用好这个平台与大家共同成长呢？思考着这些问题，我们采访了<span style="color: rgba(62, 62, 62, 0.82);background-color: rgb(255, 249, 206);box-sizing: border-box;">14位程序媛，问了7个有趣的问题</span>，分享给大家。希望这样的对话也能勾起大家的一些回忆，带给大家一些启发~</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="color: rgb(178, 178, 178);">全文总计4736个字，阅读大概需要8分钟。</span></p></section></section><section style="text-align: right;margin: 0px 0% -8px;transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 13px;height: 13px;vertical-align: top;overflow: hidden;background-color: rgba(0, 0, 0, 0);border-radius: 100%;border-width: 2px;border-style: solid;border-color: rgb(253, 230, 162);box-shadow: rgb(189, 189, 189) 0px 0px 0px;letter-spacing: 1px;padding: 1px;box-sizing: border-box;line-height: 0;"><br></section></section></section></section><section style="margin: 20px 0% 10px;text-align: left;font-size: 5px;transform: translate3d(44px, 0px, 0px);-webkit-transform: translate3d(44px, 0px, 0px);-moz-transform: translate3d(44px, 0px, 0px);-o-transform: translate3d(44px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(255, 230, 230);display: inline-block;text-align: center;width: 2.5em;height: 2.5em;line-height: 2.5em;border-radius: 100%;margin-left: auto;margin-right: auto;font-size: 9px;color: rgb(255, 255, 240);box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section><section style="margin: 0px 0% 20px;font-size: 5px;transform: translate3d(-15px, 0px, 0px);-webkit-transform: translate3d(-15px, 0px, 0px);-moz-transform: translate3d(-15px, 0px, 0px);-o-transform: translate3d(-15px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgb(199, 220, 178);display: inline-block;width: 2.5em;height: 2.5em;line-height: 2.5em;border-radius: 100%;margin-left: auto;margin-right: auto;font-size: 9px;color: rgb(255, 255, 240);box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section><section style="margin: 0px 0% -60px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;vertical-align: top;width: 33%;box-sizing: border-box;"><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="text-align: right;justify-content: flex-end;font-size: 13px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 20px;height: 20px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 0px;border-style: none;border-color: rgb(62, 62, 62);background-color: rgb(255, 241, 207);box-sizing: border-box;"><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section><section style="display: inline-block;vertical-align: top;width: 33%;box-sizing: border-box;"><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section><section style="display: inline-block;vertical-align: top;width: 33%;box-sizing: border-box;"><section style="margin: 0px 0%;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 35px;height: 35px;vertical-align: top;overflow: hidden;border-radius: 100%;border-width: 2px;border-style: dotted;border-color: rgba(252, 215, 215, 0.97);background-color: rgba(216, 67, 61, 0);box-sizing: border-box;"><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="margin: 20px 0% 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="color: rgba(252, 215, 215, 0.97);font-size: 30px;line-height: 1;letter-spacing: 5px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">最认可自己的一点</strong></p></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">在过去的一年中，相信每一个你都有很多自己的小确幸，让我们从最近丧丧的心情中抽出来，分享一些最值得认可，给自己鼓鼓掌的事儿～</p>
<section><qqmusic class="js_editor_qqmusic qqmusic_iframe js_uneditable custom_select_card" musicid="1439506" mid="002rdQQx0vliRT" albumurl="https://y.gtimg.cn/music/photo_new/T002R68x68M000000F63Uj3alZTi.jpg" audiourl="http://isure.stream.qqmusic.qq.com/C200002x82Lr0Rk7GK.m4a?guid=2000001731&amp;vkey=6F7753A1E6CA0E8C7B980B2B665ACD0FCBE333BB6BB4423E7C6EF74907C9BC32B0921A5231B5C1B474219332CDD933F04A92E38F3531265C&amp;uin=&amp;fromtag=50" music_name="When You Believe" singer="Mariah Carey;Whitney Houston - The Essential Mariah Carey" play_length="272" src="https://localhost:3000/cgi-bin/readtemplate?t=tmpl/qqmusic_tmpl&amp;singer=Mariah%20Carey%3BWhitney%20Houston%20-%20The%20Essential%20Mariah%20Carey&amp;music_name=When%20You%20Believe&amp;albumurl=https%3A%2F%2Fy.gtimg.cn%2Fmusic%2Fphoto_new%2FT002R68x68M000000F63Uj3alZTi.jpg&amp;musictype=1" musictype="1" otherid="002rdQQx0vliRT" albumid="000F63Uj3alZTi" jumpurlkey="" data-pluginname="insertaudio"></qqmusic></section></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: center;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="text-decoration: underline;box-sizing: border-box;"><strong style="box-sizing: border-box;"><span style="text-decoration: underline;color: rgba(62, 62, 62, 0.82);background-color: rgb(255, 249, 206);box-sizing: border-box;">强大的执行力</span></strong></span></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">很多人按自己的计划在重要的时间节点完成了阶段性目标，Naomi上了10节私教课瘦了5kg；搬入了自己的房子，四处充满着自己装修的小巧思；雅可比考到了驾照，驾驶技能终于在城市点亮；Lu顺利博士毕业了；Summer也是越来越上镜了，美照我就私藏不分享了嘻嘻～</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: center;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="text-decoration: underline;box-sizing: border-box;"><strong style="box-sizing: border-box;"><span style="text-decoration: underline;color: rgba(62, 62, 62, 0.82);background-color: rgb(255, 249, 206);box-sizing: border-box;">良好的生活习惯</span></strong></span></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">Ashley终于克服了多年的晚睡强迫症，养成了早睡早起的作息习惯，精力值和效率都有神奇的提高；也甩掉了天天吃外卖的饮食状况，大多数时间自己做饭，省钱又健康，而且做饭的过程还意外成为了每天的减压时刻。</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: center;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="text-decoration: underline;box-sizing: border-box;"><strong style="box-sizing: border-box;"><span style="text-decoration: underline;color: rgba(62, 62, 62, 0.82);background-color: rgb(255, 249, 206);box-sizing: border-box;">与家人的时光<br style="box-sizing: border-box;"></span></strong></span></p>
<p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">与家人的时光总是非常珍贵的，许多人带爸妈一起出国旅游，已经感觉自己是他们的依靠啦。Yangyang带爸妈出了远门，去了加拿大的班夫和贾斯珀国家公园；陪他们爬了雷尼尔山，看到了土拨鼠和漫山野花；带他们出海钓鱼，看他们开心得像小孩；拉他们一起跳Just Dance，妈妈说比广场舞难；带他们参加年轻一辈的公园野餐，让他们不用担心我在这里会孤单。爸妈说：“以前想都没想过我们可以走这么远，看这样的美景。你小时候我们带你出去玩的不多，也走的不远，倒是<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">现在你来带我们看更大的世界了。</span>” 值得为自己骄傲一会儿。</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: center;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="text-decoration: underline;box-sizing: border-box;"><strong style="box-sizing: border-box;"><span style="text-decoration: underline;color: rgba(62, 62, 62, 0.82);background-color: rgb(255, 249, 206);box-sizing: border-box;">工作上的成长<br style="box-sizing: border-box;"></span></strong></span></p>
<p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">很多人在职场上迅速成长。Jiajia作ICCV AC处理了30篇论文，每篇论文都仔细看过，还读了buddy AC那边的30篇。认真写了3遍meta review的comments，有个超级靠谱和专业的buddy AC真是鞭策，收获良多！在这个AI投稿满天飞，AC &amp; reviewer都严重超负荷的时代，<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">做好本分也是对别人工作和自己的褒奖。</span></p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">有人在学业上不断进步。不想当栋梁的篱笆在韩国读硕博，因为韩语英语都不十分精通，每次做presentation，defense都特别心累，被老师说不够格， 好在2019年出了好几篇论文，感觉自己的能力得到了认可，老师们态度也180度转换。觉得自己好了不起。</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: center;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="text-decoration: underline;box-sizing: border-box;"><strong style="box-sizing: border-box;"><span style="text-decoration: underline;color: rgba(62, 62, 62, 0.82);background-color: rgb(255, 249, 206);box-sizing: border-box;">限制与突破</span></strong></span></p>
<p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">因为之前的成绩，Redhouse被限制在对自己过高的期待中，总希望做出更加了不起的工作来，但其实每件事成功也有运气的成分在。幸运的是，在2019年，<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">她放下以前的包袱，不对结果做过高的期待，只想应该投入哪些有意思和有价值的方向，然后去做，结果却出乎意料的好。</span></p></section></section><section style="text-align: right;margin: 0px 0% -8px;transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 13px;height: 13px;vertical-align: top;overflow: hidden;background-color: rgba(0, 0, 0, 0);border-radius: 100%;border-width: 2px;border-style: solid;border-color: rgb(253, 230, 162);box-shadow: rgb(189, 189, 189) 0px 0px 0px;letter-spacing: 1px;padding: 1px;box-sizing: border-box;line-height: 0;"><br></section></section></section></section><section style="margin: 0px 0%;transform: translate3d(-4px, 0px, 0px);-webkit-transform: translate3d(-4px, 0px, 0px);-moz-transform: translate3d(-4px, 0px, 0px);-o-transform: translate3d(-4px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 15px;line-height: 1.8;letter-spacing: 3px;color: rgb(253, 230, 162);padding: 0px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">●●●●●●●</p></section></section><section style="margin: 20px 0% 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="color: rgba(252, 215, 215, 0.97);font-size: 30px;line-height: 1;letter-spacing: 5px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">最沮丧的一件事</strong></p></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">如果问大家最近最丧的事儿是什么，大概大家的答案都是相似的——新型冠状病毒，希望一切能够很快好起来！</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">在我们收集这道题的答案的时候，武汉还没有封城，病毒也还没有引起足够的重视，在那些令现在的我们无比怀念的普通日常生活里，那些令人沮丧的：生病，胖了，过敏了，晒伤了；压力大失眠；跟父母相隔太远，担心却不知道怎么解决；期待的旅行取消；工作上自己的想法和努力被忽视；等等。很多的小细节，相信都让大家感同身受。</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">曾在外企做研发工作，几年前跳槽开始带团队做产品的Redhouse跟我们分享的故事，是印象最深的一个。</p></section></section><section style="text-align: justify;line-height: 1;box-sizing: border-box;" powered-by="xiumi.us"><section style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;line-height: normal;"><span style="box-sizing: border-box;font-size: 12px;">这一年是对人的信任崩溃的一年。跳槽之后，经过一段时间战战兢兢的适应，这一年里在多件事情上开始认识到中间的利害关系。自己一开始是忍让，但忍的很辛苦，每天晚上即使是在家陪孩子做作业，脑袋也经常被负面情绪所占据。同僚之间明争暗斗其实也不难理解，毕竟每个人都希望自己的利益最大化。比较让我想不通的是领导也似乎没有意识需要管理和平衡。也许是因为自己并不重要吧。人很奇怪，一旦有事情想不通，觉得无解的时候，就会陷入反复思量的状态。休息不好，精神也会更差。由于没有斗争经验，也没有得到支持，我一度会在脑海里想象出另一个自己，来鼓励自己。无论别人如何无视或贬低我的工作，我靠另一个自己来温暖自己。但这其实是很无奈的一种自我保护吧。</span></section><section style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;line-height: normal;"><span style="box-sizing: border-box;font-size: 12px;">让我没继续滑下去的最后一棵稻草，是敞开心扉。我在无法再忍耐的时候，和一位比较投缘的同僚倾诉了我大半年的难过与郁闷。我当时也做了一个决定，不再忍，不再依靠领导，靠自己去反击。我那时候应该是气到了极点，或者说被逼到墙角，突然生出最后一股破罐子破摔的勇气，我用暴力沟通的方式进行了反击。那段时间，我也时常会跟他讲一讲我的想法和做法。很感谢他愿意听我讲，也并没有拿大道理来劝阻我。在我发泄的差不多的时候，他才跟我说，我的本性并非喜欢争斗与人为难的人。那之后，他在工作上也给予了我信任和帮助，我渐渐的找到了一些支撑。 </span></section><section style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;line-height: normal;"><span style="box-sizing: border-box;font-size: 12px;">有一天，老公跟我讲起囚徒困境。我们把都不供认罪行叫做合作，把供认叫做背叛，背叛也是非常合乎人性的选择，甚至如果这种困境只发生一次，那么选择背叛的收益更大。但生活中，很多博弈关系不止一次，经过理论和实验验证，最好的策略是以牙还牙且报复之后不记仇。也就是说，第一次选择合作，但如果某一个人选择背叛，再次与其处于囚徒困境时，我方应该选择背叛，但第三次，可以不计前嫌选择合作。这样的策略长期最为有效。一年的暗黑经历，让我明白，老好人也是很难持久的。持续受损是很难消化的，选择反击一次，既可以得到心理上的释放和平衡，也可以给背叛的一方一个警示，有利于下一次的合作。</span></section><section style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;line-height: normal;"><span style="box-sizing: border-box;font-size: 12px;">我很开心自己没有继续黑化下去，让自己成为孤立的刺猬。相反，抱着首先要合作的态度，主动和友好的同僚交流合作，也和领导做了一些沟通，了解他们的难处，请他们做力所能及的帮助。我把之前过多关注公司目标，改换成设定了自己的小目标，每天又重新有了动力去克服困难，实现梦想。当我把注意力从讨厌的人身上移开，我也发现很多喜欢和支持自己的人。我很感谢那位帮助我的同僚，是他让我没有失去本心，在提高了处理矛盾的经验值的同时，仍然是那个目光温暖的自己。</span></section><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="text-align: right;margin: 0px 0% -8px;transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);justify-content: flex-end;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 13px;height: 13px;vertical-align: top;overflow: hidden;background-color: rgba(0, 0, 0, 0);border-radius: 100%;border-width: 2px;border-style: solid;border-color: rgb(253, 230, 162);box-shadow: rgb(189, 189, 189) 0px 0px 0px;letter-spacing: 1px;padding: 1px;box-sizing: border-box;line-height: 0;"><br></section></section></section></section><section style="margin: 0px 0%;transform: translate3d(-4px, 0px, 0px);-webkit-transform: translate3d(-4px, 0px, 0px);-moz-transform: translate3d(-4px, 0px, 0px);-o-transform: translate3d(-4px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 15px;line-height: 1.8;letter-spacing: 3px;color: rgb(253, 230, 162);padding: 0px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">●●●●●●●</p></section></section><section style="margin: 0px 0%;box-sizing: border-box;" powered-by="xiumi.us"><section style="color: rgba(252, 215, 215, 0.97);font-size: 30px;line-height: 1;letter-spacing: 5px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">最喜欢的一本书</strong></p></section></section><section style="margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 80%;height: auto;border-style: dashed;border-width: 0px;border-radius: 0px;border-color: rgb(62, 62, 62);box-sizing: border-box;"><img data-ratio="1.7814815" data-w="1080" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNowugtYSYhFx0TKQFNwu2GDN9FtricpT4vSicCaIWZkLELudSslPwdSiaoaLw19cJPficYyf6lgUhReSQ/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="jpeg"></section></section><section style="margin: 0px 0%;transform: translate3d(-4px, 0px, 0px);-webkit-transform: translate3d(-4px, 0px, 0px);-moz-transform: translate3d(-4px, 0px, 0px);-o-transform: translate3d(-4px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 15px;line-height: 1.8;letter-spacing: 3px;color: rgb(253, 230, 162);padding: 0px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">●●●●●●●</p></section></section><section style="margin: 20px 0% 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="color: rgba(252, 215, 215, 0.97);font-size: 30px;line-height: 1;letter-spacing: 5px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">最爱的一首歌</strong></p></section></section><section style="margin: 0px 0% 15px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 14px;color: rgb(94, 93, 93);line-height: 2.5;letter-spacing: 2px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="background-color: rgb(255, 249, 206);box-sizing: border-box;">扫描下方二维码获取QQ音乐歌单！</span></p>
<p style="text-align: center;"><img class="rich_pages" data-ratio="0.9932885906040269" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNowugtYSYhFx0TKQFNwu2GDfYvCJBtOSA0wjkwLESU0uWg7ImEPr1geCu0JP5QJY9adS7GwOmzwUg/640?wx_fmt=jpeg" data-type="jpeg" data-w="149" style="width: 81px;height: 80px;"><span style='background-color: rgb(255, 249, 206);font-family: mp-quote, -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'></span></p></section></section><section style="margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 80%;height: auto;box-sizing: border-box;"><img data-ratio="0.7277778" data-w="1080" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNowugtYSYhFx0TKQFNwu2GDPiaAaW471IT5B85xhib4tVYiaK4J2PGIuNFAuuaqVO6hf3br8saNVVjBQ/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section><section style="margin: 0px 0%;transform: translate3d(-4px, 0px, 0px);-webkit-transform: translate3d(-4px, 0px, 0px);-moz-transform: translate3d(-4px, 0px, 0px);-o-transform: translate3d(-4px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 15px;line-height: 1.8;letter-spacing: 3px;color: rgb(253, 230, 162);padding: 0px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">●●●●●●●</p></section></section><section style="margin: 20px 0% 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="color: rgba(252, 215, 215, 0.97);font-size: 30px;line-height: 1;letter-spacing: 5px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">新</strong><strong style="box-sizing: border-box;">get</strong><strong style="box-sizing: border-box;">的技能点</strong></p></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 14px;color: rgba(62, 62, 62, 0.82);text-align: left;padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">2019年大家学会了很多生活中增加幸福感的技能，比如带娃、养柴犬、驾驶、健身、跳韩舞、手冲咖啡、咖啡拉花和单手抱吉他等等。也有另一些小伙伴“没有新学习什么具体的技能，但是有一些不足为外人道的内心成长。”<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">内心上的改变不仅能帮助我们提高执行力，还能帮助我们积累对生活的掌控感和对自己的认可度。</span></p>
<section><qqmusic class="js_editor_qqmusic qqmusic_iframe js_uneditable custom_select_card" musicid="233805163" mid="00130L8j1JSjQI" albumurl="https://y.gtimg.cn/music/photo_new/T002R68x68M000002BK2WZ4XQc46.jpg" audiourl="http://isure.stream.qqmusic.qq.com/C20000130L8j1JSjQI.m4a?guid=2000001731&amp;vkey=E8474FFBE609A4431350BF01910972F910F02AC0EC0E8A2994487AC2391B2B040EB28F3973495BECB8D1721D0112E3184FD748D2471248A3&amp;uin=&amp;fromtag=50" music_name="Speechless (Full)" singer="Naomi Scott - Aladdin (Svenskt Original Soundtrack)" play_length="208" src="https://localhost:3000/cgi-bin/readtemplate?t=tmpl/qqmusic_tmpl&amp;singer=Naomi%20Scott%20-%20Aladdin%20(Svenskt%20Original%20Soundtrack)&amp;music_name=Speechless%20(Full)&amp;albumurl=https%3A%2F%2Fy.gtimg.cn%2Fmusic%2Fphoto_new%2FT002R68x68M000002BK2WZ4XQc46.jpg&amp;musictype=1" musictype="1" otherid="00130L8j1JSjQI" albumid="002BK2WZ4XQc46" jumpurlkey="" data-pluginname="insertaudio"></qqmusic></section></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">最喜人的分享是一位可爱的姑娘在2019年成功战胜了抑郁症。</p>
<p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-size: 12px;box-sizing: border-box;">得病的时候就是容易很丧，感觉自己很差劲，经常想哭，上班躲在会议室里哭，晚上回家被窝里哭… 去北医六院看过两次，第一次药有副作用，耽误了一些时间，后来换了个医生，换药后感觉像获得重生一样！每天的阳光都是美好的！整个过程还是比较艰难的，甚至吃药的时候也经常想放弃治疗，幸好结局还不错~ 想了很多很多问题，也看了很多书，对生命有了新的认知。</span><span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);font-size: 12px;box-sizing: border-box;">新的一年希望把每一天都过得开开心心的，不辜负爱和阳光！</span></p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">Redhouse 也带来了最细腻的一个分享。</p>
<p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-size: 12px;box-sizing: border-box;">我开始能够</span><span style="font-size: 12px;color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">欣赏失控之美</span><span style="font-size: 12px;box-sizing: border-box;">。2019年之前，自己还是学霸型的人格，对自己、别人和事情有很完美的追求。这一年，让我看到学霸的盲点，人不是书本上的理想的人，而是活生生有血肉有欲望的人。事情不是说只有百分百有把握的时候才值得去做，也不是只有自己都懂的时候才能去做。太可控的目标虽然让人心里踏实，也鲜有风险，但是收益其实未必大。正是那些看似冒险难解的问题，如果可以组织资源去一步步接近和实现它，才会有意想不到的收获。这就像小孩子，不怕摔不断的调皮和探索，才能不断的成长。虽然我们已经是成年人，但在工作生活中，如果能像孩子一样，不怕失控，才会有新的成长机会。即使摔了，也不要给自己太多的心理负担，认为自己会很糗。其实就像你看到一个调皮的孩子摔倒了，爬起来继续跑，那种酷酷的感觉，成年人心里是有一种羡慕和欣赏的。</span></p></section></section></section></section><section style="margin: 0px 0%;transform: translate3d(-4px, 0px, 0px);-webkit-transform: translate3d(-4px, 0px, 0px);-moz-transform: translate3d(-4px, 0px, 0px);-o-transform: translate3d(-4px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 15px;line-height: 1.8;letter-spacing: 3px;color: rgb(253, 230, 162);padding: 0px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">●●●●●●●</p></section></section><section style="margin: 20px 0% 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="color: rgba(252, 215, 215, 0.97);font-size: 30px;line-height: 1;letter-spacing: 5px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">印象最深刻的瞬间</strong></p></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">如果问你们2019年印象最深刻的瞬间会是什么呢？也许是一次旅行中大自然的馈赠带给我们由内而外的感动，比如“在土耳其坐热气球升起来的时候”心里的感受一定是充盈又神奇吧。或者是另一个人对自己的好想让你永生珍藏，比如“男朋友帮我剥珍宝蟹（全程不用我动手）”，以及还有两位小伙伴在2019年与另一半许下了终身的约定。亦或是一件小事给了我们满满的力量。下面分享一些瞬间给大家，希望也能勾起你相似的回忆和感受：</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box;"><img data-ratio="0.6666667" data-w="1080" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNowugtYSYhFx0TKQFNwu2GDmC1D56ynaNJB2xKW7GGC2nIDyUvnbTH6KWbIzWu5hldn4VNCjGrRlw/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">在火山口湖国家公园看到太阳从山顶升起的瞬间。四月的山上还有厚厚的积雪，山顶的风特别大，吹得人直往后退，我们被冻得缩在一起。从起床到看到日出大概有三个小时，冻着等着，所有的辛苦在最后看到太阳出现的那一刻都被冲散了，非常值得。</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 80%;height: auto;box-sizing: border-box;"><img data-ratio="0.6833977" data-w="518" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNowugtYSYhFx0TKQFNwu2GDV6zhibviam8ibIfC68NZMIicDXib7Dz8OWBEsxPXzP1jhgTJ1DVDV7dOPFw/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">9月份我在自己的公众上给8年前的“网红面经”发了个续篇，很多同事朋友深有共鸣、写了长篇的评论给我。在刷屏的朋友圈里，我看见一位多年没有联系的大学同学转发时只写了四个字，“<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">成为自己</span>”——这就是我一直在追索的，和希望大家能从推文里体会的事，能被理解和被这样简单有力地提炼，真是太好了。</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="transform: rotate(0deg);-webkit-transform: rotate(0deg);-moz-transform: rotate(0deg);-o-transform: rotate(0deg);box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="vertical-align: middle;display: inline-block;line-height: 0;width: 60%;height: auto;box-sizing: border-box;"><img data-ratio="0.9979167" data-w="480" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNowugtYSYhFx0TKQFNwu2GD1AoUtgEjyhEAcMFRbDiaIiaMFLIOWWpvdgepiaBI6SWpFN1hJUkaReKvA/640?wx_fmt=gif" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="gif"></section></section></section><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">面临重要的职业选择时，第一次在办公室掉眼泪，不是嚎啕大哭，而是像个大人一样冷静的说出自己的决定时，眼角滑下的一滴泪。</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">有一天我心情低落，不知道自己是否还有必要在现在的公司继续下去，我问上小学的女儿：“你在学校有没有讨厌的同学？”她说：“有啊。”那阵子，她也会跟我讲，班里有个女生很霸道，习惯抢别人的东西，可是同学们都好像很怕她，甚至她的好朋友被欺负多次之后，却主动和那个霸道的女生做起好朋友，疏远了她。我又问她：“那你想过转学吗？”她说：“没有呀，我想过让那个讨厌的同学转学。”那一瞬间，我被撞到了一样，是呀，当我们碰到不喜欢的人和事，转学（或着跳槽）可能是自己可控的一种方式；可是这里也有不少你喜欢的人和事，那如果那个讨厌的人能离开岂不是更好？小孩子天真直率的想法，却是很有力量的答案，在黑暗中给了我一道质朴的光。</p></section></section></section></section><section style="margin: 0px 0%;transform: translate3d(-4px, 0px, 0px);-webkit-transform: translate3d(-4px, 0px, 0px);-moz-transform: translate3d(-4px, 0px, 0px);-o-transform: translate3d(-4px, 0px, 0px);box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 15px;line-height: 1.8;letter-spacing: 3px;color: rgb(253, 230, 162);padding: 0px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">●●●●●●●</p></section></section><section style="margin: 20px 0% 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="color: rgba(252, 215, 215, 0.97);font-size: 30px;line-height: 1;letter-spacing: 5px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><strong style="box-sizing: border-box;">送给自己的寄语</strong></p></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">这个问题反映出了大家对于自身成长的关注和期待。有四个小伙伴都表示2020想让脾气变好一点，平常心多一点，焦虑少一点，向<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">情绪自由</span>而努力。另外一些小伙伴期待着自己能更强大，去面对生活中的负能量、挫折与危机。下面其中几个她们的宣言：</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;line-height: normal;"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 0px 0% 15px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 14px;color: rgb(94, 93, 93);line-height: 2.5;letter-spacing: 2px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="background-color: rgb(255, 249, 206);box-sizing: border-box;">给自己创造新的挑战，并且最终达成它！</span></p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 0px 0% 15px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 14px;color: rgb(94, 93, 93);text-align: left;line-height: 1.6;letter-spacing: 2px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="background-color: rgb(255, 249, 206);box-sizing: border-box;">希望自己特立但不独行，有能力爱自己，有余力爱别人！</span></p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">新的一年里，我希望自己能够有实用的研究成果，能够多花些时间陪伴家人，能够多多锻炼身体，能够在闲暇时光读几本好书，会三五好友，能够做到，<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">不惑、不忧、不惧</span>，在今后的日子里，对自身和对生活的双向探索，能让自己变得自信，充实，充满安全感。<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">因为内心的勇敢而变得自由，不会因为时间流逝感到焦虑。</span></p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">去年冬天曾经给自己一个期限，一年发不出顶会我就打消自己继续读博科研的想法。但现在只有一篇三作的文章，按道理应该承认自己没有科研天分要果断放弃了。却被自己一步步的规划悄无声息地推到了申请季。还没有正式的offer, 却总是畅想以后在一个陌生的地方生活，觉得很兴奋，把所有的顾虑都抛在脑后。<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">未来和未知对我的吸引力太大了，让我忘记自己的局限性，厚着脸皮想去看看和试试。</span>我相信这一年会给我以后的人生带来很大很大的改变。</p></section></section></section></section><section style="margin: 0px 0% -10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 40%;vertical-align: top;border-left: 5px solid rgb(253, 230, 162);border-bottom-left-radius: 0px;border-right: 5px solid rgb(253, 230, 162);border-top-right-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section style="text-align: justify;line-height: 2;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section style="justify-content: center;box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 82%;vertical-align: top;background-color: rgb(255, 255, 255);border-style: dashed;border-width: 2px;border-radius: 5px;border-color: rgb(253, 230, 162);padding: 10px;height: auto;overflow: hidden;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: justify;font-size: 14px;color: rgba(62, 62, 62, 0.82);padding: 0px 10px;line-height: 1.8;box-sizing: border-box;"><p style="text-align: left;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">2019年我平均每天睡眠时间差不多达到了9小时，玩了前大半年，后几个月的工作也是轻松愉快。其实我知道，这是在享受2017，2018年的劳动成果，<span style="color: rgb(94, 93, 93);background-color: rgb(255, 249, 206);box-sizing: border-box;">我享受得很开心，心安理得</span>。2020年最大的目标是坚持锻炼，前半年也是享受玩耍基调，后半年计划在工作上努力一下，2021年的目标是升职。</p></section></section></section></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 100%;vertical-align: top;box-sizing: border-box;"><section style="margin: 0px 0% 15px;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 14px;color: rgb(94, 93, 93);line-height: 2.5;letter-spacing: 2px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="background-color: rgb(255, 249, 206);text-align: left;box-sizing: border-box;"><span style="color: rgb(253, 230, 162);font-size: 15px;letter-spacing: 3px;text-align: center;background-color: rgb(255, 254, 250);">●●●●●●●</span></span></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="background-color: rgb(255, 249, 206);text-align: left;box-sizing: border-box;">2020已经过去2个月了，大家有在向着自己期待的方向努力么？看完这些的你，愿意分享你的 resolution 给我们吗？愿疫情早日过去，让我们恢复到过去精彩的日常生活。期待我们都有一个难忘又不悔的 2020！</span></p></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 18%;vertical-align: top;height: auto;box-sizing: border-box;"><section style="margin: 10px 0% 20px;font-size: 10px;box-sizing: border-box;" powered-by="xiumi.us"><section style="background-color: rgba(252, 226, 131, 0.53);height: 3px;box-sizing: border-box;line-height: 0;"><br></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 94%;vertical-align: top;box-sizing: border-box;"><section style="margin: 10px 0%;box-sizing: border-box;" powered-by="xiumi.us"><section style="font-size: 12px;line-height: 2.3;color: rgb(160, 160, 160);letter-spacing: 3px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">特别感谢 </p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">Ashley、Bibi、不想当栋梁的篱笆、CY、HQ、Jiajia、Lu、Naomi、Redhouse、Summer、小烦羊、雅可比、Yangyang、Yuxuan</p></section></section><section style="text-align: justify;font-size: 12px;line-height: 2.5;letter-spacing: 1px;box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><span style="color: rgb(160, 160, 160);box-sizing: border-box;">- </span> <span style="color: rgb(255, 255, 255);background-color: rgb(255, 238, 186);box-sizing: border-box;"> E N D </span>  <span style="color: rgb(160, 160, 160);box-sizing: border-box;">-</span><br style="box-sizing: border-box;"></p></section><section style="text-align: justify;box-sizing: border-box;" powered-by="xiumi.us"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="display: inline-block;width: 100%;vertical-align: top;box-sizing: border-box;"><section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box;"><img data-ratio="0.1384892" data-w="556" src="https://mmbiz.qpic.cn/mmbiz_png/OaWJibLCM5zLq03Kib7r9IJCmFgyZYANH72s9sgAGY6sFicMibJpQPKic2AxYDrh7w4LWjciaJtfLIN6ppN07vAMeiafg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="png"></section></section><p style="text-align: center;"><img class="rich_pages" data-ratio="0.4514877102199224" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNpKic4DiblDA75KduXD0EeqoZ8McBkcR5DUfib3SQ5IiaNYpzLW98Valvic1XIQmFEIbibgFwZUPNb69vcg/640?wx_fmt=png" data-type="png" data-w="773" style=""></p>
<section style="margin: 10px 0% 0px;box-sizing: border-box;" powered-by="xiumi.us"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;box-sizing: border-box;"><br></section><br></section></section></section></section></section></section></section></section></section>
</div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693307&amp;idx=1&amp;sn=1e67ccb5999d915fe0ca5ccbbe143c77</link>
      <pubDate>Mon, 02 Mar 2020 00:01:25 -0800</pubDate>
      <guid isPermaLink="false">255261</guid>
    </item>
    <item>
      <title>日常 | 后厂村程序媛相亲实录</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740" style=""></p>
<section style="box-sizing: border-box;font-size: 16px;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px;border-radius: 0px;border-color: rgb(225, 238, 226);background-color: rgba(191, 224, 216, 0.58);box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;vertical-align: bottom;width: 18%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 40px;height: 40px;vertical-align: top;overflow: hidden;background-color: rgb(147, 195, 186);border-bottom: 0px none rgb(62, 62, 62);border-bottom-right-radius: 100%;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 3px;margin-right: 0%;margin-left: 0%;transform: translate3d(4px, 0px, 0px);box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 60%;box-sizing: border-box;"><svg class="svg" xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 64.7 40.8" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><path d="M36.2,27c0,0.2,0,0.4,0,0.6l0,0c0.5,7.4,6.7,13.3,14.2,13.3c7.9,0,14.3-6.4,14.3-14.3c0-7.2-5.3-13.1-12.2-14.1   C53,8.9,54.2,4,57,0c0,0-19,3-20.7,24.7c-0.1,0.6-0.1,1.2-0.1,1.8C36.2,26.7,36.2,26.9,36.2,27z" fill="rgb(225,238,226)" style="box-sizing: border-box;"></path><path d="M0,27c0,0.2,0,0.4,0,0.6l0,0C0.6,35,6.7,40.8,14.3,40.8c7.9,0,14.3-6.4,14.3-14.3c0-7.2-5.3-13.1-12.2-14.1   C16.8,8.9,18,4,20.8,0c0,0-19,3-20.7,24.7C0,25.3,0,26,0,26.6C0,26.7,0,26.9,0,27z" fill="rgb(225,238,226)" style="box-sizing: border-box;"></path></g></svg></section></section></section></section></section></section></section><section style="display: inline-block;vertical-align: bottom;width: 82%;padding-top: 5px;padding-right: 5px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="transform: translate3d(-10px, 0px, 0px);-webkit-transform: translate3d(-10px, 0px, 0px);-moz-transform: translate3d(-10px, 0px, 0px);-o-transform: translate3d(-10px, 0px, 0px);box-sizing: border-box;"><section style="color: rgb(82, 126, 118);font-size: 14px;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">后厂村程序媛相亲实录 </p></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;padding: 10px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-59516-1550217541511" style="font-size: 15px;color: rgb(46, 101, 91);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="font-family: Optima-Regular, PingFangTC-light;box-sizing: border-box;">随着身边的程序媛们年纪渐长，大家都逐渐进入了被催婚的行列，今天来跟朋友们分享几个真实的相亲小故事，博大家一笑，让繁忙的周一有个愉快的开头，也祝福各位姑娘们早日找到自己的Mr.Right !</span></p></section></section></section></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(123, 172, 163);border-width: 4px;border-radius: 8px;border-style: solid;border-color: rgb(123, 172, 163);overflow: hidden;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 0.5em;margin-bottom: 0.5em;text-align: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;box-sizing: border-box;"><section style="width: 6px;height: 6px;display: inline-block;vertical-align: middle;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: middle;padding-right: 5px;padding-left: 5px;color: rgb(208, 241, 233);box-sizing: border-box;"><p style="box-sizing: border-box;">女汉子般的程序媛甲</p></section><section style="display: inline-block;vertical-align: middle;width: 6px;height: 6px;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(255, 255, 255);padding: 10px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">A男发来一张号称高清无码的相亲照片，在CCTV的总理发布会的观众席圈了一个圈，说那是他…</p>
<p style="white-space: normal;box-sizing: border-box;">甲女（这句不知道怎么接？！）</p></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 100%;box-sizing: border-box;"><svg class="svg" xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 73.9 2.7" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><path style="box-sizing: border-box;" d="M38.4,1.4c0,0.8-0.6,1.4-1.4,1.4c-0.8,0-1.4-0.6-1.4-1.4C35.7,0.6,36.3,0,37,0   C37.8,0,38.4,0.6,38.4,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M70.3,1.4c0,0.5-0.4,0.8-0.8,0.8c-0.5,0-0.8-0.4-0.8-0.8c0-0.5,0.4-0.8,0.8-0.8   C69.9,0.5,70.3,0.9,70.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M72.3,1.4c0,0.3-0.3,0.6-0.6,0.6c-0.3,0-0.6-0.3-0.6-0.6c0-0.3,0.3-0.6,0.6-0.6   C72.1,0.8,72.3,1.1,72.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M73.9,1.4c0,0.2-0.2,0.4-0.4,0.4c-0.2,0-0.4-0.2-0.4-0.4c0-0.2,0.2-0.4,0.4-0.4   C73.8,1,73.9,1.2,73.9,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M68.4,0L68.4,0c-0.8,0-1.3,0.5-1.4,1.2H56.5C52.3,1,46.4,0.8,39.7,0.8c0,0.2,0.1,0.4,0.1,0.6   c0,0.2-0.1,0.4-0.1,0.6c6.6,0,12.5-0.2,16.6-0.4H67c0.1,0.6,0.6,1.2,1.3,1.2c0,0,0,0,0.1,0c-0.3-0.4-0.6-0.8-0.6-1.4   C67.8,0.8,68,0.4,68.4,0z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M3.6,1.4c0,0.5,0.4,0.8,0.8,0.8c0.5,0,0.8-0.4,0.8-0.8c0-0.5-0.4-0.8-0.8-0.8C4,0.5,3.6,0.9,3.6,1.4   z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M1.6,1.4c0,0.3,0.3,0.6,0.6,0.6c0.3,0,0.6-0.3,0.6-0.6c0-0.3-0.3-0.6-0.6-0.6   C1.8,0.8,1.6,1.1,1.6,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M0,1.4c0,0.2,0.2,0.4,0.4,0.4c0.2,0,0.4-0.2,0.4-0.4C0.8,1.2,0.6,1,0.4,1C0.2,1,0,1.2,0,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M5.6,0L5.6,0c0.8,0,1.3,0.5,1.4,1.2h10.5C21.6,1,27.5,0.8,34.2,0.8c0,0.2-0.1,0.4-0.1,0.6   c0,0.2,0.1,0.4,0.1,0.6c-6.6,0-12.5-0.2-16.6-0.4H6.9C6.8,2.2,6.3,2.7,5.6,2.7c0,0,0,0-0.1,0c0.3-0.4,0.6-0.8,0.6-1.4   C6.1,0.8,5.9,0.4,5.6,0z" fill="rgb(128,178,163)"></path></g></svg></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="font-size: 15px;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="letter-spacing: 0px;box-sizing: border-box;">B男：我最近已经成熟到可以给弟弟妹妹制定人生规划了。我弟是学环境科学的，你知道人工智能吧？我看书，说是有个模型调调参数就行，我就跟我弟说你一定要往这个方向靠，这样钱就都自动送到你面前了。</span><br style="box-sizing: border-box;"></p>
<p style="box-sizing: border-box;">甲女（这大概是AI被黑得最惨的一次）</p></section></section></section></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 100%;box-sizing: border-box;"><svg class="svg" xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 73.9 2.7" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><path style="box-sizing: border-box;" d="M38.4,1.4c0,0.8-0.6,1.4-1.4,1.4c-0.8,0-1.4-0.6-1.4-1.4C35.7,0.6,36.3,0,37,0   C37.8,0,38.4,0.6,38.4,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M70.3,1.4c0,0.5-0.4,0.8-0.8,0.8c-0.5,0-0.8-0.4-0.8-0.8c0-0.5,0.4-0.8,0.8-0.8   C69.9,0.5,70.3,0.9,70.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M72.3,1.4c0,0.3-0.3,0.6-0.6,0.6c-0.3,0-0.6-0.3-0.6-0.6c0-0.3,0.3-0.6,0.6-0.6   C72.1,0.8,72.3,1.1,72.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M73.9,1.4c0,0.2-0.2,0.4-0.4,0.4c-0.2,0-0.4-0.2-0.4-0.4c0-0.2,0.2-0.4,0.4-0.4   C73.8,1,73.9,1.2,73.9,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M68.4,0L68.4,0c-0.8,0-1.3,0.5-1.4,1.2H56.5C52.3,1,46.4,0.8,39.7,0.8c0,0.2,0.1,0.4,0.1,0.6   c0,0.2-0.1,0.4-0.1,0.6c6.6,0,12.5-0.2,16.6-0.4H67c0.1,0.6,0.6,1.2,1.3,1.2c0,0,0,0,0.1,0c-0.3-0.4-0.6-0.8-0.6-1.4   C67.8,0.8,68,0.4,68.4,0z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M3.6,1.4c0,0.5,0.4,0.8,0.8,0.8c0.5,0,0.8-0.4,0.8-0.8c0-0.5-0.4-0.8-0.8-0.8C4,0.5,3.6,0.9,3.6,1.4   z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M1.6,1.4c0,0.3,0.3,0.6,0.6,0.6c0.3,0,0.6-0.3,0.6-0.6c0-0.3-0.3-0.6-0.6-0.6   C1.8,0.8,1.6,1.1,1.6,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M0,1.4c0,0.2,0.2,0.4,0.4,0.4c0.2,0,0.4-0.2,0.4-0.4C0.8,1.2,0.6,1,0.4,1C0.2,1,0,1.2,0,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M5.6,0L5.6,0c0.8,0,1.3,0.5,1.4,1.2h10.5C21.6,1,27.5,0.8,34.2,0.8c0,0.2-0.1,0.4-0.1,0.6   c0,0.2,0.1,0.4,0.1,0.6c-6.6,0-12.5-0.2-16.6-0.4H6.9C6.8,2.2,6.3,2.7,5.6,2.7c0,0,0,0-0.1,0c0.3-0.4,0.6-0.8,0.6-1.4   C6.1,0.8,5.9,0.4,5.6,0z" fill="rgb(128,178,163)"></path></g></svg></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(123, 172, 163);border-width: 4px;border-radius: 8px;border-style: solid;border-color: rgb(123, 172, 163);overflow: hidden;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 0.5em;margin-bottom: 0.5em;text-align: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;box-sizing: border-box;"><section style="width: 6px;height: 6px;display: inline-block;vertical-align: middle;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: middle;padding-right: 5px;padding-left: 5px;color: rgb(208, 241, 233);box-sizing: border-box;"><p style="box-sizing: border-box;">人美手巧的程序媛乙</p></section><section style="display: inline-block;vertical-align: middle;width: 6px;height: 6px;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(255, 255, 255);padding: 10px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">C男<span style="letter-spacing: 0px;box-sizing: border-box;">自称来自某专注于AI相关的VC，在校时是数学系第一名，P大在读硕士</span></p>
<p style="white-space: normal;box-sizing: border-box;">[加了微信]</p>
<p style="white-space: normal;box-sizing: border-box;">C男: 前几天我还在准备P大的考试，比较忙</p>
<p style="white-space: normal;box-sizing: border-box;">乙女：（？？不是在读吗）…好的</p>
<p style="white-space: normal;box-sizing: border-box;">C男：下午有空吗要不我去你公司找你聊聊？</p>
<p style="white-space: normal;box-sizing: border-box;">乙女：…行。（半晌）那我是不是给你约个访客？</p>
<p style="white-space: normal;box-sizing: border-box;">C男：对，名字XXX，手机XXXX</p>
<p style="white-space: normal;box-sizing: border-box;">乙女：...好</p>
<p style="white-space: normal;box-sizing: border-box;">[见面]</p>
<p style="white-space: normal;box-sizing: border-box;">C男：你们工作会接触很多的人吧</p>
<p style="white-space: normal;box-sizing: border-box;">乙女：...算是吧</p>
<p style="white-space: normal;box-sizing: border-box;">…（省略半小时难挨的对话过程）</p>
<p style="white-space: normal;box-sizing: border-box;">[微信]</p>
<p style="white-space: normal;box-sizing: border-box;">C男：转发链接「你在一个女人上寻求的是什么」</p>
<p style="white-space: normal;box-sizing: border-box;">乙女：？？</p></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 100%;box-sizing: border-box;"><svg class="svg" xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 73.9 2.7" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><path style="box-sizing: border-box;" d="M38.4,1.4c0,0.8-0.6,1.4-1.4,1.4c-0.8,0-1.4-0.6-1.4-1.4C35.7,0.6,36.3,0,37,0   C37.8,0,38.4,0.6,38.4,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M70.3,1.4c0,0.5-0.4,0.8-0.8,0.8c-0.5,0-0.8-0.4-0.8-0.8c0-0.5,0.4-0.8,0.8-0.8   C69.9,0.5,70.3,0.9,70.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M72.3,1.4c0,0.3-0.3,0.6-0.6,0.6c-0.3,0-0.6-0.3-0.6-0.6c0-0.3,0.3-0.6,0.6-0.6   C72.1,0.8,72.3,1.1,72.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M73.9,1.4c0,0.2-0.2,0.4-0.4,0.4c-0.2,0-0.4-0.2-0.4-0.4c0-0.2,0.2-0.4,0.4-0.4   C73.8,1,73.9,1.2,73.9,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M68.4,0L68.4,0c-0.8,0-1.3,0.5-1.4,1.2H56.5C52.3,1,46.4,0.8,39.7,0.8c0,0.2,0.1,0.4,0.1,0.6   c0,0.2-0.1,0.4-0.1,0.6c6.6,0,12.5-0.2,16.6-0.4H67c0.1,0.6,0.6,1.2,1.3,1.2c0,0,0,0,0.1,0c-0.3-0.4-0.6-0.8-0.6-1.4   C67.8,0.8,68,0.4,68.4,0z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M3.6,1.4c0,0.5,0.4,0.8,0.8,0.8c0.5,0,0.8-0.4,0.8-0.8c0-0.5-0.4-0.8-0.8-0.8C4,0.5,3.6,0.9,3.6,1.4   z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M1.6,1.4c0,0.3,0.3,0.6,0.6,0.6c0.3,0,0.6-0.3,0.6-0.6c0-0.3-0.3-0.6-0.6-0.6   C1.8,0.8,1.6,1.1,1.6,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M0,1.4c0,0.2,0.2,0.4,0.4,0.4c0.2,0,0.4-0.2,0.4-0.4C0.8,1.2,0.6,1,0.4,1C0.2,1,0,1.2,0,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M5.6,0L5.6,0c0.8,0,1.3,0.5,1.4,1.2h10.5C21.6,1,27.5,0.8,34.2,0.8c0,0.2-0.1,0.4-0.1,0.6   c0,0.2,0.1,0.4,0.1,0.6c-6.6,0-12.5-0.2-16.6-0.4H6.9C6.8,2.2,6.3,2.7,5.6,2.7c0,0,0,0-0.1,0c0.3-0.4,0.6-0.8,0.6-1.4   C6.1,0.8,5.9,0.4,5.6,0z" fill="rgb(128,178,163)"></path></g></svg></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="letter-spacing: 0px;box-sizing: border-box;">D男：请问您有户口吗？我是没有，留京好难。打扰了。</span></p>
<p style="white-space: normal;box-sizing: border-box;">乙女：？？</p></section></section></section></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(123, 172, 163);border-width: 4px;border-radius: 8px;border-style: solid;border-color: rgb(123, 172, 163);overflow: hidden;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 0.5em;margin-bottom: 0.5em;text-align: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;box-sizing: border-box;"><section style="width: 6px;height: 6px;display: inline-block;vertical-align: middle;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: middle;padding-right: 5px;padding-left: 5px;color: rgb(208, 241, 233);box-sizing: border-box;"><p style="box-sizing: border-box;">音乐是命的程序媛丙</p></section><section style="display: inline-block;vertical-align: middle;width: 6px;height: 6px;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(255, 255, 255);padding: 10px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="letter-spacing: 0px;box-sizing: border-box;">E男，长辈介绍的朋友的孩子。</span><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">相亲配置：丙女，E男及E男的爸妈，介绍人。</p>
<p style="white-space: normal;box-sizing: border-box;">相亲吃饭过程平平无奇，直到最后：</p>
<p style="white-space: normal;box-sizing: border-box;">介绍人：“小E小丙我给你们拍个照片吧，小丙你爸爸妈妈都没有来，让他们也看看。”</p>
<p style="white-space: normal;box-sizing: border-box;">丙女（好尴尬但还是要保持围笑）</p>
<p style="white-space: normal;box-sizing: border-box;">拍完照，介绍人给传群里，E男爸爸端一杯酒走过来，</p>
<p style="white-space: normal;box-sizing: border-box;">E男爸爸：“叔叔不强求你们，任由你们年轻人自己发展。但叔叔还是在这里预祝你们有个好结果，百年好合！”</p>
<p style="white-space: normal;box-sizing: border-box;">丙女<span style="letter-spacing: 0px;box-sizing: border-box;">（……，应该说“叔叔同喜同喜吗”）</span></p></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 100%;box-sizing: border-box;"><svg class="svg" xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 73.9 2.7" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><path style="box-sizing: border-box;" d="M38.4,1.4c0,0.8-0.6,1.4-1.4,1.4c-0.8,0-1.4-0.6-1.4-1.4C35.7,0.6,36.3,0,37,0   C37.8,0,38.4,0.6,38.4,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M70.3,1.4c0,0.5-0.4,0.8-0.8,0.8c-0.5,0-0.8-0.4-0.8-0.8c0-0.5,0.4-0.8,0.8-0.8   C69.9,0.5,70.3,0.9,70.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M72.3,1.4c0,0.3-0.3,0.6-0.6,0.6c-0.3,0-0.6-0.3-0.6-0.6c0-0.3,0.3-0.6,0.6-0.6   C72.1,0.8,72.3,1.1,72.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M73.9,1.4c0,0.2-0.2,0.4-0.4,0.4c-0.2,0-0.4-0.2-0.4-0.4c0-0.2,0.2-0.4,0.4-0.4   C73.8,1,73.9,1.2,73.9,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M68.4,0L68.4,0c-0.8,0-1.3,0.5-1.4,1.2H56.5C52.3,1,46.4,0.8,39.7,0.8c0,0.2,0.1,0.4,0.1,0.6   c0,0.2-0.1,0.4-0.1,0.6c6.6,0,12.5-0.2,16.6-0.4H67c0.1,0.6,0.6,1.2,1.3,1.2c0,0,0,0,0.1,0c-0.3-0.4-0.6-0.8-0.6-1.4   C67.8,0.8,68,0.4,68.4,0z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M3.6,1.4c0,0.5,0.4,0.8,0.8,0.8c0.5,0,0.8-0.4,0.8-0.8c0-0.5-0.4-0.8-0.8-0.8C4,0.5,3.6,0.9,3.6,1.4   z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M1.6,1.4c0,0.3,0.3,0.6,0.6,0.6c0.3,0,0.6-0.3,0.6-0.6c0-0.3-0.3-0.6-0.6-0.6   C1.8,0.8,1.6,1.1,1.6,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M0,1.4c0,0.2,0.2,0.4,0.4,0.4c0.2,0,0.4-0.2,0.4-0.4C0.8,1.2,0.6,1,0.4,1C0.2,1,0,1.2,0,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M5.6,0L5.6,0c0.8,0,1.3,0.5,1.4,1.2h10.5C21.6,1,27.5,0.8,34.2,0.8c0,0.2-0.1,0.4-0.1,0.6   c0,0.2,0.1,0.4,0.1,0.6c-6.6,0-12.5-0.2-16.6-0.4H6.9C6.8,2.2,6.3,2.7,5.6,2.7c0,0,0,0-0.1,0c0.3-0.4,0.6-0.8,0.6-1.4   C6.1,0.8,5.9,0.4,5.6,0z" fill="rgb(128,178,163)"></path></g></svg></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">F男，长辈介绍的朋友的孩子。</p>
<p style="white-space: normal;box-sizing: border-box;">相亲配置：丙女，F男</p>
<p style="white-space: normal;box-sizing: border-box;">在KTV唱歌，丙女找到一首当时还挺喜欢但没在别的KTV找到过的歌，很开心。</p>
<p style="white-space: normal;box-sizing: border-box;">然后一口气唱了5遍……</p>
<p style="white-space: normal;box-sizing: border-box;">然后F男<span style="letter-spacing: 0px;box-sizing: border-box;">再没和</span><span style="letter-spacing: 0px;box-sizing: border-box;">丙女</span><span style="letter-spacing: 0px;box-sizing: border-box;">联系过……</span></p>
<p style="white-space: normal;box-sizing: border-box;">那首歌叫《易燃易爆炸》。</p></section></section></section></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(123, 172, 163);border-width: 4px;border-radius: 8px;border-style: solid;border-color: rgb(123, 172, 163);overflow: hidden;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 0.5em;margin-bottom: 0.5em;text-align: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;box-sizing: border-box;"><section style="width: 6px;height: 6px;display: inline-block;vertical-align: middle;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: middle;padding-right: 5px;padding-left: 5px;color: rgb(208, 241, 233);box-sizing: border-box;"><p style="box-sizing: border-box;">身高175的程序媛丁</p></section><section style="display: inline-block;vertical-align: middle;width: 6px;height: 6px;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(255, 255, 255);padding: 10px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">在公司内网加了G男微信，聊了两天，觉得不太合适就一直没见面。丁女<span style="letter-spacing: 0px;box-sizing: border-box;">头像朋友圈都是自己照片，G</span><span style="letter-spacing: 0px;box-sizing: border-box;">男一张照片都没有。</span></p>
<p style="white-space: normal;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">此后G男每隔两天就发一条信息，</p>
<p style="white-space: normal;box-sizing: border-box;">“你是不是在二楼”</p>
<p style="white-space: normal;box-sizing: border-box;">“我在大堂看见你了”</p>
<p style="white-space: normal;box-sizing: border-box;">“你这几天好像长胖了，这可不行”</p>
<p style="white-space: normal;box-sizing: border-box;">丁女从未理会，G男从未放弃。</p>
<p style="white-space: normal;box-sizing: border-box;">最牛逼的是发来了偷拍丁女的照片，“你刚才在饭堂吃饭了”。</p>
<p style="white-space: normal;box-sizing: border-box;">（请问现在报警来得及吗）</p></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 100%;box-sizing: border-box;"><svg class="svg" xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 73.9 2.7" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><path style="box-sizing: border-box;" d="M38.4,1.4c0,0.8-0.6,1.4-1.4,1.4c-0.8,0-1.4-0.6-1.4-1.4C35.7,0.6,36.3,0,37,0   C37.8,0,38.4,0.6,38.4,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M70.3,1.4c0,0.5-0.4,0.8-0.8,0.8c-0.5,0-0.8-0.4-0.8-0.8c0-0.5,0.4-0.8,0.8-0.8   C69.9,0.5,70.3,0.9,70.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M72.3,1.4c0,0.3-0.3,0.6-0.6,0.6c-0.3,0-0.6-0.3-0.6-0.6c0-0.3,0.3-0.6,0.6-0.6   C72.1,0.8,72.3,1.1,72.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M73.9,1.4c0,0.2-0.2,0.4-0.4,0.4c-0.2,0-0.4-0.2-0.4-0.4c0-0.2,0.2-0.4,0.4-0.4   C73.8,1,73.9,1.2,73.9,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M68.4,0L68.4,0c-0.8,0-1.3,0.5-1.4,1.2H56.5C52.3,1,46.4,0.8,39.7,0.8c0,0.2,0.1,0.4,0.1,0.6   c0,0.2-0.1,0.4-0.1,0.6c6.6,0,12.5-0.2,16.6-0.4H67c0.1,0.6,0.6,1.2,1.3,1.2c0,0,0,0,0.1,0c-0.3-0.4-0.6-0.8-0.6-1.4   C67.8,0.8,68,0.4,68.4,0z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M3.6,1.4c0,0.5,0.4,0.8,0.8,0.8c0.5,0,0.8-0.4,0.8-0.8c0-0.5-0.4-0.8-0.8-0.8C4,0.5,3.6,0.9,3.6,1.4   z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M1.6,1.4c0,0.3,0.3,0.6,0.6,0.6c0.3,0,0.6-0.3,0.6-0.6c0-0.3-0.3-0.6-0.6-0.6   C1.8,0.8,1.6,1.1,1.6,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M0,1.4c0,0.2,0.2,0.4,0.4,0.4c0.2,0,0.4-0.2,0.4-0.4C0.8,1.2,0.6,1,0.4,1C0.2,1,0,1.2,0,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M5.6,0L5.6,0c0.8,0,1.3,0.5,1.4,1.2h10.5C21.6,1,27.5,0.8,34.2,0.8c0,0.2-0.1,0.4-0.1,0.6   c0,0.2,0.1,0.4,0.1,0.6c-6.6,0-12.5-0.2-16.6-0.4H6.9C6.8,2.2,6.3,2.7,5.6,2.7c0,0,0,0-0.1,0c0.3-0.4,0.6-0.8,0.6-1.4   C6.1,0.8,5.9,0.4,5.6,0z" fill="rgb(128,178,163)"></path></g></svg></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="letter-spacing: 0px;box-sizing: border-box;">丁女工作关系结识了H男。</span><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">H男表示好感，丁女屡屡回避。</p>
<p style="white-space: normal;box-sizing: border-box;">最后一次，丁女明确拒绝，H男拉黑丁女。</p>
<p style="white-space: normal;box-sizing: border-box;">H男时隔数周又加回丁女，询问近况。丁女隔天回复，发现已再次被拉黑。</p>
<p style="white-space: normal;box-sizing: border-box;">数周后再次收到H男回复，丁女隔天回复，发现再被拉黑。</p>
<p style="white-space: normal;box-sizing: border-box;">（大哥到底让不让人说话...）</p></section></section></section></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(123, 172, 163);border-width: 4px;border-radius: 8px;border-style: solid;border-color: rgb(123, 172, 163);overflow: hidden;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 0.5em;margin-bottom: 0.5em;text-align: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;box-sizing: border-box;"><section style="width: 6px;height: 6px;display: inline-block;vertical-align: middle;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: middle;padding-right: 5px;padding-left: 5px;color: rgb(208, 241, 233);box-sizing: border-box;"><p style="box-sizing: border-box;">拍照超强的程序媛<span style="letter-spacing: 0px;box-sizing: border-box;">戊</span></p></section><section style="display: inline-block;vertical-align: middle;width: 6px;height: 6px;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(255, 255, 255);padding: 10px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">平日工作压力和强度比较大，个人时间精力很有限，一般父母亲朋介绍的相亲能推就推。有一次相亲对象是<span style="letter-spacing: 0px;box-sizing: border-box;">戊女妈妈</span><span style="letter-spacing: 0px;box-sizing: border-box;">同学的儿子，阿姨到北京出差，一再说要请她儿子和戊女一起吃个饭，认识一下交个朋友。戊女妈妈念叨了三回你必须得去你必须得去你必须得去，没办法下班早走了一会儿，顶着好几天加班睡眠不足的黑眼圈，目光呆滞素面朝天的去相亲。</span></p>
<p style="white-space: normal;box-sizing: border-box;">餐厅是一家排队巨可怕的烤鸭店，三个人坐在等位的椅子上静默无言，阿姨直勾勾的盯着戊女笑而不语，小哥四十五度看天四十五度看地，我也只好一边笑而不语一边分神祈祷实习生别把昨天误删数据库的骚操作再来一遍。</p>
<p style="white-space: normal;box-sizing: border-box;">就这么着尬了不知道多长时间，阿姨突然说你们都饿了吧，这家酸奶好像还挺好吃的，我们先来点尝尝？</p>
<p style="white-space: normal;box-sizing: border-box;">那一瞬间可能就是传说中的福至心灵，也可能是走神太远猝不及防，戊女特别顺溜地接了一句，没事阿姨不用麻烦了，我自己带了。然后特别顺溜地从包里摸出来早上出门带的但是一直没来得及喝的酸奶。阿姨直勾勾地盯着戊女不言不语，小哥继续四十五度看天四十五度看地。</p>
<p style="white-space: normal;box-sizing: border-box;">后来得知这件事的戊女妈妈情绪十分激动，专门视频过来气壮山河地吼，说出息了啊你，上大街上问问去有谁相亲自己带酸奶的？！</p>
<p style="white-space: normal;box-sizing: border-box;">虽然过程不怎么愉快，但这之后妈妈再也不提相亲这个事儿了。可能她也对戊女绝望了吧。</p></section></section></section></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(123, 172, 163);border-width: 4px;border-radius: 8px;border-style: solid;border-color: rgb(123, 172, 163);overflow: hidden;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 0.5em;margin-bottom: 0.5em;text-align: center;box-sizing: border-box;"><section style="display: inline-block;vertical-align: middle;box-sizing: border-box;"><section style="width: 6px;height: 6px;display: inline-block;vertical-align: middle;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: middle;padding-right: 5px;padding-left: 5px;color: rgb(208, 241, 233);box-sizing: border-box;"><p style="box-sizing: border-box;">男友力爆棚的程序媛己</p></section><section style="display: inline-block;vertical-align: middle;width: 6px;height: 6px;border-radius: 100%;background-color: rgb(255, 255, 255);box-sizing: border-box;"></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(255, 255, 255);padding: 10px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">J男：你是学什么专业的？</p>
<p style="white-space: normal;box-sizing: border-box;">己女：我是学计算机的</p>
<p style="white-space: normal;box-sizing: border-box;">J男从此消失了，过年发微信祝春节快乐的时候，发现被删除好友了…</p></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;line-height: 0;width: 100%;box-sizing: border-box;"><svg class="svg" xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 73.9 2.7" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%"><g style="box-sizing: border-box;"><path style="box-sizing: border-box;" d="M38.4,1.4c0,0.8-0.6,1.4-1.4,1.4c-0.8,0-1.4-0.6-1.4-1.4C35.7,0.6,36.3,0,37,0   C37.8,0,38.4,0.6,38.4,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M70.3,1.4c0,0.5-0.4,0.8-0.8,0.8c-0.5,0-0.8-0.4-0.8-0.8c0-0.5,0.4-0.8,0.8-0.8   C69.9,0.5,70.3,0.9,70.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M72.3,1.4c0,0.3-0.3,0.6-0.6,0.6c-0.3,0-0.6-0.3-0.6-0.6c0-0.3,0.3-0.6,0.6-0.6   C72.1,0.8,72.3,1.1,72.3,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M73.9,1.4c0,0.2-0.2,0.4-0.4,0.4c-0.2,0-0.4-0.2-0.4-0.4c0-0.2,0.2-0.4,0.4-0.4   C73.8,1,73.9,1.2,73.9,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M68.4,0L68.4,0c-0.8,0-1.3,0.5-1.4,1.2H56.5C52.3,1,46.4,0.8,39.7,0.8c0,0.2,0.1,0.4,0.1,0.6   c0,0.2-0.1,0.4-0.1,0.6c6.6,0,12.5-0.2,16.6-0.4H67c0.1,0.6,0.6,1.2,1.3,1.2c0,0,0,0,0.1,0c-0.3-0.4-0.6-0.8-0.6-1.4   C67.8,0.8,68,0.4,68.4,0z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M3.6,1.4c0,0.5,0.4,0.8,0.8,0.8c0.5,0,0.8-0.4,0.8-0.8c0-0.5-0.4-0.8-0.8-0.8C4,0.5,3.6,0.9,3.6,1.4   z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M1.6,1.4c0,0.3,0.3,0.6,0.6,0.6c0.3,0,0.6-0.3,0.6-0.6c0-0.3-0.3-0.6-0.6-0.6   C1.8,0.8,1.6,1.1,1.6,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M0,1.4c0,0.2,0.2,0.4,0.4,0.4c0.2,0,0.4-0.2,0.4-0.4C0.8,1.2,0.6,1,0.4,1C0.2,1,0,1.2,0,1.4z" fill="rgb(128,178,163)"></path><path style="box-sizing: border-box;" d="M5.6,0L5.6,0c0.8,0,1.3,0.5,1.4,1.2h10.5C21.6,1,27.5,0.8,34.2,0.8c0,0.2-0.1,0.4-0.1,0.6   c0,0.2,0.1,0.4,0.1,0.6c-6.6,0-12.5-0.2-16.6-0.4H6.9C6.8,2.2,6.3,2.7,5.6,2.7c0,0,0,0-0.1,0c0.3-0.4,0.6-0.8,0.6-1.4   C6.1,0.8,5.9,0.4,5.6,0z" fill="rgb(128,178,163)"></path></g></svg></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-77336-1550217870709" style="font-size: 15px;line-height: 1.8;color: rgb(46, 101, 91);box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="letter-spacing: 0px;box-sizing: border-box;">K男（刚加微信没看过照片）：你不是大美女吧，我气质比较没有，配不上。</span><br style="box-sizing: border-box;"></p>
<p style="white-space: normal;box-sizing: border-box;">己女（一脸懵逼，这句接不上）</p>
<p style="white-space: normal;box-sizing: border-box;">转话题聊兴趣</p>
<p style="white-space: normal;box-sizing: border-box;">K男：PM太会玩，RD跟不上…</p>
<p style="white-space: normal;box-sizing: border-box;">己女……</p></section></section></section></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 10px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;background-color: rgb(191, 224, 216);box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="border-bottom: 1px solid rgb(123, 172, 163);box-sizing: border-box;"></section><section style="width: 0px;margin: auto;border-top: 9px solid rgb(123, 172, 163);border-left: 9px solid transparent !important;border-right: 9px solid transparent !important;box-sizing: border-box;"></section><section style="width: 0px;margin-top: -10px;margin-right: auto;margin-left: auto;border-top: 9px solid rgb(255, 255, 255);box-sizing: border-box;border-left: 9px solid transparent !important;border-right: 9px solid transparent !important;"></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;width: 100%;vertical-align: top;padding: 15px 10px;box-sizing: border-box;"><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section class="tn-yzk-fuid-text-34768-1548224337401" style="font-size: 15px;color: rgb(15, 79, 67);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="font-family: Optima-Regular, PingFangTC-light;box-sizing: border-box;">感谢大家的收看～也欢迎大家私信后台分享自己的故事哟！么么哒～下期再见！</span></p></section></section></section></section></section></section></section></section></section><section style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: 10px;margin-left: 0%;box-sizing: border-box;"><section style="border-bottom: 1px solid rgb(123, 172, 163);box-sizing: border-box;"></section><section style="width: 0px;margin: auto;border-top: 9px solid rgb(123, 172, 163);border-left: 9px solid transparent !important;border-right: 9px solid transparent !important;box-sizing: border-box;"></section><section style="width: 0px;margin-top: -10px;margin-right: auto;margin-left: auto;border-top: 9px solid rgb(191, 224, 216);box-sizing: border-box;border-left: 9px solid transparent !important;border-right: 9px solid transparent !important;"></section></section></section></section><p><img class="" data-copyright="0" data-ratio="0.4514877102199224" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNpKic4DiblDA75KduXD0EeqoZ8McBkcR5DUfib3SQ5IiaNYpzLW98Valvic1XIQmFEIbibgFwZUPNb69vcg/640?wx_fmt=png" data-type="png" data-w="773" style='font-family: -apple-system-font, BlinkMacSystemFont, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;'><br></p>
<p><br></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693287&amp;idx=1&amp;sn=006a8ec5d0a478965b3cd32cfedba067#wechat_redirect</link>
      <pubDate>Sun, 03 Mar 2019 17:57:28 -0800</pubDate>
      <guid isPermaLink="false">103874</guid>
    </item>
    <item>
      <title>日常 | GHC 18 参会感想</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p><span style="font-size: 15px;">大家好，我是 Yangyang，目前在 Microsoft Vancouver 工作了一年多。今年是我第一次参加 GHC，</span><em style="font-size: 15px;">很受鼓舞和启发</em><span style="font-size: 15px;">，在此与大家分享我的所得：GHC速览 &amp; GHC 18 感想 &amp; 参会Tips。 </span><br></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 270px;height: 90px;vertical-align: top;overflow: hidden;background-position: 50% 50%;background-repeat: no-repeat;background-size: cover;background-attachment: scroll;background-image: url("https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyyIBpNXGAgYNvjUiaFE0GfrSdPTIo4cdN32nXfYcS2Dom5bgDFFQBzuQ/640?wx_fmt=gif");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 45px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="font-size: 18px;color: rgb(98, 82, 63);box-sizing: border-box;"><p style="box-sizing: border-box;"><strong style="box-sizing: border-box;">GHC 速览</strong></p></section></section></section></section></section></section></section><p><strong><span style="font-size: 15px;">GHC</span></strong><span style="font-size: 15px;"> 的全称为 <em>GraceHopper Celebration of Women in Computing</em>，官网为https://ghc.anitab.org/。GHC 由 AnitaB.org 和 ACM 主办，从1994年至今已经举办了18届，2006年开始每年举办一次，会期多在九月和十月。<br>GHC 主要包括 <strong>technical 和 career 两大类 session</strong>，同时也有 <em>career fair, awards ceremony, competitions </em>等多个环节，聚集了来自全球各地的女性科技工作者以及支持 GHC 社区的同盟（GHC也非常欢迎非技术类妹子参加）。<br><strong>GHC 18 于9月26-28日在 Houston, Texas 举办，参会人数为 20,000+。</strong></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.66484375" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy4Jp2NYbCibarPakIeTNz0PE5mlqxHqm4KDFibj5Yc5EKCHIVKdIKiaicFg/640?wx_fmt=jpeg" data-type="jpeg" data-w="1280" style=""></p></section></section><p style="text-align: center;"><span style="font-size: 12px;color: rgb(178, 178, 178);">主会场 - GRB 外观</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 270px;height: 90px;vertical-align: top;overflow: hidden;background-position: 50% 50%;background-repeat: no-repeat;background-size: cover;background-attachment: scroll;background-image: url("https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyTSnJxbdWZAdDan0Ash9GlhBnDLwjWuEs9H6icY1qiaPKiaodGyMPZAibNg/640?wx_fmt=gif");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 45px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="font-size: 18px;color: rgb(98, 82, 63);box-sizing: border-box;"><p style="box-sizing: border-box;"><strong style="box-sizing: border-box;">GHC 18 感想</strong></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;"></span></p>
<p><span style="font-size: 15px;">在 GHC 18 的三天里，我每天给自己安排了6-7个 session，<em>主要是关于 career 的，想获得一些 early-career 的发展建议</em>。<br>九月的最后一天，我回想过去的九月和过去的二十几年，<strong>问了自己的很多问题，重新去认识和面对自己</strong>，然后明白了我从 GHC 18 学到的最重要的事是：<strong>要真正地面对自己，相信自己，想说什么就要说出来，想要什么就要开口，主动向前一步。</strong></span></p>
<p style="text-align: center;"><br></p>
<p><span style="font-size: 15px;">那天我有点激动，写了很长的朋友圈，下午又写了英文版（点击阅读原文可查看）以分享给更多人，在这里分享最主要的三个方面给大家：</span><br></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 20px 0% 10px;box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;padding-right: 8px;padding-bottom: 8px;padding-left: 8px;background-color: rgb(245, 213, 66);box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: -10px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="padding: 3px;display: inline-block;border-bottom: 1px solid rgb(62, 62, 62);line-height: 1;letter-spacing: 0.8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 15px;text-decoration: none;">Confront Yourself Honestly and Trust Yourself from Your Deep Heart</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;"> 在 GHC18 的 Session:Nail Your Promotion 里有一个环节：和你身边的人互相练习要用“I”开头的句子来描述你最近工作里的 achievements，</span><strong style="font-size: 15px;">一定是“I”而不是“We”</strong><span style="font-size: 15px;">。我跟同事 H 坐在一起，一开始我们不知道说什么，后来一句一句越说越多，到最后我们都有点想哭。因为我们是同组的同事，互相了解对方的工作，所以后面都帮对方补充了一些“I”句子。</span><em style="font-size: 15px;">那一瞬间，我意识到，我原来这么不相信自己，原来我做了这么多，原来别人也认可我的工作，只有自己没有给自己认可。</em></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">在这个 session 过后的几天里，我一直在问自己：什么让我觉得自豪？做什么让我真正地从内心感到快乐？我在害怕什么？<strong>我为什么不相信自己？我想成为什么样的人？</strong>我过去是怎么做的？有什么是我想保持的和想改变的？后来我写下了非常多的“我”句子，这里分享一部分吧： </span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: 64.8852%;background-attachment: scroll;letter-spacing: 0px;padding: 15px;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyjk8sdm4lhbanj6jb0lzruVT6GWXdbOfKJMhrzrAAPN11GicOr7hiacRQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">我是个好姑娘，完全是。我热爱生活，努力工作，想要多挣钱，带爸妈去看世界，给家庭带来更好的未来。我唱歌很好听，特别是唱邓丽君的歌。我拍照也不错，我有发现美的眼睛，我把那些美好的片刻分享出来，让快乐和感动翻倍翻倍再翻倍。我是个生活和工作都比较 organized 的人，我把家里收拾得窗明几净，我也很会打包行李。我有点强迫症，有时其实这给我带来了一些麻烦，但我现在意识到了，也乐于去改进。我积累着自己的生活和工作经验，方方面面，而且我非常乐于分享给别人，帮助别人，这也让我感到和快乐。我是个helpful的人，未来我想做个 knowledgeable 的人，然后依然 helpful，给他人支持和帮助！</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;"> </span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">我有些胖，但不是那么胖，我尝试去健身，但我确实毅力上还有些问题。我害怕所有和高速有关的冒险，比如滑雪比如开卡丁车。我也很害怕未知，现在游泳还不敢去深水区，因为我总在脑子里想象一万遍自己呛水然后被救生员捞起来的场景。我发现克服心理障碍真的不容易，不要因为别人说这很容易啊然后觉得自己不行，不要为自己此刻做不到而且苛责自己。我害怕开车，未来我当然要学会在高速开车，因为这是我生活的一部分，我会找个靠谱的教练，多上几节课，问问我的女性朋友们怎么做到的，寻求她们的帮助。我也挺开心我尝试过一些对我来说有点难的冒险，然后跪求自己不要再逼自己尝试了，因为我真的不喜欢，比如卡丁车。有个朋友在看我开过卡丁车后对我说：这对你来说是折磨，我完全同意，也不想再尝试，let it go。我会继续学滑雪，即使我今年一月摔裂了膝盖右侧韧带，过了半年才好，因为我觉滑雪场风景太好了，我也有点享受滑雪的感觉。我会学着安全地摔倒，学着冷静一些，学着 trust my body，然后不再阻止我老公去蓝道探索，而是对他说：Have fun and take care of yourself。我也会继续学自由泳，学踩水，即使可能需要很久才能学会，要过很久才能在深水区游个来回。</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">在这样几番的自我对话之后，我开始相信自己是很棒的人，是值得被爱的人，也是认真去爱的人。如果你是一个不够自信的人，那么就尝试向自己提问吧，一定要让内心深处的自己来回答。<strong>Confidence is the stuff that turns our ideas into actions. What we believe matters.</strong> 诚实地面对自己，相信自己，接受自己，是自我成长的重要一步，有时甚至是第一步。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 20px 0% 10px;box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;padding-right: 8px;padding-bottom: 8px;padding-left: 8px;background-color: rgb(245, 213, 66);box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: -10px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="padding: 3px;display: inline-block;border-bottom: 1px solid rgb(62, 62, 62);line-height: 1;letter-spacing: 0.8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 15px;">Be Active and Lean In</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">这是我在 GHC18 的多个 session 听到的共同点，听起来有点鸡汤，但从前我都是“道理我都懂，生活还是过成了这样”。<em>向前一步，曾经真的很难，尤其对于女生。</em>多少个时刻，我在犹豫，我在胆怯，我欲言又止，然后什么都没有改变，自己也不开心。在 GHC18 的经历告诉我：</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;text-align: left;"><section style="display: inline-block;vertical-align: top;width: 29%;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-left: 0%;text-align: right;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"><section style="display: inline-block;vertical-align: bottom;width: 10px;margin-bottom: 5px;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: top;margin-left: 2px;color: rgb(132, 132, 132);font-size: 14px;text-align: left;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 12px;">1<br style="box-sizing: border-box;"></span></p></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 71%;border-left: 5px solid rgb(245, 213, 66);border-bottom-left-radius: 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 15px 0% 8px;box-sizing: border-box;"><section style="border-width: 0px 0px 1px;border-style: solid;border-bottom-color: rgb(132, 132, 132);color: rgb(132, 132, 132);text-align: justify;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="font-size: 12px;">Ask For What You Want</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">第一次听到这句话是在Session: Nail Your Promotion。回想自己工作满一年的时候有个升职节点，我想知道自己有没有机会，但我总等着 manager 来跟我聊，而我同组的两个哥们一直在催我主动跟 manager 聊。我犹犹豫豫好几次，我怕尴尬，我怕失望。终于在一次1/1上跟 manager 问了，manager 给了我反馈和认可，也解释了为什么这次没升，也聊到了其他人都做怎么样。听过这个 Session，我回想起来，这件事最后并不尴尬，反而从中获得了很多有用的信息。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">所以，当你想退缩时，问问自己：我在害怕什么呢？事情真的这么复杂吗？最坏的结果是什么呢？我能做什么让结果不那么坏？然后你会发现，有时你只需要 Say Hello，只需要 Ask Questions。所以啊，<strong>Don’t wait for the door to open. Open doors and go through them. </strong></span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;width: 29%;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-left: 0%;text-align: right;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"><section style="display: inline-block;vertical-align: bottom;width: 10px;margin-bottom: 5px;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"></section></section><section style="display: inline-block;vertical-align: top;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;margin-left: 2px;color: rgb(132, 132, 132);font-size: 14px;text-align: left;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 12px;">2</span></p></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 71%;border-left: 5px solid rgb(245, 213, 66);border-bottom-left-radius: 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 15px 0% 8px;box-sizing: border-box;"><section style="border-width: 0px 0px 1px;border-style: solid;border-bottom-color: rgb(132, 132, 132);color: rgb(132, 132, 132);text-align: justify;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="font-size: 12px;">Measure Your Achivements</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">在 Keynote 里 Padmasree Warrior 提到：Eyes on the star, feet on the ground。当时我就想，这不就是“仰望星空，脚踏实地”！可是还缺点什么呢，<em>怎么才能更接近自己的星星，怎么更好地脚踏实地？</em>后来我从其他的 session 得到了一种答案：Measure you achievements。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.665625" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyoXRAnaXLKmWruibp12adLy5LWl2HegvqoYL2vNR9BsHbyNaPR5I6IGA/640?wx_fmt=jpeg" data-type="jpeg" data-w="1280" style=""></p>
<p style="text-align: center;"><span style="font-size: 12px;color: rgb(178, 178, 178);">Keynote 现场</span></p></section></section><p><span style="font-size: 15px;">这是一条很有实践性的建议。比如工作中，你会不会觉得自己做了很多事情，但是好像太琐碎，别人也看不到？这时量化你的工作成果就很重要。你可以问自己，这个月我做了几次 Demo？我手中的项目有几个阶段，我做了哪几项工作，推进到了哪里？这个项目的结果有什么指标可以衡量，我期望以及实际提升了多少？比如生活中，你会不会觉得每年都过得好快啊，我怎么好像什么也没做？你同样可以尝试量化生活中的 achievements。比如问自己，今年去了哪几个地方，拜访了哪些人，读了哪些书，有哪些是我今年的大事件？<strong>Measure you achievements, 其实是在沉淀和分析你的所得，也帮你去设定新的目标，审视自己，也展示给他人。</strong></span><br></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;width: 29%;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-left: 0%;text-align: right;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"><section style="display: inline-block;vertical-align: bottom;width: 10px;margin-bottom: 5px;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: top;margin-left: 2px;color: rgb(132, 132, 132);font-size: 14px;text-align: left;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="box-sizing: border-box;">3</p></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 71%;border-left: 5px solid rgb(245, 213, 66);border-bottom-left-radius: 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 15px 0% 8px;box-sizing: border-box;"><section style="border-width: 0px 0px 1px;border-style: solid;border-bottom-color: rgb(132, 132, 132);font-size: 12px;color: rgb(132, 132, 132);text-align: justify;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">Communicate And Connect: GIVE &amp; GAIN</p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">在 GHC 18 的第三天上午，我取消了之前注册的 session，我花了一上午在 Career Fair 的各个展台之间穿梭，我问不同公司的工程师：How do you feel about yourcompany? Do you like your job? What’s your most interesting project recently? 朋友开玩笑说，你不是去参加面试，你是去面试 ta 们呀！再回想整个九月，我跟很多人交流了很多事情，比如，和不同的朋友聊工作聊家庭聊自己，和小伙伴们在 KTV 聊唱歌的经验，和当时的 Airbnb 房东聊中西文化差异，在 GHC 和同事聊参会感想，和爸妈聊对过去的和现在的自己的想法。这些都让我觉得很开心。</span><br><span style="font-size: 15px;"></span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><strong><span style="font-size: 15px;">我想，交流的意义在于连接，在于给予和获得的正向循环，在于建立属于自己的Sponsor Network。</span></strong><span style="font-size: 15px;">我越来越愿意说出自己的想法，获得他人的反馈，无论是支持还是反对；我也倾听和了解 ta 们，向 ta 们分享我的经验，给出我的反馈。<em>曾经我觉得“Impact”是一个很大很远的词，现在我觉得只要有一个人因为我分享的东西而受到启发，这就很有意义。</em>我也感谢他人给我带来 ta 们的 impact。这些都是你我所处世界的碰撞、拓展和延伸。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 20px 0% 10px;box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;padding-right: 8px;padding-bottom: 8px;padding-left: 8px;background-color: rgb(245, 213, 66);box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: -10px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="padding: 3px;display: inline-block;border-bottom: 1px solid rgb(62, 62, 62);line-height: 1;letter-spacing: 0.8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 15px;">Embrace Vulnerability</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">很多个 Session 都提到，要 walk across boundaries，要 push yourself to your boundaries，也提到 Growth is inversely proportional to comfort。我在从 GHC 18 回来之后，跟当时的 Airbnb 房东聊起这个话题，她给我推荐了 TED 演讲：<em>The power of vulnerability | Brené Brown</em>。我看完之后的体会是，<strong>如果我想要突破自己的极限，push 自己一把，首先要做的是接受自己的不完美。</strong></span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">就像我在上面那段“我”句子所说的，我有很多不完美的地方。而从前我是那么那么那么在意别人的看法，典型的讨好型人格。我把自己限制在自己想要的“完美”里，然后对自己的弱点深恶痛绝。我总是要让自己做到 well-prepared，然后对那偏离计划的10%而焦躁不安，而责备自己。现在我明白了，因为我不曾相信自己，不曾接受自己，我总想100%的去 control，去 predict，其实完全不可能做到。我要做的是去调研、去交流、去发现、去解决，去化解100%的压力，去把0%的信心变成60%的把握，然后认真地努力地向着100%行动。即使这件事成功之后，再回头看，你会发现过程本来就会 messy，再重头来过，也没法做到每个细节100%。Jesscia O. Matthews 在 Keynote 里说：<em>Just because it’s not your plan doesn’t mean it’s not your destiny。</em>所以我与自己和解，接受过去和现在的自己，接受不完美的自己，不苛求，有期待，有行动。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.66328125" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyZyicsibKg6zntaZ43ffwMZicGQY1QTMSs6FLePqeEutUF1pzlA6GF6BrA/640?wx_fmt=jpeg" data-type="jpeg" data-w="1280" style=""></p>
<p style="text-align: center;"><span style="font-size: 12px;color: rgb(178, 178, 178);">主会场 - GRB 涂鸦</span></p></section></section><p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">还想再说一些琐碎的感想：</span><br></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: 48.2185%;background-attachment: scroll;letter-spacing: 0px;padding: 15px;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUythBJicdAZFf5HbpstE9I0mGho7LKYdpKYUtLQzbkiaSg7DwJFtJjEvRg/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">1. 社区真的很重要。在 GHC 18，我时常感受到“啊，原来我不是一个人，原来我不孤单”，这从内心为我带来了勇气和支持。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">2. 我观察了 Keynote 和 Featured Session 的这些speaker，她们很明白自己想要给这个世界带来什么改变，也相信自己可以。自信自强自立的人真的会发光。(<em style="box-sizing: border-box;">Keynote的视频可以在YouTube上观看。</em>)</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">3. 关于 Career Fair：现场真的很热闹，逛不同公司的展台，观察他们的风格差异，是很有趣的事情。也感受到找工作竞争越来越大，佩服那些在展台和招聘人员非常流利地讲解自己简历的妹子们，我当年找工作时做不到这么自信从容。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">4. 在 Career Fair 的 Indeed 和 LinkedIn 展台可以拍那种职业头像，我在参会的第二天下午也去拍了一张。照片出来后我很惊讶，当时我问自己：这个看起来还挺自信的笑着的妹子，是我？我觉得这是 GHC 给我带来的即时改变。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">5. 我还需要从文化角度去提升英语交流能力。虽然我在温哥华工作了一年，基本的交流没有问题，对 keynote 和 session 中大部分梗的理解可以和大家同步，但有时还会有延迟。以前在国内读书，看了美剧英剧；来温哥华的一年，反而主要看的是国内剧集和综艺。借口包括“我要和国内朋友们维持话题”和“工作已经都是英语交流了，回家我想放松”，其实是自己不愿走出舒适区。这是我需要改变的地方。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">6. 我对很多印度妹子在 session 里的积极和主动感到很惊讶，听到了一些惊艳的想法和观点，体会到了职场中印度 vs 中国的竞争感。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">7. 在我参加的 session 里，台上的亚裔（我指的中日韩）面孔略少，但在颁奖环节又有好几个亚裔妹子获奖，让人感到惊喜。期待未来能够在世界舞台感受到更多亚裔女性工作者的影响力。</span></p></section></section></section></section></section></section></section><section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 270px;height: 90px;vertical-align: top;overflow: hidden;background-position: 50% 50%;background-repeat: no-repeat;background-size: cover;background-attachment: scroll;background-image: url("https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyyIBpNXGAgYNvjUiaFE0GfrSdPTIo4cdN32nXfYcS2Dom5bgDFFQBzuQ/640?wx_fmt=gif");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 45px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="font-size: 18px;color: rgb(98, 82, 63);box-sizing: border-box;"><p style="box-sizing: border-box;"><strong style="box-sizing: border-box;">GHC 参会 Tips</strong><span style="font-size: 15px;"> </span></p></section></section></section></section></section></section></section><section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">注册 &amp; 抢票</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">今年没有 guaranteed tickets，需要自己守在电脑前注册抢票。注册开放时间为 9AM PST。主办方今年更新了注册算法，记录每人最后一次打开的浏览器窗口中的最后一次点击为最后排队顺序。所以找一个网络靠谱的地方，刷出注册按钮后，一次点击即可，大概1-5分钟后会出现下一步。整个注册全部完成大概花了10分钟。每一步的响应都比较慢，要耐心。最好给最终显示注册成功的页面截个图，以防万一。最后收到 confirmation 邮件，才算是完全确保注册成功。</span><br></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">GHC 18 注册类型包括 general, academic, student三种，general 价格为1,150USD，后两者价格会便宜很多，听说学生注册是 300 USD，但我没有注册过，猜测会需要相应的身份证明。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">酒店</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">今年 GHC 的会场主要在 George R.Brown Convention Center (GRB), Marriott Marquis (MMQ) 和 Hilton Americas-Houston (HIL)，GRB 的两端有天桥分别连接 MMQ 和 HIL。GHC 官网有推荐的酒店列表，也有 shuttle 连接各个酒店和会场。我跟同事妹子 H 很幸运抢到了 HIL 的房间，省去了每天 shuttle 来回的时间。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.66640625" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUybwCQ13Vk01e3PHl1BzhuQHXNyiaNQtbfByNLuFahexcJhicodC7VYygg/640?wx_fmt=jpeg" data-type="jpeg" data-w="1280" style=""></p>
<p style="text-align: center;"><span style="font-size: 12px;color: rgb(178, 178, 178);">休斯顿 - 有轨电车</span></p></section></section><p><span style="font-size: 15px;">我们实际订酒店的过程有点慌乱。由于我们公司报销的 approver 多次强调确认拿到票了才可以订酒店，我们收到 confirmation 之后再去看酒店列表时，发现 HIL 每人一间会超额度。于是马上跟 approver 联系，确认两人一间没问题，便赶紧下手，最后在注册后的半小时抢到了一个 HIL 的房间。有国内小伙伴告诉我，当时她抢完票国内已经半夜了，想着第二天一早订酒店，结果第二天只能订到 shuttle 最远处的酒店了。</span><br></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">建议想参加 GHC 的小伙伴，先提前查看官网推荐的酒店列表信息，确认公司/学校报销限额，列出可选酒店；注册成功就马上订酒店。如果无需考虑报销限额，或不怕注册失败而酒店可能无法取消的问题，可以提前订酒店。</span><span style="font-size: 15px;"> </span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">Pre-Registration</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">GHC 18 可以通过网站或 APP 提前注册感兴趣的 session，获得 reserved spot。Keynotes 无需注册，featured sessions 的 reserved spot 很快会被约满。入场时有两个队列：reserverd, walk-up。Reserved 队列先入场，walk-up 队列后入场，一般也都有位置。</span><br></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">我是在 GHC 18 iOS APP 注册 reserved spot，Schedule 页面点击 session 右侧的+号即可。重叠时段只能注册一个，之后可以在 My Schedule 看到自己选择的 session。建议慎选20min session，个人感觉 career 类的20min session 比较水，内容基本和相似的1h session 重叠。<em>（另外同事反映，今年她在 Web 端的 reserved spot全部丢失，Andriod APP 经常卡死闪退。希望今年会有人对此给出反馈，GHC 19 能做得好些。）</em></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">航班 &amp; 时差</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">尽量选择直飞航班，不得不转机的话就尽量多留一些时间，飞机还是很容易延误的；如果没有买买买需求，带个小的登机箱会比较方便，转机也不用担心行李跟不上。</span><br></p>
<p><br></p>
<p><span style="font-size: 15px;">国内过来的小伙伴，中美时差很长，而会期很短，所以可能刚倒过来时差，回去又得倒回来。建议借助褪黑素，尽量撑到晚上睡。比如上飞机之前吃1-2粒褪黑素，尽量在飞机上多睡；落地后困了也尽量撑到晚上再睡；睡前1小时可以吃1粒褪黑素。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">行装准备</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">建议关注出发地和目的地天气，考虑会场有空调，必备薄外套和雨伞。比如：Houston 当时气温大概是23-30℃，偶尔有雷阵雨。室外很湿热，会场里面很冷，所以我当时是：夏装+外套+雨伞。<em>（我跟好几个妹子都吐槽了，GHC 作为面向女性的大会，空调温度上的关怀还不够啊。）</em></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">吃饭</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">第一天没有考虑人很多的问题，比较狼狈。酒店本身的自助早餐和附近的星巴克都排着很长的对，转了好几个地方，我们才在 GRB 二楼找到队伍稍短些的小餐吧买到吃的，一路奔去丰田中心听 Keynote。午饭时间依然到处排长队，因为我和同事都有 pre-registration，想听的 session 间隔很短，之后放弃了午饭。下午主办方供应了一些甜点、面包、蔬菜沙拉，可以拿来充饥。晚饭时间比较宽裕，可以寻找合适的餐馆，如果去参加 party 也有食物。建议参加 GHC 的妹子，每晚买好第二天的早饭或者订好酒店的早餐客房服务，同时也买点零食面包，免得因为来不及排队而吃不上午饭。</span><span style="font-size: 15px;"> </span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">After-hour Party</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">GHC 三天会期，前两天晚上有些公司会安排自己的 party, 第三晚是 GHC 官方的 party。第一天我去了 Microsoft party，公司也发了注册的 link 给 FTE，鼓励大家转发&amp;邀请其他参会者，位置离主会场步行20min可达。第二天有 Google 主办的 WomenTech Worker 的 party，需要先排队登记，然后排队乘大巴去 party 场地。第三天GHC 官方 party 就安排在主会场，凭借注册时发的票入场。也听说了 Twitter 等其他公司有 party，有兴趣参加的小伙伴可以尝试去 career fair 的各公司展台直接提问。</span><br></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 640 18.8" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" width="640"><g style="box-sizing: border-box;"><path style="box-sizing: border-box;" d="M1.8,15c9.6,0,19.2,0.1,28.8,0.2l0.2-3c-9.7-0.1-19.3-0.2-29-0.2C-0.6,12-0.6,15,1.8,15z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M75.7,13.1l-25.4-0.5l-0.2,3l25.6,0.5c3.7,0.1,7.5,0.1,11.2,0.1l0.4-3l-5.8,0L75.7,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M138,13.1l-5.8,0.1l-25.4,0.1h-0.1l-0.3,2.9c10.5,0,21,0,31.6-0.2l3.4-0.1l0-3L138,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M161,15.6c10.9-0.2,21.8-0.4,32.7-0.5c0.1-1,0.2-2,0.2-3c-11,0.1-22,0.3-33,0.5L161,15.6z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M291.5,13.2l-5.8-0.1l-13.2-0.3l0.3,3l12.9,0.3c8.2,0.2,16.3,0.2,24.5,0.2l-0.5-3L291.5,13.2z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M348.1,13.1l-5.8,0.1l-12.8,0.1c0.2,1,0.3,2,0.5,3c6,0,12-0.1,18.1-0.2l20.7-0.4l-0.2-3L348.1,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M422,12c-11.4,0-22.7,0.1-34.1,0.3l0.2,3c11.3-0.2,22.6-0.3,33.9-0.3C424.4,15,424.4,12,422,12z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M213.3,12c0,1-0.1,2-0.2,3c13.4,0,26.7,0.2,40.1,0.4c-0.1-1-0.2-2-0.3-3   C239.7,12.2,226.5,12.1,213.3,12z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M1.8,5.5c9.8,0,19.7,0.1,29.5,0.2l0.2-3C21.6,2.6,11.7,2.5,1.8,2.5C-0.6,2.5-0.6,5.5,1.8,5.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M75.7,3.5L51,3l-0.2,3l24.9,0.5c4,0.1,8.1,0.1,12.1,0.2c0.1-1,0.1-2,0.2-3l-6.5,0L75.7,3.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M138,6.5l3.3-0.1l0-3L138,3.5l-5.8,0.1l-24.7,0.1l-0.2,2.9C117.6,6.7,127.8,6.7,138,6.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M160.8,3.1l0,3c11-0.2,22-0.4,33.1-0.5c0-1-0.1-2-0.2-3C182.8,2.7,171.8,2.8,160.8,3.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M251.9,5.9c-0.2-1-0.4-2-0.6-3c-12.7-0.2-25.5-0.3-38.2-0.3c0.1,1,0.2,2,0.2,3   C226.2,5.5,239.1,5.7,251.9,5.9z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M308.9,6.7c-0.1-1-0.2-2-0.3-3l-17.1-0.1l-5.8-0.1l-14.6-0.3l0.5,3l14.1,0.3   C293.5,6.7,301.2,6.7,308.9,6.7z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M367.9,6.1c-0.2-1-0.3-2-0.5-3l-19.2,0.4l-5.8,0.1l-14.1,0.1c0.1,1,0.2,2,0.3,3   c6.5,0,13-0.1,19.5-0.2L367.9,6.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M422,2.5c-11.7,0-23.4,0.1-35.1,0.3c0.2,1,0.3,2,0.4,3c11.6-0.2,23.1-0.3,34.7-0.3   C424.4,5.5,424.4,2.5,422,2.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M447.8,15c9.6,0,19.2,0.1,28.8,0.2l0.2-3c-9.7-0.1-19.3-0.2-29-0.2C445.5,12,445.5,15,447.8,15z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M521.8,13.1l-25.4-0.5l-0.2,3l25.6,0.5c3.7,0.1,7.5,0.1,11.2,0.1l0.4-3l-5.8,0L521.8,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M584.1,13.1l-5.8,0.1l-25.4,0.1h-0.1l-0.3,2.9c10.5,0,21,0,31.6-0.2l3.4-0.1l0-3L584.1,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M607.1,15.6c10.9-0.2,21.8-0.4,32.7-0.5c0.1-1,0.2-2,0.2-3c-11,0.1-22,0.3-33,0.5L607.1,15.6z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M447.8,5.5c9.8,0,19.7,0.1,29.5,0.2l0.2-3c-9.9-0.1-19.8-0.2-29.7-0.2   C445.5,2.5,445.5,5.5,447.8,5.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M521.8,3.5L497.1,3l-0.2,3l24.9,0.5c4,0.1,8.1,0.1,12.1,0.2c0.1-1,0.1-2,0.2-3l-6.5,0L521.8,3.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M584.1,6.5l3.3-0.1l0-3l-3.2,0.1l-5.8,0.1l-24.7,0.1l-0.2,2.9C563.7,6.7,573.9,6.7,584.1,6.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M639.8,2.6c-11,0.1-21.9,0.3-32.9,0.5l0,3c11-0.2,22-0.4,33.1-0.5C640,4.6,639.9,3.6,639.8,2.6z" fill="rgb(193,154,107)"></path></g></svg></section></section></section></section><p><span style="font-size: 15px;">以上就是 Yangyang 此次参加 GHC 的主要感想啦。</span></p>
<p><span style="font-size: 15px;">当我们收到 Yangyang 的投稿时，真的非常感动。这样长的感受，这样深刻的自我认识，这样无私的分享，无不传递出大会给人带来的积极影响，以及源自她内心深处的能量。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">如果大家看后有所收获，也希望大家积极把这篇文章分享给周围的朋友~</span></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240" style=""></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693282&amp;idx=1&amp;sn=0f19cd6c63c7ea874ead46a78ac92eba#wechat_redirect</link>
      <pubDate>Tue, 09 Oct 2018 18:18:35 -0700</pubDate>
      <guid isPermaLink="false">2291</guid>
    </item>
    <item>
      <title>职场 | 我的十年</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740" style=""></p>
<p style="margin-right: 8px;margin-bottom: 15px;margin-left: 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="color: rgb(136, 136, 136);letter-spacing: 1px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;">08年，</span><span style="color: rgb(136, 136, 136);letter-spacing: 1px;">我刚加入公司的时候，被告知在公司工作十年以上就有资格享受12周的年资假。当时并没有觉得这个福利和自己有什么关系，毕竟太过遥远。</span><br></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">可似乎转眼间，十年就这样悄无声息走到了面前。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6DfOLQMW7bppOPCbqcWu0RZnB9Sibia6HholhXpW0aKoJPa3fo5m3D6QQ/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="940" data-cropy1="0" data-cropy2="528.9667896678966" data-ratio="0.5627659574468085" data-s="300,640" data-type="jpeg" data-w="940" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6Y54P5paia3QnG8iaRYicVb5dUTEuaia4kHvxVlMmicQbzpxOUzK8fn6JLmw/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">十年中，我的角色发生过很多次变化，刚入职的时候是QA（质量保证分析师） ，后来做项目管理，之后因为意想不到的机会转职做People Lead（人力资源总监），从此一发不可收拾，我在转型的路上越走越远。十年来，我尝试过的角色不少于五个，技术、人力、运营、软件交付、咨询，还不包括兼职做下来了社区创立者、全球创新项目负责人以及培训师。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">回首过去的十年，每一次转型都有很大的跨度，也正是因为这些跨度让我这十年很丰富。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有一位前辈曾经和我说过：这家公司给不了你一个看得到的未来，但过程中的未知和不确定会让你终身难忘。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我用十年的时间验证了这句话的正确性，作为一个喜欢做梦，热爱思考，经常发呆也乐于自省的职场女性，有一些关于职场和成长的感悟，想掏心掏肺和大家分享。</span></p>
<h3 style="margin: 20px 8px 15px;font-weight: bold;font-size: 18px;max-width: 100%;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;-webkit-font-smoothing: antialiased;cursor: text;font-family: Helvetica, arial, sans-serif;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;letter-spacing: 1px;color: rgb(0, 0, 0);font-size: 16px;box-sizing: border-box !important;word-wrap: break-word !important;">1 从软件测试到人力资源总监，经历了什么</span></strong></h3>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">当初来ThoughtWorks是因为猎头的推荐，真正让我下定决心加入是因为ThoughtWorks践行的自动化测试，当时我预见自己会在测试领域摸爬滚打好多年，甚至想着测试将会是自己专注一生的领域，因为业内层出不穷的新技术、新工具、新观点让我一直很兴奋，不断涌现的市场需求也给我带来安全感，当时身边也有一些牛人指路，让我看到了上升空间，所以那时候我的志向差不多就是做测试界的百科全书。 </span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但是在我加入三年后，突然有一天，我的Sponsor（导师）过来找我，说有一个机会，问我愿不愿意尝试中国区 People Lead（人力资源总监）这个职位。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">当时这个职位已经空缺几个月了，前任离职后，中国区花了很长时间去招聘，但一直没有找到合适的人。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6q4ia3Qa2ibvIQNVn9XrdSxadZZQwRYILxGick2m1t6PsIaHhTbzcdM2sw/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="1280.0000000000002" data-cropy1="205.46125461254613" data-cropy2="935.2029520295202" data-ratio="0.5703125" data-s="300,640" data-type="jpeg" data-w="1280" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6Kic3ibgibKHba5Qlzn2yqlF9P4xZiaz42nYcW9rDubf3w1TMkyiaxwMFFNg/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">ThoughtWorks特殊的文化决定了对这个角色的要求和普通的人力资源总监不一样，这是一家专业技术公司，绝大多数的员工都是技术人员，作为人力资源总监，只有充分了解技术人员，了解他们的工作，能听懂并且理解他们讲的故事，才能真正发挥同理心，更好的关注到他们，给他们提供帮助并且支持他们的发展。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">中国区决定在内部寻找，而我就是在这个时候被识别出的候选人。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我当时很惊讶，完全没有做好准备，这个转型是我从来没有考虑过的事情。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但是有很多同事过来找我，告诉我他们的想法。</span></p>
<section class="" data-tools="135编辑器" data-id="86005" style='margin-left: 8px;max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.544px;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);border-width: 0px;border-style: none;border-color: initial;box-sizing: border-box !important;word-wrap: break-word !important;'><section class="" style="margin-bottom: 16px;padding: 16px;max-width: 100%;word-wrap: normal;overflow: auto;font-size: 15px;line-height: 1.45;border-radius: 3px;color: rgb(51, 51, 51);background-color: rgb(247, 247, 247);box-sizing: border-box !important;"><p style="margin: 5px 8px;max-width: 100%;min-height: 1em;-webkit-print-color-adjust: exact;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-family: Helvetica, arial, sans-serif;font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有人和我说，机会不错，能学到很多东西；</span></p>
<p style="margin: 5px 8px;max-width: 100%;min-height: 1em;-webkit-print-color-adjust: exact;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-family: Helvetica, arial, sans-serif;font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有人和我说，在这个职位上能和更多视野高的人合作，成长快；</span></p>
<p style="margin: 5px 8px;max-width: 100%;min-height: 1em;-webkit-print-color-adjust: exact;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-family: Helvetica, arial, sans-serif;font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有人和我说，试一下吧，失败了可以重来，年轻有什么可怕的；</span></p>
<p style="margin: 5px 8px;max-width: 100%;min-height: 1em;-webkit-print-color-adjust: exact;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-family: Helvetica, arial, sans-serif;font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">还有人和我说，这是ThoughtWorks的风格，鼓励每个人挑战意料之外的事情。</span></p></section></section><p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">真正影响我做决定的是这句话：多年以后，你可能不会后悔自己做过什么，但是你可能会后悔自己没有做过什么。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">总之，没有一个人告诉我放弃。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">直到现在，我还是很喜欢ThoughtWorks保留下来的这种文化，有时看似是一个人的事情，但会有很多人过来提供建议，他们不会干预你的决定，但是他们会毫无保留的和你分享他们的想法。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">他们所有的人都鼓励你勇敢，尝试新的事情，他们容忍失败，认可每个人突破自己。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">很多年以后，我发现自己成长最快的时候都是从做大胆决定那一刻开始的。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6gUuZEFLwJ8CmV1OPYoCLKoibg2YIApq5kvKYzqcVgk9BaskcjyrVCqg/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="1050" data-cropy1="0" data-cropy2="534.4086021505376" data-ratio="0.5095238095238095" data-s="300,640" data-type="jpeg" data-w="1050" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6eiamo2xc5HSCLECM2LTGDV7l1ZKibSuaMFPllVcKmlEkDbLnOias3n1NA/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">在新的岗位上，我接手的第一件事情就是给员工销假，我怎么也没有想到，本来是怀着为人民服务的理想踏上岗位的，但上岗之后做的第一件事情竟然是做“恶人”，赤裸裸“剥夺”员工的利益，“剥夺”以前和我紧密合作的同事的利益。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那时候公司的规模开始增长，但是各种制度还不健全，有的制度还成了公司发展的硬伤，比如休假制度，以前的规定是员工可以无限期延长自己的年假，有的老员工累积的年假已经超过50天，这在财务上是一种负债，对于一家专业服务公司来讲，这种负债越大，公司运营的风险就越高。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">当时中国区还不盈利，一直是全球的投资在支持，所以即使销掉员工的一部分假期，也没有办法按照百分百的比例给予补偿。对于假期累计很多天的同事来讲，这是难以接受的事情，包括我自己。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">在职场上，甚至在人生的某些个阶段，你可能都有机会遇到这样的事情，理性告诉你这样去做，但是情感上你说服不了自己，或者站在一个立场上你是在做正确的事情，但换个角度你就罪大恶极，这种问题没有答案，考验你的是如何在获得他人理解的基础上解决问题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">骨头非常硬，但还是要啃。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我和所有受影响的同事一对一沟通，解释上下文，解释为什么做这个决定，倾听他们的抱怨，接纳他们的情绪，有一位老外同事在我面前拍了桌子，指着我的鼻子发泄他的愤怒。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那段时间，每个晚上我都失眠，感觉自己站到了喜欢的那些人的对立面。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那时候，我想到了放弃或者离职，但我上岗不过才一个星期。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-copyright="0" data-ratio="0.553125" data-s="300,640" data-type="jpeg" data-w="1280" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6qAQibH6ZlADckvoazWngPDVH6RvIc70yZohrQFxqfaic4GQG3g8gk83w/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">后来想想，那其实就是不舒适区，突然跳进去，没有半点过渡，我由一个关注事情的人突然变成要关注人，有太多的技能我需要学习，沟通能力，影响力，自我管理等都需要提升。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">没有办法的话，死扛着也会有转机。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我完成了所有的谈话，也完成了设定的目标，这中间有些人让我很感动，他们理解这个决定也非常配合我推进这件事情，我知道他们是基于对我的信任。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">后来，那个老外过来和我道歉，在他理解了之后，还帮我推动了一些事情。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那件事情结束之后，我就开启了疯狂补课的模式，学习人力资源的系统知识，做各种主题阅读，参加研讨会学习行业经验，考取了人力资源从业资格证，我还给在校大学生做了一百个小时以上的职业咨询，在几个人力资源平台做签约作者。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">真正到了一年以后，我才敢和别人说，我是人力资源从业者，尽管那个时候我仍然在不舒适区，因为我还有很多东西需要学习，我也还有很多解决不了的问题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但时间是不等人的，新的机会总是会在你还没有准备好的时候就出现了。</span></p>
<h3 style="margin: 20px 8px 15px;font-weight: bold;font-size: 18px;max-width: 100%;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;-webkit-font-smoothing: antialiased;cursor: text;font-family: Helvetica, arial, sans-serif;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;letter-spacing: 1px;color: rgb(0, 0, 0);font-size: 16px;box-sizing: border-box !important;word-wrap: break-word !important;">2 再一次归零，失败了会怎样</span></strong></h3>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">2013 年，公司在全球的领导团队发生结构调整，时任中国区MD（中国区总经理）的郭晓调任全球CEO，中国区的其他运营角色也随之发生变动，我由People Lead 岗位转为PSM（专业服务资源管理）。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这是一个为所有项目匹配资源的角色，前提是要对中国区所有的专业人员和客户项目有充分的了解。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">项目匹配的复杂度之高是难以想象的，因为项目的需求各不一样，客户领域不一样，工作地点不一样，周期和起始时间不仅不一样还不断动态调整，相应的，在人员的分配上，不但要考虑每个人的技能，还要考虑机会是否符合个人发展意愿，是否需要出差协调家庭。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">归根结底，这份工作协调的是供求关系，困难的是供和求都在动态调整。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">不仅如此，这个工作的边界也很难定义，从密切关注商机预测，到把供求关系转化为培训计划，招聘优先级和启动全球支持，都需要参与。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那个时候中国区已经进入了一个快速发展的通道，规模的扩大已经带来了一些问题，以前是一个人处理的事情，现在需要一个团队协作来完成。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">在新的岗位上，我第一件事情就是组建团队。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-copyright="0" data-ratio="0.6666666666666666" data-s="300,640" data-type="jpeg" data-w="1050" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6WhlUyhzLn0AqCXNN168ws35RM4Ocuq9J8NKTW198ViaC1ibszNPccibUQ/640?wx_fmt=jpeg" style="background-color: rgb(238, 237, 235);border-width: 1px;border-style: solid;border-color: rgb(238, 237, 235);background-size: 22px;background-position: center center;background-repeat: no-repeat;height: 440.667px !important;box-sizing: border-box !important;word-wrap: break-word !important;width: 661px !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">团队里面的第一位RM（资源协调经理），是我们从业务部门识别出来的，我们俩在没有前人经验的基础上，摸着石头过河，与人力资源比起来，这个角色不属于任何领域，市场上也找不到相关的行业经验，所以在这个岗位上，发展空间难以预测，职业安全感相对也较低。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">三个月后，我唯一的团队成员提出了辞职，我像一个创业失败的人重新回到起点，那段时间，我自己也曾想着放弃，觉得自己不擅长从零做到一，那是在那个职位上最迷茫的一个阶段。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">后来在中国区总经理胡凯的帮助下，我去了印度，当时印度是唯一一个有RM团队的国家。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我加入了印度的RM团队，花了三个月的时间和他们一起处理印度的资源调配，和每个RM以及业务部门密切合作，和他们的专业人员沟通项目详情。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">他们也没有成熟的经验，但是我还是学到了很多东西，团队是如何协作的，有什么流程，在规模扩大的时候可能会面临一些什么样的坑。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">现在想想，我之所以坚持了下来，是看到了这家组织对失败的宽容，要想改变，就不可能一帆风顺。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">而且当时整个中国区的管理团队也都比较年轻，在没有人指路的情况下，大家是在一起战胜困境，探讨可能的。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我从印度回来后，立即开始组建团队，半年后，每个大的办公室都有了专职的RM，小的办公室也有兼职协调资源的人。在这个过程中，我也有了一些思考，因为工作内容复杂多变，所以很难给从事这份工作的人做严格的职责定义，但这份工作实践起来却非常锻炼人的抗压能力，应变能力，适应性和敏感性。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">更为重要的是，这份工作能让人看清专业服务公司是如何运作的。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">若干年后，我再次进入业务部门，和一家战略咨询公司合作，在一次吃饭的闲谈中，我理清了他们团队背后的配置逻辑，也帮他们解答了专业服务公司如何配置资源的问题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6zyXZvGnLcqWG13loC1dao97gRWU8BiaEot0tvoLeuBpjfkxrpUodEdg/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="1280" data-cropy1="0" data-cropy2="781.6974169741698" data-ratio="0.61015625" data-s="300,640" data-type="jpeg" data-w="1280" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6H7aWGg9T59wnicFG5RjpJFhTQEJygiavo9Go76Se1XyvEnKxX3UqjSxQ/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那时候想到的是《法华经》里的一个词“功不唐捐”。我们学到的知识，做过的事情，都会在未来的某一天以经验的方式回馈给我们。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">现在，Staffing（资源协调）这个团队仍在公司担任着非常重要的运营职责，这个团队也走出来一些很出色的leader（带头人和管理者），想起来是件很欣慰的事情。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我在这个岗位做了三年，直到后来休产假。</span></p>
<h3 style="margin: 20px 8px 15px;font-weight: bold;font-size: 18px;max-width: 100%;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;-webkit-font-smoothing: antialiased;cursor: text;font-family: Helvetica, arial, sans-serif;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;letter-spacing: 1px;color: rgb(0, 0, 0);font-size: 16px;box-sizing: border-box !important;word-wrap: break-word !important;">3 做不到家庭和工作平衡的话，就做一个无所畏惧的女性</span></strong></h3>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">还在做People Lead的时候，有一天，有位女同事找我倾诉，她的孩子那时不满一周岁，她说：</span></p>
<section style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;box-sizing: border-box !important;word-wrap: break-word !important;'><blockquote label="Copyright Jason Ng of Kenengba and knbknb" style="margin-right: 30px;margin-left: 32px;max-width: 100%;border-left-width: 5px;border-left-color: rgb(38, 189, 207);line-height: 25.6px;box-sizing: border-box !important;word-wrap: break-word !important;"><p style="margin-right: 5px;margin-bottom: 15px;margin-left: 5px;max-width: 100%;min-height: 1em;white-space: pre-wrap;line-height: 1.65em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;font-family: Helvetica, arial, sans-serif;font-size: 14px;box-sizing: border-box !important;word-wrap: break-word !important;">我加班的时候觉得对不起娃，在家里陪娃又觉得对不起我的工作，做女性为什么这么难，做什么都feel guilty（带着负罪感）。</span></p></blockquote></section><p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这就是我当时还理解不了的处境。在几年之后我理解了，因为我也成了母亲。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">当了母亲之后，就特别关注工作生活平衡的话题，这可能也是绝大多数职场女性回避不了的话题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">《奇葩大会》有一期，女强人刘楠和 5 位男性CEO一起被采访，男性CEO被问到的问题，都是关于用户，关于利润的，到了刘楠，她被问到的问题是，你是如何平衡工作和生活的？</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这揭示了当下社会和职场的一个普遍认知，男性只要管好自己就行了，女性则需要平衡事业和家庭。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">产假回来，我再一次转型，调到内部IT 部门做项目管理，共有5 个团队，服务全球。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这个转型的决定其实也是考虑了我产后回归，更好的平稳过渡而做出的。这个工作在内容上我已经很熟悉，但实际中仍然需要克服一些问题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">最大的问题就是时差，在一个全球的团队里，很多的会议不是发生在清晨就是发生在晚上，这也就预示着在家里开会是一个常态，而我家的小朋友很小，还是听不进任何道理的年纪。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">开会的时候我通常把自己关在屋里，任凭他在外面捶打着门撕心裂肺哭，我也不能分身，但这会极大的影响我的注意力和效率，我觉得私人时间里陪伴不了他，对不起小娃，开会的时候做不到全神贯注，又愧疚于信任我的同事。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">生活就是个矛盾的综合体。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6gkGxMyoU9zDTRIaah8PJVTDuViaicvcS3n7qTziaU8K0vnLjtjYvok4ibQ/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="1280" data-cropy1="40.147601476014756" data-cropy2="805.3136531365313" data-ratio="0.596875" data-s="300,640" data-type="jpeg" data-w="1280" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6g7ypE3Ec27YoYEItcIBLeLFdsoKOL1Vnj6vV9mElyIiaw2KbWE9DBsw/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">第二个问题就是个人发展，因为孩子小，在养育中必然会分心，所以我其实是做了一个放缓自己发展的决定。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这是一个取舍的过程。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">Facebook的首席运营官谢丽尔.桑德伯的畅销书《向前一步》里介绍了很多方法，鼓励女性冲破性别障碍，做好工作生活平衡。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有些女性觉得这本书确实给她们带来了改变，但也有一些女性做不到，在兼顾工作和家庭的同时，她们既忙碌又焦虑，因为大部分的女性之所以家庭和事业都美满，是因为自己的另一半或者家里的老人们在为她们负重前行。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">也就是说，所谓的平衡，不是自己做出妥协，就是家人做出牺牲。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">就连桑德伯自己也在书里提到，她能在职场打拼，没有后顾之忧，是因为她的丈夫分担了很多家庭责任。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这其实是一种方法，或许是最能支持到女性在职场发展的方法了。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">前面提到我所在的团队的特殊性，晚上开会的时候如果不是特别晚，七八点钟的会议我大都选择在公司开，这样的话我就可以在五六点钟的时候去健身房跑跑步练练器械，回到公司正好赶上开会，回家的路上我会觉得精神焕发、动力十足，这段时间是我的家里人在承担责任。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但周末的时候，我会全心全意陪伴孩子，陪伴家人。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">所以，在家庭工作平衡这件事情上，我是比较幸运的，既放飞了自我，也没有远离家庭，因为我有非常支持我的家里人。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但对于一些女性来讲，可能并不会如此幸运，网络上有很多建议，我比较认可的做法是这样的：</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">如果真的无法平衡，那就选一个自己最想成为的角色，起码是现阶段最想成为的角色。</span></strong><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但是不管你选了哪一个，都意味着你要放弃另一个，不要怕，也不要担心，做一个无所畏惧的女性。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有时想想，我们的人生就像坐标系，横轴已经订好了，区别在于纵轴上如何发挥，十年可以很短，转瞬即逝，十年也可以很长，因为成长永无止境。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">最后，18年伊始，用送给自己的话来结束吧：千帆过尽，愿自己仍是那个明眸善睐的女子，永葆一颗少年心！</span></p>
<p style="margin: 0em 8px 20px;max-width: 100%;min-height: 1em;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;font-family: 微软雅黑;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">本文转载自ThoughtWorks公司。<br></span></strong><span style="color: rgb(0, 0, 0);"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">作为「全球最佳女性科技雇主」ThoughtWorks向来不拘泥于常规而致力于打破框架，突破极限，寻求机遇。程序媛的日常在去年也曾经跟ThoughtWorks合作过</span></strong></span></strong></span><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692638&amp;idx=1&amp;sn=27ab27d01c75acb08683a1a91fc03ed1&amp;chksm=81233927b654b0311fe2d060ddd9fc8808157115986ffbde22ca5989425189908a487ccf6487&amp;scene=21#wechat_redirect" target="_blank">线下活动 | 全球最佳女性科技雇主 ThoughtWorks 的秘密</a></span></strong></span></strong><span style="color: rgb(0, 0, 0);"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">。</span></strong></span></strong></span></p>
<p style="margin: 0em 8px 20px;max-width: 100%;min-height: 1em;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;font-family: 微软雅黑;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(0, 0, 0);max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">ThoughtWorks 希望验证女性能在技术领域承担一些非常关键的岗位。但是现在行业里并没有许多这样的女性技术人员，所以需要自己培养，所以 ThoughtWorks 开展这些活动和方式来自己培养。这是远大的 ambitious vision。</span></strong></span></strong></p>
<p style="margin: 0em 8px 20px;max-width: 100%;min-height: 1em;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;font-family: 微软雅黑;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(0, 0, 0);max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">说起短期目标，就是因为 ThoughtWorks 要求每年校招进来的都是男女比1：1，这很难，只能自己先尽早培养。而个人认知——靠个人，在这一点上，ThoughtWorks 和我们 GirlsWhoCode 是非常相似的。我们都希望让大家认识到女性程序员也是多姿多彩的，不是千篇一律的，不是形象差、目光呆滞的。</span></strong></span></strong></p>
<p style="margin: 0em 8px 15px;max-width: 100%;min-height: 1em;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;font-family: 微软雅黑;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="color: rgb(0, 0, 0);"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">想要进一步了解吗？</span></strong></span></strong></span><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">ThoughtWorks校招火热进行中，详情戳“阅读原文”，欢迎大家投简历～推荐人写“大S”据说有惊喜哟！</span></strong></span></strong></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.4514877102199224" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNpKic4DiblDA75KduXD0EeqoZ8McBkcR5DUfib3SQ5IiaNYpzLW98Valvic1XIQmFEIbibgFwZUPNb69vcg/640?wx_fmt=png" data-type="png" data-w="773" style=""></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693273&amp;idx=1&amp;sn=7cbc76c4277172e260b00d1269d75ad1#wechat_redirect</link>
      <pubDate>Wed, 12 Sep 2018 02:47:03 -0700</pubDate>
      <guid isPermaLink="false">2290</guid>
    </item>
    <item>
      <title>不用 KL annealing，将注意力机制建模为隐变量</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p><em><span style="font-size: 15px;">作者：Harvard NLP group (harvardnlp)</span></em></p>
<p><em><span style="font-size: 15px;">论文：</span></em><em><strong><span style="font-size: 15px;">Latent Alignment and Variational Attention</span></strong></em></p>
<p><em><span style="font-size: 15px;"><span style="font-size: 15px;">地址：</span></span></em><span style="font-size: 15px;"><span style="font-size: 15px;"><span style="font-size: 15px;">https://arxiv.org/pdf/1807.03756.pdf</span></span></span></p>
<p><em><span style="font-size: 15px;"><span style="font-size: 15px;">代码：</span></span></em><span style="font-size: 15px;"><span style="font-size: 15px;">https://github.com/harvardnlp/var-attn </span></span></p>
<p><em><span style="font-size: 15px;">摘要：</span></em><span style="font-size: 15px;">attention 注意力模型在神经网络中被广泛应用。在已有的工作中，attention 机制一般是决定性的而非随机变量。我们提出了将 attention 建模成隐变量，并应用 VAE 和 policy gradient 训练模型。在不使用 KL annealing 等 trick 的情况下训练，在 IWSLT 14 German-English 上建立了新的 <span style="font-size: 15px;">state-of-the-art</span>。</span></p>
<p><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 10px;margin-bottom: 10px;text-align: center;box-sizing: border-box;"><section class="" style="border-bottom: 1px dotted rgb(160, 160, 160);padding-top: 3px;width: 100%;margin-bottom: -3px;box-sizing: border-box;"></section><section class="" style="height: 5px;line-height: 5px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;background-color: rgb(255, 255, 255);box-sizing: border-box;"><span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> </section><section style="clear: both;box-sizing: border-box;"></section></section></section></section></section><p><span style="font-size: 15px;">近年来很多文章将 VAE 应用到文本生成上，通过引入隐变量对文本中的一些不确定性（diversity，如文章风格主题、蕴含情感等）进行建模。<strong>这样做往往会遇到一个常见的问题—— KL collapsing。</strong>这个问题最早在16年时由 Bowman 指出 [1]，其描述的现象是直接训练 VAE 得到的 KL 接近0——这也就意味着近似后验和先验一样，使得隐变量被模型忽略 [5]。Bowman的解决办法 [1] 是使用 KL annealing（KL 项的权重从0开始逐渐增加到1）或者 word dropout（不常用在此略过）。随后，17年 Yang 等人对 KL collapsing 的问题进行了更细致的分析 [2]，并提出降低 decoder 的 contextual capacity 改善这个现象。此外 Zhao 等人提出 bag-of-word loss 去解决这个问题 [3]。18年 Graves 等人也对 KL collapsing 进行了分析 [4]。（我们在过去也介绍过许多解决 KL collapsing 的工作，<a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692981&amp;idx=1&amp;sn=37ed31cc6ec1ded76a53a26503ba4f93&amp;chksm=8123384cb654b15a6582beb3efb3bae1f350a12077456aded80ddb3d1993be4354bde3fc16a3&amp;scene=21#wechat_redirect" target="_blank">点此查看</a>）。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">在我们的工作中，attention 被建模成隐变量。</span><strong><span style="font-size: 15px;">值得注意的是，我们将 attention 建模成隐变量并不是为了单纯应用 VAE 这个工具，而是因为我们认为将 attention 建模成隐变量可以为 decoder 提供更 clean 的 feature，从而在不增加模型参数的情况下提高模型的表达能力</span></strong><span style="font-size: 15px;">（注意 inference network在测试阶段不被使用因此不计入模型参数）。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><strong><span style="font-size: 15px;">以下是一个简单的直觉。</span></strong><span style="font-size: 15px;">下图蓝色部分展示的是传统 attention，<span style="font-size: 15px;">下图红色部分展示的我们提出的隐变量 attention</span>。传统的 attention 机制仅能通过之前生成的单词确定当前即将生成单词的 attention，而因为存在多种翻译方式，所以会出现 attend 到和实际翻译的单词并不对应的位置的问题。而在红色部分展示的我们提出的隐变量 attention 中，我们可以通过全部的源文本和目标文本去得到更准确的后验 attention，因此 attention 和实际翻译应该 attend 的源单词对应得更好。并且，这样得到的更好的后验 attention 可以提供给 decoder，从而使 decoder 拿到更 clean 的 feature，藉此可以得到更好的模型。</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.5475285171102662" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3nsr4p7W1ibTTk2Qn5F3HI9HyiahcvT9OzUWyrsS9Xf6PvLsRjic9r1Jvg/640?wx_fmt=png" data-type="png" data-w="526" style="width: 411px;height: 225px;"></p>
<p><span style="font-size: 15px;"> </span></p>
<p><strong><span style="font-size: 15px;">基于这个直觉，我们将注意力 attention 建模成隐变量。</span></strong><span style="font-size: 15px;">假定 x 是源文本，y 是目标文本，z 是 attention，根据标准的 VAE 方法，我们引入 inference network q(z | x, y)去近似后验，那么 ELBO 可以表达为（为了简单我们考虑目标文本只有一个单词的情况）：</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.05363528009535161" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3N6qmf6mmcfHJVYf16x6vTcdaZfdDBVQTXS9rcOGkQyqnoia4icS1mqIw/640?wx_fmt=png" data-type="png" data-w="839" style="width: 435px;height: 23px;"></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">上面不等式的右侧是 ELBO，其中第一项是从 q(z | x, y) 中采样出 attention，使用采样出的 attention 作为 decoder 的输入计算 cross entropy loss，第二项是确保后验分布接近先验分布。这里值得注意的是，我们此处的先验和一般的 VAE 不同，我们的先验是和模型一起学习的。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><strong><span style="font-size: 15px;">因为我们的 p(z | x) 和 q(z | x, y) 都是 categorical 分布，所以我们使用 policy gradient 去优化上面的目标函数。</span></strong><span style="font-size: 15px;">由于 decoder 和 encoder 之间的主要信息传输通道是通过 attention，如果忽略了这个隐变量，就会无法得到源文本的信息而得到很大的 penalty。这与之前的许多工作中直接把隐变量加入到每个 decoding step 不同，因为那样即使 decoder 忽略了隐变量，也可以达到很好的模型表现 [5]。因此通过直接优化目标函数这个隐变量也不易被忽略，我们的实验完全验证了这一点。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 10px;margin-bottom: 10px;text-align: center;box-sizing: border-box;"><section class="" style="border-bottom: 1px dotted rgb(160, 160, 160);padding-top: 3px;width: 100%;margin-bottom: -3px;box-sizing: border-box;"></section><section class="" style="height: 5px;line-height: 5px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;background-color: rgb(255, 255, 255);box-sizing: border-box;"><span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"><span style="font-size: 15px;"> </span></span> </section></section></section></section></section><p><strong><span style="font-size: 15px;">由于我们的后验 q 能看到全部的 x 和 y，因此后验中采样的 attention 可以比先验 p(z | x)好。</span></strong><span style="font-size: 15px;">比如以下的例子：</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.9883527454242929" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3ic0T01mM6eYlRtIh1IFcDOIUXadFhbYl8ia8vVXovmdL1pO8QnQKHqcg/640?wx_fmt=png" data-type="png" data-w="601" style="width: 466px;height: 461px;" data-backw="233" data-backh="230" data-before-oversubscription-url="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3ic0T01mM6eYlRtIh1IFcDOIUXadFhbYl8ia8vVXovmdL1pO8QnQKHqcg/640?wx_fmt=png"></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">这里我们把德语（纵向）翻译成英语（横向）。红色代表先验，即只观测到 x 而不观测到 y 的 p(z | x)，蓝色代表后验，即观测到全部信息的 p(z | x, y)。注意到在第二个单词 <em>actually</em> 处，红色的先验试图 attend 到 <em>nun</em> 后面的逗号<em>“,”</em>，从而试图生成一个 <em>“well, ” </em>的翻译结果。然而实际的英语翻译中并没有逗号，反而直接是 <em>well actually</em>，由于后验 q(z | x, y) 可以看到实际的翻译，因此蓝色的后验正确 attend 到了 <em>tatsachlich </em>上。注意到训练目标 ELBO 中我们从 q 中采样 attention 给 decoder，因此通过使用 VAE 的方法，decoder 得到了更准确的 attention 信号，从而可能提高模型的表达能力。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 10px;margin-bottom: 10px;text-align: center;box-sizing: border-box;"><section class="" style="border-bottom: 1px dotted rgb(160, 160, 160);padding-top: 3px;width: 100%;margin-bottom: -3px;box-sizing: border-box;"></section><section class="" style="height: 5px;line-height: 5px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;background-color: rgb(255, 255, 255);box-sizing: border-box;"><span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"><span style="font-size: 15px;"> </span></span> </section></section></section></section></section><p><strong><span style="font-size: 15px;">实验上，我们在 IWSLT 14 German-English 上得到了新的 state-of-the-art。</span></strong><span style="font-size: 15px;">其中 KL大约在0.5，cross entropy loss大约在1.26，而且我们人工比较了很多后验和先验也很符合我们的建模直觉。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4455909943714822" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3PtSJQPTvxjMkpcibgzTqDALv1gspl8eXiahZkrAUmwdURObzvvefOyHA/640?wx_fmt=png" data-type="png" data-w="1066"></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">欢迎尝试我们的代码，项目地址在 https://github.com/harvardnlp/var-attn。我们提供了能复现我们 state-of-the-art 效果的 preprocessing、training、evaluation的command，以及我们报告的模型。</span><br></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><strong><span style="font-size: 15px;">相比过去的大部分工作是<span style="font-size: 15px;">从 attention 计算出来的固定维度的 context vector，</span><span style="font-size: 15px;">我们提出了将 attention 建模成隐变量，即在 simplex 上的 attention 本身。</span></span></strong><span style="font-size: 15px;"><span style="font-size: 15px;">由于</span></span><span style="font-size: 15px;"><span style="font-size: 15px;"><span style="font-size: 15px;">我们的工作是对 attention 注意力机制的改进，</span></span>因此理论上可以被应用到一切包含 attention 的 task 中。文章里除了机器翻译外我们也做了个视觉问答系统的实验。我们的具体模型和 inference network 的结构请参见我们的论文和代码。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">限于作者的水平，本文中有错误和纰漏在所难免，望读者朋友多多包涵！也欢迎发邮件给我 dengyuntian@seas.harvard.edu 交流。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><em><strong><span style="font-size: 15px;">References </span></strong></em></p>
<p><span style="font-size: 15px;">[1] Bowman et al, "Generating Sentences from a Continuous Space".</span></p>
<p><span style="font-size: 15px;">[2] Yang et al, "Improved Variational Autoencoders for Text Modeling using Dilated Convolutions".</span></p>
<p><span style="font-size: 15px;">[3] Zhao et al, "Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders".</span></p>
<p><span style="font-size: 15px;">[4] Graves et al, "Associative Compression Networks for Representation Learning".</span></p>
<p><span style="font-size: 15px;">[5] Zhang et al, "Variational Neural Machine Translation".</span></p>
<p><br></p>
<p><strong><span style="font-size: 15px;">更多相关文章阅读</span></strong></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><span style="font-size: 15px;"> <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692981&amp;idx=1&amp;sn=37ed31cc6ec1ded76a53a26503ba4f93&amp;chksm=8123384cb654b15a6582beb3efb3bae1f350a12077456aded80ddb3d1993be4354bde3fc16a3&amp;scene=21#wechat_redirect" target="_blank">干货 | 你的 KL 散度 vanish 了吗</a></span></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692433&amp;idx=1&amp;sn=b880134068f670cfa59a81b59d432045&amp;chksm=81230668b6548f7eb5bf73a696b5e918d0486e96dca8df55740a947df827d2263e951bcb2136&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">ACL 2017 | VAE for Dialog Response Generation</span></a></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692200&amp;idx=1&amp;sn=5211ddee37bc3a3cad0da93641a7081f&amp;chksm=81230751b6548e476ec2e2c91d023dfd1c7213e34dd085fa57a3f2aa4bd3253920a57aeb5281&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">ICLR 2017 | Attention 和 Memory Networks</span></a></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=212457976&amp;idx=1&amp;sn=ab503fb91c6a126fbe767f0dd37fb48a&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">干货 | Applied Attention-based Models in NLP（一）</span></a></p></li>
<li><p><span style="font-size: 15px;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=212530824&amp;idx=1&amp;sn=55ca55ddb7ce7d2cd9b316c598b6da03&amp;scene=21#wechat_redirect" target="_blank">干货 | Applied Attention-based Models in NLP（二）</a><br></span></p></li>
</ol></blockquote>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693269&amp;idx=1&amp;sn=c281f9ad65a38ce43ce88c578c12bc74#wechat_redirect</link>
      <pubDate>Fri, 20 Jul 2018 18:05:14 -0700</pubDate>
      <guid isPermaLink="false">2289</guid>
    </item>
    <item>
      <title>CVPR 2018 | 加速模型收敛的新思路（控制理论+深度学习）</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p><span style="font-size: 15px;"></span>今天要分享的论文是一篇 <span style="font-size: 15px;">CVPR 2018 Spotlight，</span><strong><em><span style="font-size: 15px;">A PID Controller Approach for Stochastic Optimization of Deep Networks</span></em></strong><span style="font-size: 15px;">. 荣幸的是，我们请到了这篇论文的第一作者 Wangpeng An 为大家带来第一角度解读：从想法的萌生，到投稿的坎坷——走近一篇 CVPR 论文的诞生。 </span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"></section></section></section></section></section></section><p><span style="font-size: 15px;">具有近百年历史的“比例-积分-微分”控制方法（简称 PID 控制器）是一种通过控制系统的误差来调整机器的反馈控制方法。它是迄今为止自动控制中应用最为广泛的控制方法。这个算法被改进了将近200年无人撼动，几乎所有自动化的设备都是使用 PID 算法进行控制。那么它是怎么来的呢，请看下图：</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1.334140435835351" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGsEaLM0CT1o95dCfia0YJZtf8Oibb9ulglhEUDKIzzeHibqGU3bBic8qvJw/640?wx_fmt=png" data-type="png" data-w="413" style="width: 239px;height: 319px;"></p>
<p><span style="font-size: 15px;">这个东西就叫离心机，控制给引擎送多少油，刚开始被用在怀特兄弟造的飞机 [6]。从研究离心机如何供油，控制理论慢慢地被运用到各种各样的机器控制中，<strong>并且引导了第三次工业革命：自动化——如何控制机器让其离开人自动的运行（这个特性好像跟人工智能是一样的）。</strong>如何控制机器的算法就叫控制理论，而 PID 一直在工业控制里占统治地位，它被广泛的应用在工厂 [1]、无人机 [2] 以及无人车 [3] 中。到目前为止，90%的控制器都是 PID。<br></span></p>
<p><br></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.5093085106382979" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaG8gcswWdN92oLDARBKHUAaeIAonibezXyibtZSFec4V3MhwIny6CZsQSg/640?wx_fmt=png" data-type="png" data-w="752"></p>
<p style="text-align: center;"><span style="font-size: 15px;color: rgb(136, 136, 136);">大家熟知的大疆无人机也不例外</span></p>
<p><span style="font-size: 15px;"> </span><br></p>
<p><strong><span style="font-size: 15px;">下面就来说说控制理论和深度学习优化的关系。</span></strong><span style="font-size: 15px;">在控制理论里，<span style="font-size: 15px;">“误差”即</span>受控机器的输出与期望值之间的距离。如何根据“误差”让机器做出下一步更新的算法就是控制器。<br></span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">对于深度学习优化，将模型输出和期望输出的误差来定义为损失函数。损失函数通过执行“误差的反向传播” [1] 得到梯度来影响权重。就像 Hinton 老先生在其发表在 Nature [1] 的文章里那样描述的，梯度是每个权值对总的损失函数贡献，其与损失函数为线性关系。</span></p>
<p><br></p>
<p><strong><span style="font-size: 15px;">把深度学习里的”梯度”看成是当前权重与最优权重的“误差”，这样 SGD-Momentum就与 PID 优化器通过”梯度”与“误差”的联系而联系到了一起。</span></strong><span style="font-size: 15px;">如下图，可以看到控制过程与深度学习的优化过程的联系，可以发现控制系统里的反馈、误差、PID控制器与深度学习里的反响传播、梯度、优化器有着很高的相似度。</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.563076923076923" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGVOu2CzyTIjicEa8HsAicnEJ3DeeFUBYpJb68iczwiaFY8WNdOJPmyu8BPQ/640?wx_fmt=png" data-type="png" data-w="975"></p>
<p style="text-align:center;"><br></p>
<p><span style="font-size: 15px;">如果把控制里的误差用梯度替换掉的话，那么深度学习优化器 SGD-Momentum 和反馈控制器 PID 就有如下的关系：</span><br></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.9282945736434108" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGFwH8EqOMTdRUo8ibcXntvS2KT2wicc5RPH47QNOT7KA30MWwrqhJTTLQ/640?wx_fmt=png" data-type="png" data-w="516" style="width: 379px;height: 352px;"></p>
<p><span style="font-size: 15px;"> </span><br></p>
<p><strong><span style="font-size: 15px;">这样 SGD-Momentum 可以等效为 PI 控制器。</span></strong><span style="font-size: 15px;">而在控制理论中，PI 控制有超调的问题，也就是说 SGD-Momentum 有超调问题，这一点其实很容易理解，因为 I（Integral）是历史梯度的积累。也就是说如果历史梯度朝着同一个方向变化的话，那么朝着该方向的变化就会不断地加速，但是如果要改变方向，那么由于历史梯度还是朝着相反的方向变化，就会产生滞后反应或者也叫超调（Overshoot）的问题。</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.52" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGc7kmIkTMZISWbRALlLNsCLvkSPEebd5nauic3cqzoDEiaJ56bMZuzFEQ/640?wx_fmt=png" data-type="png" data-w="600"></p>
<p><span style="font-size: 15px;">超调也可以简单的描述为在该改变方向的时候还是沿着之前错误的方向继续行进，也就是执着于“过去”。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><strong><span style="font-size: 15px;">而 PID 控制算法中的 D 的这一项就专门是用来较小超调这个问题，因为 D 是梯度的变化趋势，其放眼“未来”。</span></strong><span style="font-size: 15px;">下面的 gif 是我见过解释 PID 这三个参数意义最好的一个动画。从中可以看出 PID 控制器中 PID 前的三个超参数如何根据误差从初始值（0）收敛到最优值（1）的一个动画。</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.75" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3vdJkdoNWSBzUe8ky5nyibNUmmoiccenuTkgaKzB9fWzXuHcAX5V9xVRw/640?wx_fmt=gif" data-type="gif" data-w="400"></p>
<p><span style="font-size: 15px;"></span></p>
<p><span style="font-size: 15px;">根据这个观察我们就顺其自然地提出了 PID 优化器，也就是过去、现在、未来同时存在。过去、现在、未来，究竟哪个比较重要我们不知道，但是我们认为“未来”是必须要有的！</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.34150076569678406" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGAq3jwR2a9ayCbRz2Sg85pNFtP1licibtbniaMPib5d71DwEz0OYhmQbVew/640?wx_fmt=png" data-type="png" data-w="653" style="width: 405px;height: 138px;"></p>
<p><span style="font-size: 15px;"> </span><br></p>
<p><span style="font-size: 15px;">由于我们证明了 SGD-Momentum 其实是 PI 控制器，我们新的算法只在 SGD-Momentum 上加入了梯度的变化这一项。我们也并没有彻底地探索如何更好的设置 P I D这三个参数，为以后的创新也留有广阔空间。：）</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">为了验证我算法能在真实数据上起作用，我们在2维数据上做了可视化的研究。</span></p>
<p><span style="font-size: 15px;">优化一个包含 x 和 y 的实验，目标函数以及结果如下：</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.3815413891531874" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGkqZPiaeRoYUkIVYFHNqezIalveO3sGFCvnomZELQQEVJPic3ta2osFrg/640?wx_fmt=png" data-type="png" data-w="1051"></p>
<p><span style="font-size: 15px;">其中背景的颜色代表f的大小，越红的地方其值越大。黄色的轨迹是 SGD-Momentum的收敛路线，D 前的超参数为 K_d，红色是在不同 K_d 下 PID 优化器的收敛路线。可以从图中看出在同样步长的条件下，K_d 这个超参数可以减小甚至消除超调问题。</span><br></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">然后是在 MNSIT 数据集上的结果:</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.7538265306122449" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGEt2F2TIhZYFSzG2icRibAXNLInicVIxic7k9zPeSKtZflrbGqNwRlD8cTQ/640?wx_fmt=png" data-type="png" data-w="784" style="width: 471px;height: 355px;"></p>
<p><span style="font-size: 15px;">可以看到 PID 优化器明显地好于 SGD-Momentum。</span><br></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"><span style="font-size: 15px;"> </span><br></section></section></section></section></section></section><p><span style="font-size: 15px;">觉得在控制理论里下面的工作中也可以把其误差用梯度替代掉从而提出新的深度学习优化算法：</span></p>
<p><span style="font-size: 15px;">1. 现代控制理论。</span></p>
<p><span style="font-size: 15px;">2. 模糊控制器。</span></p>
<p><span style="font-size: 15px;">3. 模型预测控制。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"><br></section></section></section></section></section></section><p><strong><span style="font-size: 15px;">下面说说做这个工作是如何产生以及最后投稿被接收的：</span></strong></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">其起初就是在折腾 SGD-Momentum 的公式的时候，用了一个高中就会学到的数列求和的技巧化简了一下，发现有一项是历史梯度的求和……在这个过程中不断地把东西往一块凑，不断地丰富实验，丰富理论。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">但截稿前其实没有写完论文，投出去之后发现虽然做了两个实验，然而其中一个结果还忘了贴上去，论文中提到的一个图还并不存在。当时心情复杂，很多师兄师姐都告诉我毕竟是第一次投还只写了7页可以原谅，老师建议我再好好改一下准备投 ECCV。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;"></span></p>
<p class="p1"><span style="font-size: 15px;">在跑 CIFAR 10 实验的时候发现加入梯度的变化那一项并没有 SGD-Momentum 好，当时心情沮丧觉得弄不好做了这么长时间的 idea 可能就是不 work。</span></p>
<p class="p1"><br></p>
<p class="p1"><span style="font-size: 15px;">之后就开始了很长一段时间每天起床，出门，实验室坐下打开 Youtube 和 Github 找事情做的过程。有一次在 Youtube 看到刘媛媛的励志演讲，看完充满了力量异常的有动力。联想起自己做的工作，脑中一闪发现其实并没有人限制我是不是要给梯度的变化这一项前面加的那个超参数必须是设定好不可调节的……于是我就开始了漫长的调参之路。</span></p>
<p class="p1"><br></p>
<p class="p1"><span style="font-size: 15px;"><strong>其实现在也觉得如何面对可能不 work 的 idea 是一门艺术，那种被偶然发现，纯天然无污染的 idea 太少， 大部分可能都是人造 idea。</strong>只要意志坚强，总能做完投出去。</span></p>
<p class="p1"><br></p>
<p class="p1"><span style="font-size: 15px;">最后把她说的一句话送给大家：</span></p>
<p class="p1"><span style="font-size: 15px;"><em>命里有时终须有，命里无时须强求。努力就有 CVPR！</em></span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"><br></section></section></section></section></section></section><p><span style="font-size: 15px;">概括一下，这篇论文（<span style="font-size: 15px;">CVPR 2018 Spotlight，</span><strong style="white-space: normal;"><em><span style="font-size: 15px;">A PID Controller Approach for Stochastic Optimization of Deep Networks</span></em></strong><span style="font-size: 15px;">）</span>介绍了一个简单的<span style="font-size: 15px;">优化方法：</span>修改 SGD + Momentum 使模型更快收敛。其 Motivation 主要来自自动化学科的比例积分微分（PID）控制算法。作者证明了 SGD -Momentum 是一个 PI 控制器（缺乏一个D，与 PID 更新相比方程）。通过加入 D 对 SGD-Momentum 进行改进从而实现加速收敛，而 D 就对应梯度的变化。</span></p>
<p style="white-space: normal;"><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"></section></section></section></section></section></section><p style="white-space: normal;"><span style="font-size: 15px;">这篇论文的 Reviewers 们是这么说的：</span></p>
<blockquote>
<p style="white-space: normal;"><span style="font-size: 15px;">1. I think the paper makes a very interesting connection between two classical approaches, respectively from the field of feedback control and the field of optimization.</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"> </span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">2. The link between PID and SGD+Momentum opens the door to further innovations in optimization of deep networks. I wish the ML community would pick this up and go much further.</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"> </span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">3. I actually believe the authors would do better by performing additional MLP and LSTM experiments, and then submitting this paper to ICLR, NIPS. I don't think CVPR is the right venue and I fear it will not give theauthors the exposure the work deserves.</span></p>
</blockquote>
<p style="white-space: normal;"><span style="font-size: 15px;"> </span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">被多次问过，为什么研究改进 SGD-Momentum 而不是 Adam 呢。因为目前卷积神经网络里最好的模型 VGG、Resnet、DenseNet、Mask-RCNN 等都是由其训练。最近也有人指出 Adam 这种自动化的学习速率方法其实并没有 SGD-Momentum 训练的模型效果好 [2]。简单来说就是 SGD-Momentum 就是一台单反，自己调节一些参数可以达到很好的效果，而 Adam 就像是一台傻瓜相机，虽然大多数时候都行，但追求极致还是得用“单反”。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"> </span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">因为 PID 一直是控制理论中的黑箱，其参数的调节也如炼丹一样的 CNN 至今无人完全参透，所以知乎上也有这样的评论：</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">PID，CNN，双壁合一，但凡练成者，江湖人送外号：“调参圣手”。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"></section></section></section></section></section></section><p style="white-space: normal;"><strong><em><span style="font-size: 15px;">References: </span></em></strong></p>
<p><span style="font-size: 15px;">[1] D. E. Rumelhart, G. E. Hinton,and R. J. Williams, “Learning representations by back-propagating errors.”  Nature, 1986.</span></p>
<p><span style="font-size: 15px;">[2] The Marginal Value of AdaptiveGradient Methods in Machine Learning- arXiv</span></p>
<p><span style="font-size: 15px;">[3] Knospe, Carl. "PID control." IEEEControl Systems 26.1 (2006): 30-31.</span></p>
<p><span style="font-size: 15px;">[4] A. L. Salih, M. Moghavvemi,H. A. F. Mohamed, and K. S. Gaeid, “Modelling and pid controller design for aquadrotor unmanned air vehicle,” International Conference on Automation.</span></p>
<p><span style="font-size: 15px;">[5] P. Zhao, J. Chen, Y. Song, X.Tao, T. Xu, and T. Mei, “Design of a control system for an autonomous vehiclebased on adaptive-pid,”International Journal of Advanced Robotic Systems.</span></p>
<p><span style="font-size: 15px;">[6] https://en.wikipedia.org/wiki/Control_theory.  </span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;"> </span></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693262&amp;idx=1&amp;sn=a2f0dfc40664b377d1ba9b7735629cfa#wechat_redirect</link>
      <pubDate>Thu, 19 Jul 2018 19:30:24 -0700</pubDate>
      <guid isPermaLink="false">2288</guid>
    </item>
    <item>
      <title>干货 | 人脑与语言——语言模型解码与语义空间映射</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p><span style="font-size: 15px;">人类大脑认知和语言功能之间的关联奥秘，一直是研究者们关注的焦点。在本科的脑认知课程上，小S就听到老师说，人类如何处理语言、形成语言，已经是二十一世纪几乎最受关注也最“一筹莫展”的脑认知研究领域。<br></span></p>
<p><span style="font-size: 15px;"><br></span></p>
<blockquote><p><em><span style="font-size: 14px;">Computational psycholinguistics has “always been...the thing that computational linguistics stood the greatest chance of providing to humanity” (Kay, 2005).</span></em></p></blockquote>
<p><br></p>
<p><span style="font-size: 15px;">随着统计语言模型，神经网络语言模型等的出现，脑认知和自然语言处理这两个学科的交叉研究——<strong>计算神经科学</strong>——也不断涌现出来。曾经的爆款 Science 论文《Human-level concept learning through probabilistic program induction》也属于这个范畴。今天又看到三篇这个领域的研究工作，想分享给大家，包括：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><span style="font-size: 15px;"> <strong>Finding Syntax in Human Encephalography with Beam Search</strong>. ACL 2018.</span></p></li>
<li><p><strong><span style="font-size: 15px;">Navigating with Graph Representations for Fast and Scalable Decoding of Neural Language Models</span></strong><span style="font-size: 15px;">. from Microsoft.</span></p></li>
<li><p><strong><span style="font-size: 15px;">Does the brain represent words? An evaluation of brain decoding studies of language understanding</span></strong><span style="font-size: 15px;">. from MIT.</span></p></li>
</ol></blockquote>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 100%;overflow: hidden !important;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 311.5 25.1" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" width="311.5"><g style="box-sizing: border-box;"><g style="box-sizing: border-box;"><rect x="165.1" y="17.1" transform="matrix(0.7073 -0.7069 0.7069 0.7073 34.8572 124.992)" width="6.6" height="6.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><path d="M56.7,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C57.9,14,56.7,12.8,56.7,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M59.4,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C64.6,14.2,62.3,16.6,59.4,16.6z     M59.4,7.2c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C63.5,9.1,61.6,7.2,59.4,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M262.4,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C263.6,14,262.4,12.8,262.4,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M265,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C270.3,14.2,267.9,16.6,265,16.6z M265,7.2    c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C269.1,9.1,267.3,7.2,265,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><rect x="144.7" y="9.1" width="6.1" height="4.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><polygon points="16.7,18.2 0,18.2 0,17.1 16.3,17.1 22.6,10.8 145.8,10.8 145.8,11.9 23.1,11.9" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="311.5,19.5 288.7,19.5 281.1,11.9 149.8,11.9 149.8,10.8 281.6,10.8 289.1,18.3 311.5,18.3" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="282.5,8.9 271.7,8.9 271.7,7.8 282,7.8 289.5,0.4 311.3,0.4 311.3,1.5 289.9,1.5" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="47.9,7.9 22.7,7.9 15.9,1.1 0,1.1 0,0 16.4,0 23.2,6.8 47.9,6.8" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="201.6,21 167.5,21 162.2,15.7 125.8,15.7 125.8,14.6 162.7,14.6 168,19.8 201.6,19.8" fill="#000000" style="box-sizing: border-box;"></polygon></g></g></svg></section></section></section></section><p><span style="font-size: 15px;">发表在 ACL 2018 上的第一篇工作 [1] 主要关注的是语言模型解码中的语法信息，。比起我们非常熟悉的 LSTM，</span><span style="font-size: 15px;text-decoration: underline;">这篇工作 [1] 使用的基本语言模型是 RNNG，G for Grammers.</span><span style="font-size: 15px;"> 由下面的示意图可以看到，RNNG 在生成过程中，不仅生成具体的词，还生成语法结构。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.5537017726798749" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpRKEsE5e0zIBWG1991tuLCCIDMYlCVZxvFBhAyT3csjSuqkhmmdb7Yg/640?wx_fmt=png" data-type="png" data-w="959" style="width: 387px;height: 215px;"></p>
<p style="text-align: left;"><span style="font-size: 15px;">这样的结构非常适合增量过程建模（incremental processing）。可以想到，当利用这样的生成过程做语法建模时主要包含三种语法分析决策（parser action）：是否开始一个新的成分（左括号），是否结束最近的一个成分（右括号），以及是否生成一个具体的词。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="1.222707423580786" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpFFlbpAPgorXsQ4xawFh7WHyzEoYAq2JC8yPTic1YkiaTxkMJO91A06kg/640?wx_fmt=png" data-type="png" data-w="458" style="width: 373px;height: 455px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">上方给出的示意图的左上角就是三个彩色的方块就代表着三种决策动作。然而，经典的 beam search 算法用于这样结合语法信息的语言模型时，会有些不合适。毕竟语法信息和具体的词，不是一个层次上的信息，把它们放在一起去预测得到的概率分布会严重的不均衡。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.2976076555023923" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp7mZTSwbsErGsnvLDq5zKicVplExHdzFtqPySJADPichdRvuB2YIicq1QA/640?wx_fmt=png" data-type="png" data-w="1045" style="width: 428px;height: 127px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">为此，作者 [1] 使用的是 word-synchronous beam search。其主要思路是在解码时，不断搜索扩展语法结构，直到有足够多的结构都生成了具体的词，然后同步这些词，再开始下一次的结构解码。示意图见下，具体的算法描述请大家参考论文 [1]。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.42799597180261834" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp5vupXiaLPL0an1t8dvX157Y1MGnypaFdib51B6ohNtibdWEmfRUZGcp9g/640?wx_fmt=png" data-type="png" data-w="993" style="width: 389px;height: 166px;"></p>
<p style="text-align: left;"><span style="font-size: 15px;text-decoration: underline;">作者 [1] 希望探究 RNNG+word-synchronous beam search 是不是一个合适的语言认知模型（cognitive model of language processing）</span><span style="font-size: 15px;">。为此，他们比较了人脑听到一段语言时的 EEG 信号和语言认知模型计算过程中的复杂度指标（具体请见论文 Section 4）。从下图的部分实验结果示意来看，回归模型还是比较契合的。</span><br></p>
<p style="text-align:center;"><img class="" data-ratio="0.41259500542888167" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpR06gEzgiaicjLglja0mFOoITG7aYteiajwJWAPHibAYdjtraEXWkHpuvzw/640?wx_fmt=png" data-type="png" data-w="921" style="width: 465px;height: 192px;"></p>
<p><span style="font-size: 15px;">这篇工作 [1] 里还有很多有意思的 insight analysis，比如实验结果就发现，LSTM 的回归效果就比较差，说明不适合作为一个语言认知模型使用。<span style="text-decoration: underline;">而且这篇论文 [1] 给我的最大的冲击还是他们并不是使用 RNNG 直接建模 EEG 数据，而是用 RNNG 中的计算语言复杂度来建模。</span>这可能也是纯自然语言处理和计算语言处理的一个比较大的区别。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 100%;overflow: hidden !important;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 311.5 25.1" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" width="311.5"><g style="box-sizing: border-box;"><g style="box-sizing: border-box;"><rect x="165.1" y="17.1" transform="matrix(0.7073 -0.7069 0.7069 0.7073 34.8572 124.992)" width="6.6" height="6.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><path d="M56.7,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C57.9,14,56.7,12.8,56.7,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M59.4,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C64.6,14.2,62.3,16.6,59.4,16.6z     M59.4,7.2c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C63.5,9.1,61.6,7.2,59.4,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M262.4,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C263.6,14,262.4,12.8,262.4,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M265,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C270.3,14.2,267.9,16.6,265,16.6z M265,7.2    c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C269.1,9.1,267.3,7.2,265,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><rect x="144.7" y="9.1" width="6.1" height="4.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><polygon points="16.7,18.2 0,18.2 0,17.1 16.3,17.1 22.6,10.8 145.8,10.8 145.8,11.9 23.1,11.9" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="311.5,19.5 288.7,19.5 281.1,11.9 149.8,11.9 149.8,10.8 281.6,10.8 289.1,18.3 311.5,18.3" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="282.5,8.9 271.7,8.9 271.7,7.8 282,7.8 289.5,0.4 311.3,0.4 311.3,1.5 289.9,1.5" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="47.9,7.9 22.7,7.9 15.9,1.1 0,1.1 0,0 16.4,0 23.2,6.8 47.9,6.8" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="201.6,21 167.5,21 162.2,15.7 125.8,15.7 125.8,14.6 162.7,14.6 168,19.8 201.6,19.8" fill="#000000" style="box-sizing: border-box;"></polygon></g></g></svg></section></section></section></section><p><span style="font-size: 15px;">第二篇想分享的论文 [2] 则是希望通过借鉴人脑中的语义映射或者说词汇关联，来加速语法模型的解码过程（beam search decoding）。</span><span style="font-size: 15px;text-decoration: underline;">因为过去的认知科学研究多数认为，人脑中的词汇是复杂网络结构</span><span style="font-size: 15px;">，所以当给定一个语境时，人脑能联想到的词汇会锁定在网络中的某一小部分，而不是像 softmax 一样遍历一遍所有词。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;text-decoration: underline;">基于这样的启发，作者 [2] 提出了基于图的快速解码算法 Fast Graph Decoder (FGD)。</span><span style="font-size: 15px;text-decoration: none;">算法的过程也很直接，先根据词向量的语义距离构建一个小世界网络来“模拟”人脑中的语言网络。当解码时，将语境编码（隐变量）和图中的结点进行比较，从而找到最相似的结点作为图解码的起点，并启发式地进行图推理。</span><span style="font-size: 15px;text-decoration: underline;">值得一提的是，在小世界网络上的图搜索有比较成熟的快速算法，这也是这篇论文的算法能大幅提速传统 beam search 的主要原因</span><span style="font-size: 15px;text-decoration: none;">。算法的示意图如下。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpY4bOpSPvTrYhOx7ibzXshyeFCqKdyBwy8iaRTZ9WzPQGbVibcia8iaEA7ibw/640?wx_fmt=png" data-cropx1="0" data-cropx2="941" data-cropy1="0" data-cropy2="300.17562724014334" data-ratio="0.3188097768331562" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJponJT853NIewAayOkHgkz01kxC2y86polD60zcopWkqbUn0RdE15x8A/640?wx_fmt=jpeg" data-type="jpeg" data-w="941" style="width: 558px;height: 178px;"></p>
<p><span style="font-size: 15px;"></span><br></p>
<p><span style="font-size: 15px;">这样简单的思路在实验中效果还不错。在不损失准确度的情况下，其解码计算速度大幅提高。这是因为经典 beam search 主要受限于 softmax 计算，其复杂度是 O( D * |V| )，而这篇中的快速图解码 FGD 则是 O( D * log|V| )。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.34763948497854075" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp5ODkriamDrYNJcRBHS5kZSK0Me8IgKyc6jLicibshnysmxiaN0VtsdPrvA/640?wx_fmt=png" data-type="png" data-w="932"></p>
<p style="text-align:center;"><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 100%;overflow: hidden !important;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 311.5 25.1" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" width="311.5"><g style="box-sizing: border-box;"><g style="box-sizing: border-box;"><rect x="165.1" y="17.1" transform="matrix(0.7073 -0.7069 0.7069 0.7073 34.8572 124.992)" width="6.6" height="6.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><path d="M56.7,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C57.9,14,56.7,12.8,56.7,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M59.4,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C64.6,14.2,62.3,16.6,59.4,16.6z     M59.4,7.2c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C63.5,9.1,61.6,7.2,59.4,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M262.4,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C263.6,14,262.4,12.8,262.4,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M265,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C270.3,14.2,267.9,16.6,265,16.6z M265,7.2    c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C269.1,9.1,267.3,7.2,265,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><rect x="144.7" y="9.1" width="6.1" height="4.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><polygon points="16.7,18.2 0,18.2 0,17.1 16.3,17.1 22.6,10.8 145.8,10.8 145.8,11.9 23.1,11.9" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="311.5,19.5 288.7,19.5 281.1,11.9 149.8,11.9 149.8,10.8 281.6,10.8 289.1,18.3 311.5,18.3" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="282.5,8.9 271.7,8.9 271.7,7.8 282,7.8 289.5,0.4 311.3,0.4 311.3,1.5 289.9,1.5" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="47.9,7.9 22.7,7.9 15.9,1.1 0,1.1 0,0 16.4,0 23.2,6.8 47.9,6.8" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="201.6,21 167.5,21 162.2,15.7 125.8,15.7 125.8,14.6 162.7,14.6 168,19.8 201.6,19.8" fill="#000000" style="box-sizing: border-box;"></polygon></g></g></svg></section></section></section></section><p><span style="font-size: 15px;">今天想分享的最后一篇论文 [3] 则有些与众不同。其更多地是从脑认知的角度来探讨计算语言相关研究的价值。如上面提到过的，计算语言学中经常把脑认知的一些数据（EEG, fMRI等）和神经网络处理的文本数据联系起来，希望从中找到特征的相关性甚至一致性。这样的研究框架可以用下面这样的示意图来展示：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4841549295774648" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpB6UGoFqMtcZHbhSR38T6XzsqsiaFbOxHjlFdoKNlfFMmx9C7AmWRvzQ/640?wx_fmt=png" data-type="png" data-w="568"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">如果将 fMRI 数据转换后的特征和经过语言模型建模后 m_k 的文本特征分别称为 r(x_i, s_j) 和 r(x_i, m_k), 那么这两者的差距越小，说明这个模型拟合某个被试（s_j）拟合得越好。为此，作者比较了下面多种模型：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.31841652323580033" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp1LFwos6YL9L3YriaRUwRGzM3gIjxf4wtbSBYu6qJic1NGSq93Mhbr42Q/640?wx_fmt=png" data-type="png" data-w="581" style="width: 449px;height: 144px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">可以看到，这些模型主要是针对不同自然语言处理相关的任务提出的。从某种程度来说，作者希望看到的是，针对不同任务的语言模型和人类阅读某一句话时的反应是否一致。<span style="text-decoration: underline;">换句话说，人类阅读某一句话是，具体在进行着哪些语言处理（不同的语言任务）。</span></span></p>
<p><br></p>
<p style="text-align:center;"><img class="" data-ratio="0.5940246045694201" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp8xpickMUGa7IpykXhoxt9SaDfUdicmZdGA0pw6Eba1L4p60q6WlVCTMA/640?wx_fmt=png" data-type="png" data-w="569" style="width: 400px;height: 237px;"></p>
<p><span style="font-size: 15px;">实验结果来看，大部分模型学到的特征都可以一定程度上契合人脑活动数据中提取的特征。而且，表现最好的模型是 GloVe、SkipThought 等前四个模型，这些模型也可以认为是最“包罗万象”的模型。因为它们主要建模的是对下一个词的预测，并不涉及一个具体的自然语言任务（task-transferrable）。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">然而，这样的结果只能说明 GloVe 等模型学到了一些和人脑活动相关的特征，却并不能解释人脑在处理语言到底做了哪些任务。</span><span style="font-size: 15px;">如果对这方面的研究感兴趣，推荐大家看一篇综述：John Hale. 2017. </span><strong style="font-size: 15px;">Models of human sentence comprehension in computational psycholinguistics</strong><span style="font-size: 15px;">. In Mark Aronoff, editor, Oxford Research Encyclopedia of Linguistics. Oxford University Press.</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">诚然，这些工作可能很多不够研究，只是借鉴一些思路灵感。就像 CNN 的设计与视觉细胞，神经网络的设计与大脑网络，这些关联可能都不是真实准确的。但是，人工智能并不一定是复原或者说模仿人类的智能。研究工作带来的灵感可能是更宝贵的。在这方面，大家又怎么看呢？</span></p>
<p><br></p>
<p style="white-space: normal;"><span style="font-size: 15px;">更多相关阅读：<br></span></p>
<blockquote style="white-space: normal;">
<p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6&amp;chksm=81233b7fb654b269794e8246aa3c1512421a406625438a864699337a6c3da7f94afa4497fd22&amp;scene=21#wechat_redirect" target="_blank" style="text-decoration: underline;font-size: 15px;">干货 | 限定和指导 Seq2Seq 模型的生成结果（一）</a></p>
<p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693201&amp;idx=1&amp;sn=e385650379ede59df28ae504f453dc9a&amp;chksm=81233b68b654b27ee2f6c4833493e7aec053c6b14bb6afa8c1f2920df9a933e53ce95db4430a&amp;scene=21#wechat_redirect" target="_blank" style="text-decoration: underline;font-size: 15px;">干货 | 限定和指导 Seq2Seq 模型的生成结果（二）</a></p>
</blockquote>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693249&amp;idx=1&amp;sn=038815e711d5ea91fba308a6429e1785#wechat_redirect</link>
      <pubDate>Mon, 18 Jun 2018 23:03:02 -0700</pubDate>
      <guid isPermaLink="false">2287</guid>
    </item>
    <item>
      <title>日常 | 端午安康，来跟你分享大S的生日寄语</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <section style="line-height: 1.5;letter-spacing: 1.5px;padding: 0px 1px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740" style=""></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-size: 14px;">转眼距离上次午夜电台又是4个月，大S回来报道！祝福大家端午安康～</span></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 10px;margin-bottom: 10px;text-align: center;box-sizing: border-box;"><section class="" style="display: inline-block;vertical-align: middle;box-sizing: border-box;"><section class="" style="max-width: 100%;display: inline-block;vertical-align: bottom;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.7948718" data-w="351" src="https://mmbiz.qpic.cn/mmbiz_gif/tXT40RmPWYZP7ibOyLsmyyueOy3yhuiaIJus3iatZfNsxYsScBzZBh6afqFZuGSsB8XDuyoJ0RQiaMIAkXuMQibaR6w/640?wx_fmt=gif" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="gif"></section><section class="tn-yzk-fuid-text-14814-1528344263398" style="display: inline-block;vertical-align: bottom;padding-left: 5px;padding-right: 5px;line-height: 1.2em;margin-bottom: 2px;font-size: 18px;color: rgb(60, 102, 77);box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-family: PingFangTC-light;box-sizing: border-box;">来吃端午粽吧</span></p></section><section class="" style="max-width: 100%;display: inline-block;vertical-align: bottom;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.8303571" data-w="336" src="https://mmbiz.qpic.cn/mmbiz_gif/tXT40RmPWYZP7ibOyLsmyyueOy3yhuiaIJbBSiarpUJ662Qt2kqWF6toKr0tV1gn2HBicmWr9yAGoaYlibyoLbUbXWg/640?wx_fmt=gif" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="gif"></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-size: 14px;">记忆中，每年炎热的夏天，都是考试的季节，而大S的生日，也往往是在准备考试/考试/等考试成绩的过程中度过。毕业工作后，好像终于迎来了第一个不再充斥着考试的生日，但生活中其实又不是何处都有“考验”呢。</span></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.1203125" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_png/t8fqR4fCPn852ALia8OgxEe3yvLia1Nx0JdBCzqLbIGBuQVe1MT8W0IFeHgHBCbsCehWEDvGOWE2xRQPZWjBFnrg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;font-size: 13px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1.0147059" data-w="68" src="https://mmbiz.qpic.cn/mmbiz_png/9LfNnibXYmxqaPnOTmCZ9vibWFA9elPibic8hEhuI62pMDmoaiaiaKQibtEGCXE5mb1mu58waOys6sRL5exQTtFNib7yWg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px solid rgb(71, 193, 168);border-bottom-left-radius: 0px;padding: 0px 0px 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">随着年纪增长，记忆力下降、熬夜修复困难等等初老症状来袭【误】，大S和身边的朋友好像也都越来越意识到身体健康的重要性。从小不擅长运动的大S，本来是非常不喜欢运动的（也可能是因为下意识不愿意面对自己的短板or可能的出丑？）所以基本能不动就不用动…而2018年至今，最大的收获大概就是终于把健身融为了自己生活的一部分，虽然我还不能拍拍胸脯说我已经爱上了运动，但是至少逐渐习惯到了日子就该去健身房找教练报道，也渐渐有了探索未知运动的好奇心，就已经是巨大的进步了。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">为了让自己拥有<strong style="box-sizing: border-box;">健康的身体</strong>，大S今年立下了一个减肥的小目标，几个月过去了，最令人惊喜的大概就是生活方式的改变了吧。仍然热爱美食，却可以吃的更加健康。瘦的7公斤里面5公斤都是脂肪，大概就是最大的礼物啦！</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: right;margin: 20px 0% 10px;transform: translate3d(-20px, 0px, 0px);-webkit-transform: translate3d(-20px, 0px, 0px);-moz-transform: translate3d(-20px, 0px, 0px);-o-transform: translate3d(-20px, 0px, 0px);box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.959375" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8o0XXvfFKWOLqma4ia4cZ55TNwkB2k3bAtiaDZEOPM0Ou5HTAmJUtf5ow/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="png"></section></section></section></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.9677419" data-w="62" src="https://mmbiz.qpic.cn/mmbiz_png/ceU3PHgscVLmeJK16r1B7ApFWuCgTMrkyXdwRcJ1KRTLEdHfy4I6Ty3MSaAwwJtu9NryZ40DxIMJCyY5ibcfhPg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px none rgb(160, 160, 160);border-bottom-left-radius: 0px;padding: 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;border-right: 1px solid rgb(71, 193, 168);border-top-right-radius: 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">生活和工作中难免要经历各种起伏与变化，如何用平常心面对这些变化，大概是大S最近正在努力学习的功课了。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">很多时候，如果自己决定了要花时间好好去做，就不用想太多的去努力。既然是自己决定要去做的，也就不要额外的去期待什么回报，在这个过程中的收获，无论是好的坏的，都将是人生中重要的财富。只要自己付出了百分之百的努力，问心无愧地对得起自己，那么很多外部的噪音也就可以不用在意了。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="font-size: 14px;text-align: justify;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">大S其实是个偏理想主义的人，重感情，不喜欢跟别人争抢，总觉得把事情做成了比什么都重要。也在渐渐锻炼自己<strong style="box-sizing: border-box;">强大的内心</strong>，学习辨别到底什么样的人值得真心的付出，而什么样的时候也要为自己多争取一点。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="line-height: 1.8;font-size: 14px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">有时候觉得自己骨子里其实是一个内向的人，虽然生活好像把我锻造成了现在这个活泼的样子，但是闲暇的时候，还是希望自己一个人窝在咖啡馆的小角落码码字看看书。【当然了天天听我在家里废话的室友估计是不同意这段的哈哈哈</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;box-sizing: border-box;"><section class="" style="display: inline-block;vertical-align: top;width: 45%;padding: 0px 10px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: 10px 0% 20px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);margin: 5px 0% -5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);margin: 5px 0% -5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 0px 0%;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;box-shadow: rgb(0, 0, 0) 0px 0px 0px;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8F6qq2ZHvju863gDGYR9Q3lBAS7JsRHTfAvAd4yAwia347BnCdU7xicGQ/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section></section></section></section></section></section></section></section></section></section></section></section><section class="" style="display: inline-block;vertical-align: top;width: 45%;padding: 0px 10px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="transform: rotateZ(8deg);-webkit-transform: rotateZ(8deg);-moz-transform: rotateZ(8deg);-o-transform: rotateZ(8deg);box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: 20px 0% 10px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 0px 0%;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;box-shadow: rgb(0, 0, 0) 0px 0px 0px;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8sL7hPK7TSoMZCxuM2Y902JPzCPdiaUoI2tHQF385AS8Yqm2xX7cgkQQ/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section></section></section></section></section></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin: 10px 0%;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1.0147059" data-w="68" src="https://mmbiz.qpic.cn/mmbiz_png/9LfNnibXYmxqaPnOTmCZ9vibWFA9elPibic8hEhuI62pMDmoaiaiaKQibtEGCXE5mb1mu58waOys6sRL5exQTtFNib7yWg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px solid rgb(71, 193, 168);border-bottom-left-radius: 0px;padding: 0px 0px 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">      “平稳的人生，一定闷。我受不了闷，我决定活得有趣。”<br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="text-align: right;font-size: 14px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">——蔡澜</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">一直羡慕那些拥有<strong style="box-sizing: border-box;">有趣的灵魂</strong>的人，那种从骨子里散发出来的自在和“幽默”，让他们在我眼中拥有人群也遮不住的耀眼光芒。希望自己也能做个有趣的人，拥有自己的爱好与特点。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="font-size: 14px;line-height: 1.8;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">最近捡起了做手帐的习惯。重新翻出大学的时候做的黑卡，真的包含着非常多美好的回忆。这次去台湾旅行，特地找心灵手巧的同事定制了一本Junk Journal，准备记录一下这趟台湾行。有个10秒的视频给大家秀一下最终的作品：</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><iframe class="video_iframe" data-vidtype="2" allowfullscreen="" frameborder="0" data-ratio="1.7647058823529411" data-w="480" src="https://v.qq.com/iframe/preview.html?vid=t1342v9rwkl&amp;width=500&amp;height=375&amp;auto=0"></iframe><br></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="font-size: 14px;line-height: 1.8;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">累计加起来大概做了要有快20个小时吧…本来计划是旅行过程中每天晚上做，结果每天我除了记得盖章和买和纸胶带外，其他时间都累瘫，完全顾不上…【现在一切都电子化了真的不利于我们这些喜欢收集票根的人呀&gt;&lt;<br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">回来制作的过程中真的太累了…但是事后看着那些照片都会忍不住微笑，真的是美好的回忆呀！</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.9677419" data-w="62" src="https://mmbiz.qpic.cn/mmbiz_png/ceU3PHgscVLmeJK16r1B7ApFWuCgTMrkyXdwRcJ1KRTLEdHfy4I6Ty3MSaAwwJtu9NryZ40DxIMJCyY5ibcfhPg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px none rgb(160, 160, 160);border-bottom-left-radius: 0px;padding: 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;border-right: 1px solid rgb(71, 193, 168);border-top-right-radius: 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">梦想这个词，说大不大，说小也不小。相信大家都对这个词有自己的理解，只希望能够拥有自己<strong style="box-sizing: border-box;">美好的梦想</strong>，在“忙”中不“盲”，不要忘记你曾经是怎么样的小孩，不要忘记你曾希望变成怎样的大人。而大S跟大家的30岁之约，也又近了一年，不知道有多少人还记得这个约定呢，希望我们可以一起努力朝这个目标迈进。</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: right;margin: 20px 0% 10px;transform: translate3d(-20px, 0px, 0px);-webkit-transform: translate3d(-20px, 0px, 0px);-moz-transform: translate3d(-20px, 0px, 0px);-o-transform: translate3d(-20px, 0px, 0px);box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.75" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8KBnhbUDnNbQwTDOiaZMhfRP7wa7b82ELkTUCnR99CXAFVSQibqfDU4YA/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section></section></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px none rgb(160, 160, 160);border-bottom-left-radius: 0px;padding: 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;border-right: 1px solid rgb(71, 193, 168);border-top-right-radius: 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: transparent rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">顺路给大家推荐一本书“牧羊少年的奇幻之旅”（The Alchemist），这本书是巴西作家Paulo Coelho的作品，2015年还曾经被改编成电影。这本寓言式小说运用富含哲理和诗意的语言讲述了牧羊少年圣地亚哥追寻宝藏的奇幻冒险故事，其中的很多小故事都可以映射到我们的活动中，非常多的小细节都值得品味。最后一碗浓鸡汤，嗯</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: right;margin: 20px 0% 10px;transform: translate3d(-20px, 0px, 0px);-webkit-transform: translate3d(-20px, 0px, 0px);-moz-transform: translate3d(-20px, 0px, 0px);-o-transform: translate3d(-20px, 0px, 0px);box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.5625" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8byRNovaeAKNrGjOe4XEwa3353PVPHVWdP7OkR92peQfjdzPC4dZAmw/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section></section></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1.0147059" data-w="68" src="https://mmbiz.qpic.cn/mmbiz_png/9LfNnibXYmxqaPnOTmCZ9vibWFA9elPibic8hEhuI62pMDmoaiaiaKQibtEGCXE5mb1mu58waOys6sRL5exQTtFNib7yWg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px solid rgb(71, 193, 168);border-bottom-left-radius: 0px;padding: 0px 0px 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">一直都觉得很幸福，因为自己拥有<strong style="box-sizing: border-box;">可爱的家人朋友</strong>。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">每次面临选择的时候，爸妈总是跟我说，你开心最重要，想选什么就什么，我们在你背后支持你。</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">每次心情不好的时候，朋友们总是陪在我身边，虽然有时候他们觉得那些小事儿没啥大不了，但是还是耐着性子听我念叨完…有时候在电话里难过得大哭一场，也会让他们跨越整个北京来陪我，跟我说“有我们在，再难也要熬过去”。</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">在一段旅行的最后，一个天天嬉皮笑脸，打打闹闹的朋友，突然一脸真诚的对你说“你真的已经很棒了！不要有压力，要开心的做自己。”那一刻的感动好像无法用语言形容。于是在他说他pick菊姐是因为有些地方让他想到了我的时候我也只能保持微笑了。</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 60%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.159375" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_png/PlhySLk3mOGTKbibTBRr39rOiakvBRn90Ka8ssKkMaaeicq1KJoMcxm1ED48V9kva4YHRr1k3vZdvr52002L5OuvQ/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-size: 14px;">谢谢大家陪伴着大小S一起成长。愿亲爱的你们，拥有健康的身体，强大的内心，有趣的灵魂，美好的梦想，和可爱的家人朋友。下次见啦～</span></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080" style=""></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br></p>
<p><img class="" data-copyright="0" data-ratio="0.4514877102199224" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNpKic4DiblDA75KduXD0EeqoZ8McBkcR5DUfib3SQ5IiaNYpzLW98Valvic1XIQmFEIbibgFwZUPNb69vcg/640?wx_fmt=png" data-type="png" data-w="773" style=""></p></section></section></section></section>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693232&amp;idx=1&amp;sn=9c4e7a96ac7fd09c902ec97c66f6d286#wechat_redirect</link>
      <pubDate>Mon, 18 Jun 2018 05:37:34 -0700</pubDate>
      <guid isPermaLink="false">2286</guid>
    </item>
    <item>
      <title>干货 | 文本改写和风格迁移的一些思路</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p style="text-align: left;"><span style="font-size: 15px;">今天想分享几篇比较新的文本改写（Rewrite）和文本风格迁移（Style Transfer）的工作。会涵盖的论文有：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p style="text-align: left;"><span style="font-size: 14px;"><strong>Unpaired Sentiment-to-Sentiment Translation: A Cycled Reinforcement Learning Approach</strong>. ACL 2018.</span></p></li>
<li><p style="text-align: left;"><span style="font-size: 14px;"><strong>Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer</strong>. NAACL 2018.</span></p></li>
<li><p style="text-align: left;"><span style="font-size: 14px;"><strong>Generating Sentences by Editing Prototypes</strong>. TACL 2018.</span></p></li>
</ol></blockquote>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0578125" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Zv4WU8vayrn26QiaPYYkRiaZlrsxFXeEYDgP2QptHy0O2Rp4VSicg8UT4w/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p style="white-space: normal;"><span style="font-size: 15px;">第一篇工作 [1] 从题目来看就很吸引人，unpaired 对应 cycle，很自然联想到 reconstruction。而 reinforcement learning 主要是为了解决训练过程中离散文本不可求导的问题。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"><br></span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">虽然过去也有一些 text style transfer 的问题，但是因为缺少对于非 stylish 或者说“普通文本”的语义信息的显式保留，导致很容易出现下面这种情况：当把“The food is delicious”这句话从正向情感迁移到负向情感时，会得到“What a bad movie”——虽然情感被正确转化了，但是内容的主体也跟着变了。这显然是不好的。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"><br></span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">为了解决这个问题，这篇工作 [1] 使用了两个 module，一个 neutralization module 用于提取 non-emotional content，另一个 emotionalization module 用于生成或者说融入 sentiment words。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="1.1518691588785046" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Z1IWIVHPTsY1KN91b3pROG2Fnv1VngicJwH9F3wXESmQCl7LolhFb9Bw/640?wx_fmt=png" data-type="png" data-w="428" style="width: 384px;height: 442px;"></p>
<p><span style="font-size: 15px;">如上图所示，下面的 neutralization module 其实是通过<span style="text-decoration: underline;">直接删除</span> emotional words 来实现 non-emotional content 的保留。这一步中，作者 [1] 通过使用 self-attention 得到的注意力权重作为分类器特征，从而自动找出 emotional words。从实验结果来看，这个方法的分类准确率也可以达到 90% 左右（不过 sentiment classification 本身也不是很难，稍微复杂一点的模型都可以达到95%甚至更高）。接下来就是 emotionalization module，可以看到，一个 encoder 和两个 decoder。两个 decoder 对应 sentiment classification 的两个类别（positive or negative），并且分别用于训练时的 reconstruction 和测试时的 generation（emotionalization）。</span></p>
<p style="white-space: normal;"><br></p>
<p style="white-space: normal;"><span style="font-size: 15px;">由于文本的离散性，作者采用了 RL 的方法来训练这两个 module。RL 成功与否严重依赖于奖励函数的设计。在这篇论文中 [1]，作者提出了两个 reward：sentiment confidence 和 BLEU。尤其是 BLEU，主要是针对作者想解决的问题——non-emotional content perserving。最终的奖励函数如下：</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.2662192393736018" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZfgWhrGoTmpGWliaDhuiaVdgnLvoAWyWeM3wvzYqvelTfUIUKTEt4ribiaw/640?wx_fmt=png" data-type="png" data-w="447" style="width: 322px;height: 86px;"></p>
<p style="text-align: left;"><span style="font-size: 15px;">从实验结果来看，过去的工作主要都只考虑了 style transfer 的成功与否（sentiment accuracy），而没有显性地考虑 content perserving，所以确实在 BLEU 指标下表现很差。</span><br></p>
<p style="text-align:center;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZOz8RUOyhHF4A5EfT548lHZExzTuHHaxcJuT08MrGCYvz68c0WqLVSQ/0?wx_fmt=png" data-cropx1="0" data-cropx2="459" data-cropy1="0" data-cropy2="191" data-ratio="0.41612200435729846" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZcoPDZ0RMPXskSdAsI0icfpBkWaS1MsgBFJ07D8WdccPrHb5ticaHeVNw/640?wx_fmt=jpeg" data-type="jpeg" data-w="459" style="width: 373px;height: 155px;"></p>
<p style="text-align: left;"><span style="font-size: 15px;">但是也可以看到，这篇工作提出的方法 [1] 在大幅提高 BLEU 的情况下，ACC 也有一定损失。可以从这个结果看到，</span><span style="font-size: 15px;text-decoration: underline;">想要非常自然地融入 sentiment words 并且不破坏语义和语法，是很困难的事情</span><span style="font-size: 15px;">。看一些直观的转换结果：</span></p>
<p style="text-align:center;"><img class="" data-ratio="1.0369515011547343" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Z59HZ4OwfImiao3NMaHQCcoXBsjGzCzzDAY3GphmPQ98yJB7Wq0QupGQ/640?wx_fmt=png" data-type="png" data-w="433" style="width: 377px;height: 391px;"></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0578125" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Zv4WU8vayrn26QiaPYYkRiaZlrsxFXeEYDgP2QptHy0O2Rp4VSicg8UT4w/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p style="white-space: normal;"><span style="font-size: 15px;">要分享的第二篇论文 [2] 几乎也是一样的思路。但是在具体实现上要更直接一点。第二篇论文 [2] 在进行 style transfer 的时候，基于的是这样的假设：文本的一些属性（attribute），比如文本传达出来的情感信息等，可以体现在文本中的某些特定词汇上。通过改变这些词汇，就可能直接改变整句文本的性质属性（value）。</span><br></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<p style="white-space: normal;text-align: center;"><img class="" data-ratio="1.564516129032258" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKntyYJBUYibdY5JbJZVBlzfmiaOB9qqf3JMa3DygFYhq94qYpMvsjtwcWg/640?wx_fmt=png" data-type="png" data-w="310"></p>
<p style="white-space: normal;"><span style="font-size: 15px;">由上面的示意图可以看出，作者将那些具有指示功能的特定词称为 attribute marker，通过（主要）改变这些 marker，进行 attribute transfer。文章进一步提出了4种具体的方法来改变 marker。</span></p>
<p style="white-space: normal;text-align: center;"><img class="" data-ratio="0.4241935483870968" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnoLp2NaXNdpELibvYCiaIuZVbtNAKOBqHicAia1XnTh2x7Jyps27nt424gQ/640?wx_fmt=png" data-type="png" data-w="620"></p>
<p style="white-space: normal;"><span style="font-size: 15px;">这4种具体的方式也由上图很清晰地展示了出来。值得一提的是 DeleteOnly 和 DeleteAndRetrieve。在删除了具体的 marker 后，作者提出再去根据这个没有情感特定性的句子去从数据中检索出一句最相似但是 attribute 的数值相反（target attribute）的话。检索出的这句话主要用于</span><span style="font-size: 15px;text-decoration: underline;">提取新的 target attribute marker</span><span style="font-size: 15px;">，作为“重写” RNN 的输入之一。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"><br></span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">很明显，这样的工作可能有一定局限性，但在一些简化场景中，却是更可控的。其实验结果也印证了这点 [2]。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<p style="white-space: normal;text-align: center;"><img class="" data-ratio="0.7072" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnVmibed5J4FeUqRBiaWTjTtwuLAaBWRz4vNM2V2YKwGhevibMib38tyuJKw/640?wx_fmt=png" data-type="png" data-w="625"></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0578125" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Zv4WU8vayrn26QiaPYYkRiaZlrsxFXeEYDgP2QptHy0O2Rp4VSicg8UT4w/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">上面两篇工作都是进行风格转换，而在进行文本改写的时候，还有一类常见的场景就是进行非情感类改写，比如扩写句子，改写语法错误等。在进行这样的工作时，其实也可以采用类似的思路。</span><span style="font-size: 15px;text-decoration: underline;">先找到一个“中立”的、“简单”的“模板”，再在这个基础上加入希望增加的信息，如情感、一些复杂修饰词等等</span><span style="font-size: 15px;">。今天要分享的最后一篇工作 [3] 就是这样一个思路。</span></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.9395248380129589" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZkB3GxduPiatiau1xlTzBTGosNZl77HmjWFNsSxD2IvZ03xHSmQWymkoA/640?wx_fmt=png" data-type="png" data-w="463" style="width: 409px;height: 384px;"></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">从上面的示意图中可以看到，修改后的句子增加了“modiocre”和“ridiculous”这两个形容词，变得更复杂了。而重中之重就是在学习 edit vector 上。可以看到，此时的 edit vector 不再是 sentiment words，而是一种 semantic operation。如果让 edit vector 作为一种隐变量，也遵循某种分布，那么同样的 edit vector 应该符合同一种 edit operation，并且对于句子的改写是一种微小的可控的操作。</span></p>
<p style="white-space: normal;text-align: left;"><br></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">基于这样的假设和期望，作者用 VAE 来建模 z，目标函数如下：</span></p>
<p><img class="" data-ratio="0.2673913043478261" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZN9lgIJIafVnTeUhNo3fiaFco9pQGM4DqpQhLCHBVscEg4wGj4HUTb9g/640?wx_fmt=png" data-type="png" data-w="460" style="width: 359px;height: 96px;"></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">关键就是如何保证学到好的 p 和 q。这篇论文使用的方式和过去很多 VAE 的文本应用都不太一样，有兴趣的同学请一定去查阅原文。</span><span style="font-size: 15px;text-decoration: underline;">值得一提的是，这篇工作中的 edit vector z 是直接拼接在给 decoder 的每一个时刻的输入上的</span><span style="font-size: 15px;">，并没有额外的 gate 或者 transform。这也是和上面两篇工作的区别之一。</span></p>
<p style="white-space: normal;text-align: left;"><br></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">最后从结果来看，作者提出的方法确实能更自然地改写句子。这点从 case studies 还有句子的平滑度方面可以看出：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.9108695652173913" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZpmNCqkHYibTlFB3WSiaVibibV1E9q2a92zgtymiaxkV7NDibsdNMUhEzxV5Q/640?wx_fmt=png" data-type="png" data-w="460"></p>
<p style="text-align:center;"><img class="" data-ratio="0.23229166666666667" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZYoqelGU8kPpoW8umpnduAH7oCia2zzocrFjOnADnkowzy8O1gLDiaj0Q/640?wx_fmt=png" data-type="png" data-w="960"></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0578125" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Zv4WU8vayrn26QiaPYYkRiaZlrsxFXeEYDgP2QptHy0O2Rp4VSicg8UT4w/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p style="white-space: normal;text-align: left;"><br></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">更多相关阅读：<br></span></p>
<blockquote>
<p style="white-space: normal;text-align: left;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6&amp;chksm=81233b7fb654b269794e8246aa3c1512421a406625438a864699337a6c3da7f94afa4497fd22&amp;scene=21#wechat_redirect" target="_blank" style="font-size: 15px;text-decoration: underline;"><span style="font-size: 15px;">干货 | 限定和指导 Seq2Seq 模型的生成结果（一）</span></a></p>
<p style="white-space: normal;text-align: left;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693201&amp;idx=1&amp;sn=e385650379ede59df28ae504f453dc9a&amp;chksm=81233b68b654b27ee2f6c4833493e7aec053c6b14bb6afa8c1f2920df9a933e53ce95db4430a&amp;scene=21#wechat_redirect" target="_blank" style="font-size: 15px;text-decoration: underline;"><span style="font-size: 15px;">干货 | 限定和指导 Seq2Seq 模型的生成结果（二）</span></a></p>
<p style="white-space: normal;text-align: left;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692981&amp;idx=1&amp;sn=37ed31cc6ec1ded76a53a26503ba4f93&amp;chksm=8123384cb654b15a6582beb3efb3bae1f350a12077456aded80ddb3d1993be4354bde3fc16a3&amp;scene=21#wechat_redirect" target="_blank" style="font-size: 15px;text-decoration: underline;"><span style="font-size: 15px;">干货 | 你的 KL 散度 vanish 了吗？</span></a></p>
<p style="white-space: normal;text-align: left;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692433&amp;idx=1&amp;sn=b880134068f670cfa59a81b59d432045&amp;chksm=81230668b6548f7eb5bf73a696b5e918d0486e96dca8df55740a947df827d2263e951bcb2136&amp;scene=21#wechat_redirect" target="_blank" style="font-size: 15px;text-decoration: underline;"><span style="font-size: 15px;">ACL 2017 | VAE in Dialog Response Generation</span></a></p>
</blockquote>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693224&amp;idx=1&amp;sn=96e1e8b74bd9cbda4ca3896a6a95befd#wechat_redirect</link>
      <pubDate>Wed, 13 Jun 2018 02:00:35 -0700</pubDate>
      <guid isPermaLink="false">2285</guid>
    </item>
    <item>
      <title>干货 | 改进注意力机制：分解 LSTM 隐状态</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p style="text-align: left;"><span style="font-size: 15px;">注意力机制想必大家都已经很熟悉了。虽然注意力机制最初是被用在机器翻译上，但随后其在各种任务上都占有了一席之地，比如 multi-modal 相关的 image caption generation, video description，以及现在大热的机器阅读理解（machine reading comprehension）。</span></p>
<p style="text-align: left;"><br></p>
<p style="text-align: left;"><span style="font-size: 15px;">然而，注意力机制也有天然的弱点。最初提出的注意力机制，主要是利用当前状态解码器的状态 h_t 和过去多个时刻的 h_{t-i} 比较的结果，来计算其对应的权重——最终将这些信息和当前状态的信息一起综合利用，去判断当前时刻的输出。这样的机制归根结底还是将信息整合到一个向量中，不可避免会造成信息损失和弥散。</span></p>
<p style="text-align: left;"><br></p>
<p style="text-align: left;"><span style="font-size: 15px;">针对这个问题，研究者们也提出了不少解决办法。今天就想和大家分享其中几篇代表作。今天会涵盖的论文有：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><span style="font-size: 14px;"><strong>Frustratingly Short Attention Spans in Neural Language Modeling</strong>. ICLR 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Recurrent Attention Network on Memory for Aspect Sentiment Analysis</strong>. EMNLP 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Dialog Generation Using Multi-turn Reasoning Neural Networks</strong>. NAACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Natural Answer Generation with Heterogeneous Memory</strong>. NAACL 2018.</span></p></li>
</ol></blockquote>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">首先要介绍的工作 [1] 可以说是第一篇专注分析 NLM+attention 的工作。作者首先将</span><span style="font-size: 15px;text-decoration: underline;">注意力机制拆分为三部分</span><span style="font-size: 15px;">：（1）keys: 对比当前时刻的状态和过去时刻的状态；（2）values: 比较结果用于建模上下文；（3）prediction: 综合当前时刻和上下文的信息来进行预测。然而作者认为，现有的方式并不能很好地实现这三件事。所以作者 [1] 提出拆分解码器的状态。</span></p>
<p><img class="" data-ratio="0.4096989966555184" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdMq7tdibd7RcWUZmrQ2KiciaImYE1yzCZZib7gYFQc9L4Laliby01Q6iaVShg/640?wx_fmt=png" data-type="png" data-w="598"><br></p>
<p><span style="font-size: 15px;">对比上图中的 (a) 和 (b)，可以看到，作者将解码器状态 h_i 拆分为 k_i 和 v_i，并用 k_i 比较每一个时刻的解码器状态，用得到的 v_i 建模上下文和预测下一个词。这就是作者提出的第一个改进模型，key-value attention model。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">显然，这个模型中，v_i 仍然</span><span style="font-size: 15px;text-decoration: underline;">一人分饰两角</span><span style="font-size: 15px;">。所以作者 [1] 进一步提出了第二个改进模型，key-value-predict attention model：将解码器状态拆分为三个向量，其中新引入的 p_i 用于做预测。这个模型示意图请参考下方左侧 (c)。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4833887043189369" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdxR1RlaAbCibZmFpRWd76bB3JrtMEWMABolYElMk4VUp9OhoIrPso5gg/640?wx_fmt=png" data-type="png" data-w="602"></p>
<p><span style="font-size: 15px;">在实验中，作者提出的两个改进模型</span><span style="font-size: 15px;text-decoration: underline;">在大部分数字指标下</span><span style="font-size: 15px;">的表现，都超过了传统注意力机制模型，并且符合预期地：key-value-predict attention &gt; key-value attention &gt; traditional attention。然而，从</span><span style="font-size: 15px;text-decoration: underline;">可视化的结果</span><span style="font-size: 15px;">来看，这些模型的差异却并没有那么大：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.47019867549668876" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdbsiaOAZia03zeuZHBAIJFoNVAAcODetgOpeKQZfaSK4n8Rib2Cfuhn0Cw/640?wx_fmt=png" data-type="png" data-w="604"></p>
<p><span style="font-size: 15px;">作者发现，这些基于注意力机制的模型的注意力基本都局限在最近的5个状态的窗口里，也就是题目中所谓的 short span。这也直接导致了，当作者用基于 N-gram 的 RNN 代替基于注意力机制的 RNN 时，也取得了非常惊人的表现。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">作者也通过改变注意力窗口大小，再次印证了，注意力机制的局限性。当窗口增大到一定程度，也就是注意力范围扩大到一定程度后，实验结果并不会继续提高了。换句话说，</span><span style="font-size: 15px;text-decoration: underline;">注意力机制对于捕捉 long-term 范围内的信息的帮助很小</span><span style="font-size: 15px;">。实验结果可以见：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.475" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdDqANuM1yL4kvqndKUK01b90vMbTMhfBCv5C9Am2XSlodjyLoCPqepw/640?wx_fmt=png" data-type="png" data-w="360"></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;"></span><span style="font-size: 15px;">注意力机制的不足在某些任务上会显得尤为严重，尤其是那些对于句子的细粒度语义需要仔细把握的任务。而 Aspect Sentiment Analysis，也叫 Fine-grained Sentiment Analysis， 就是这样一个代表性任务。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">Aspect Sentiment Analysis 的任务是真根据某一个 product/item 的某一个 aspect 来进行情感分类（negative，neural，positive）。常见的例子是电子产品，比如手机。关于手机的一个商品评价可能是从手机的多个方面来撰写的：照相效果很不错，但是电池使用时间较短。显然，这样的任务对于</span><span style="font-size: 15px;text-decoration: underline;">建模上下文语义</span><span style="font-size: 15px;">（尤其是一些复杂句型的语义）的要求非常高。而仅仅依靠简单的注意力机制，或者说单一的注意力机制，就很难做到这点。举一个例子来说，"Except Patrick, all other actors don't play well" 这样的句子，就需要综合“except”和“don't play well”这两个信息，从而得出针对“Patrick”的情感是正向的。<br></span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">为了具有更复杂的综合信息的能力，今天要分享的第二篇论文 [2] 首先将解码器状态拆分和整合成特殊的</span><span style="font-size: 15px;text-decoration: underline;">记忆单元</span><span style="font-size: 15px;">（memory），然后<span style="font-size: 15px;">提出了一种综合多种注意力机制的</span><span style="text-decoration: underline;font-size: 15px;">非线性</span><span style="font-size: 15px;">方法。</span></span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.5818713450292398" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdZ135PDvstvXXHaE3xJh4UCibZib1OxVH3wDppjdjssniaqm2JYtyur5Hg/640?wx_fmt=png" data-type="png" data-w="684"></p>
<p><span style="font-size: 15px;">由上面的模型图可以看到，LSTM 的状态被分解成了 memory cell，并且进一步整合成了 location weighted memory（左上角）。这是因为，对于 Aspect Sentiment Analysis 来说，一句话中可能有多个 aspect。而代表 aspect 的 target word 在一句话中的位置信息有助于建模 target-based context。换句话说，memory 对于每一个 target 是一样的；</span><span style="font-size: 15px;text-decoration: underline;">而 location weighted memory 对于每一个 target 就不一样了</span><span style="font-size: 15px;">。在这篇论文 [2] 中， location weighting 的计算是启发式的（具体见论文公式7）：越靠近 target 的 sentiment words 的权重或者说影响力越大。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">有了这些基础后，就到了这篇论文 [2] 的重头戏，recurrent (multiple) attention on memory。从名字来看就知道，</span><span style="font-size: 15px;text-decoration: underline;">首先 multiple attentions 用于抽取一句话中可能存在的多个片段的有用信息；而 recurrent 用于整合这些信息</span><span style="font-size: 15px;">。更具体地，作者用 i_t 来代表每一个时刻从 memory 中提取到的（attended）有用信息，而每一次整合的结果存储在 e_t 中，不仅影响下一个时刻的 e_{t+1} 也影响着下一个时刻的 i_t。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">这样的注意力机制不仅弥补了单次单个的注意力机制的不足，还增加了非线性带来的高层语义的捕捉（只是希望）。在实验结果上也有一定的印证：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.5484351713859911" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdQSGefGBYs1RAYCOaGiaKQonNFiauFxcA41CeXGIBMJpfZl1JKHlZoHgw/640?wx_fmt=png" data-type="png" data-w="671"></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="text-align: center;vertical-align: middle;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"></section></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"></section><br></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">上面这两篇的任务中，主要还都是只有单源头的信息。接下来要分享的一篇工作 [3] 改进的是对话系统中的注意力机制。因为对话数据一般是两个说话者的，所以信息是双源头的，这也为注意力机制带来了新的困难。</span></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">和上一篇工作比较相似的地方在于，这篇工作 [3] 也是将各种上下文信息拆分到多个记忆单元（memory）中去。只不过这一次，因为是对话的上下文，或者说语境信息，所以信息的种类比较多。具体地，作者将对话双方的对话历史存在了不同的记忆单元，并且将每个人的对话历史又进一步拆分成了两个粒度的记忆——句子级别的，对话级别的（有时候也会叫 session-level，或者说 conversation-level）。与此同时，因为当前最邻近的一句输入（query）是最重要的，作者提出针对这个特殊 query 建立独立的记忆单元。</span></section><p><img class="" data-ratio="0.8851351351351351" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEduicv7VWvTY9Dzm2eGgcIy8qHADhhSNiaOiaibxKjtdsibomHiaic9f8czmtTg/640?wx_fmt=png" data-type="png" data-w="444" style=""></p>
<section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">从上图可以看到，共有四大种（四大列）记忆单元，Mh, Mc, Mq, Mr。其中 Mh 是句子级别的记忆单元，Mc 是对话级别的，Mq 和 Mr 是针对 query 的（对话双方交替建模）。有了这四大种记忆单元，作者分别（separately）对它们使用了注意力机制（multiple attentions），<span style="text-decoration: underline;">然后将四个 attention vectors 拼接在一起，用于最终的预测</span>。作者将这种模型称为 MULTI。</span></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">基于 MULTI，作者进一步提出了 multi-turn attention reasoning。其主要启发自机器阅读理解中的 multi-hop reasoning。关于这部分的细节请大家查阅原始论文 [3]。</span></section><section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">今天最后想分享的论文 [4] 也是希望整合多步注意力机制的结果，但是用了一种比较新颖的、<span style="text-decoration: underline;">和 multi-hop 不同的</span>，整合方法。multi-hop attention 的方法主要是对于同一个、单一的 context vector（可以通过各种方式得到），多次使用注意力机制（multi-turn）。而这篇工作 [4] 则是，对于多个 memory cell 多次使用注意力机制，从而最终得到针对 memory 的经过注意力整合后的 context vector。也就是说，这次的 LSTM 隐状态的拆分，首先表示在数据语言中隐藏的普通语义和外部知识库的拆分；其次针对外部知识库的隐状态，又使用了注意力机制进行知识整合。这两大类的隐状态最终拼接组合成解码器的状态，进行预测。有一些绕，还是请大家看示意图：</span></section><p style="text-align:center;"><img class="" data-ratio="0.36268068331143233" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEd3JIuMxq1Y3RK1mW5ibR11CrHicKO0rNveg3GJQgc3icDg2G5txfPqBaDg/640?wx_fmt=png" data-type="png" data-w="761"></p></section></section><section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">以上就是今天想分享的内容。从今天的内容可以发现，过分依赖注意力机制是很危险的。注意力机制的使用最好配合着特殊的信息拆分和表达。如果说曾经一时无两的神经网络大法使得大家都将各种信息一股脑丢进去而等待结果，那么现在越来越多的研究反而开始将信息重新拆分出来，各司其职也许效果更好。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">相关文章推荐阅读：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=212457976&amp;idx=1&amp;sn=ab503fb91c6a126fbe767f0dd37fb48a&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">干货 | Applied Attention-based Models in NLP（一）</span></a></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=212530824&amp;idx=1&amp;sn=55ca55ddb7ce7d2cd9b316c598b6da03&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">干货 | Applied Attention-based Models in NLP（二）</span></a></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692200&amp;idx=1&amp;sn=5211ddee37bc3a3cad0da93641a7081f&amp;chksm=81230751b6548e476ec2e2c91d023dfd1c7213e34dd085fa57a3f2aa4bd3253920a57aeb5281&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">ICLR 2017 | Attention 和 Memory Networks</span></a></p></li>
<li><p><span style="font-size: 15px;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=401930842&amp;idx=1&amp;sn=0f23752330fe37166a33dabb8e48e143&amp;scene=21#wechat_redirect" target="_blank">干货 | NIPS 2015 Reasoning, Attention, Memory Workshop</a></span></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=400074084&amp;idx=1&amp;sn=598f0b83677a276ec4ba652a0abfff97&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">干货 | 关于 Attention in Cognitive NeuroScience 你需要知道的7件事</span></a><br></p></li>
</ol></blockquote>
<p><br></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693212&amp;idx=1&amp;sn=70b63181946b17eda80f8f35b9dc0b16#wechat_redirect</link>
      <pubDate>Mon, 11 Jun 2018 23:30:45 -0700</pubDate>
      <guid isPermaLink="false">2284</guid>
    </item>
    <item>
      <title>干货 | 限定和指导 Seq2Seq 模型的生成结果（二）</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p style="white-space: normal;"><span style="font-size: 15px;">今天继续关于 Lexically Contrained Decoding 的一些工作。<a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6&amp;chksm=81233b7fb654b269794e8246aa3c1512421a406625438a864699337a6c3da7f94afa4497fd22&amp;scene=21#wechat_redirect" target="_blank">上半部分的内容请点此查看</a>。会涉及的论文有：</span></p>
<blockquote style="white-space: normal;"><ol class=" list-paddingleft-2" style="">
<li><p><span style="font-size: 14px;"><strong>Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation</strong>. NAACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Sequence to Backward and Forward Sequences: A Content-Introducing Approach to Generative Short-Text Conversation</strong>. COLING 2016.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Towards Implicit Content-Introducing for Generative Short-Text Conversation Systems</strong>. EMNLP 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Neural Response Generation with Dynamic Vocabularies</strong>. AAAI 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong style="font-size: 15px;">Globally Coherent Text Generation with Neural Checklist Models</strong>. EMNLP 2016.</span></p></li>
<li><p><strong><span style="font-size: 14px;">Learning symmetric collaborative dialogue agents with dynamic knowledge graph embeddings</span></strong><span style="font-size: 14px;">. ACL 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Dynamic Entity Representations in Neural Language Models</strong>. EMNLP 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer</strong>. NAACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Retrieve, Rerank and Rewrite: Soft Template Based Neural Summarization</strong>. ACL 2018.</span></p></li>
</ol></blockquote>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><section style="white-space: normal;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" powered-by="xiumi.us" style="box-sizing: border-box;"></section></section><p><span style="font-size: 15px;">昨天大部分论文是预测一个主题关键词，来指导和限定回复。很多时候，这样的关键词也可以从上下文中直接抽取。有一类代表词就是实体词，entity。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">关于 entity 的研究在近三四年的时间里获得了许多关注。entity 不仅是知识库的重要组成部分，也是理解很多特定体裁的文本的关键，比如小说和新闻中的各种人物关系。同时，entity 也在特定领域对话中扮演着重要角色。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">今天要分享的第一篇工作 [5] 则集中在一类更特别的文本生成——菜谱生成。菜谱生成的研究和 video description 还有 script knowledge 都有关系。而其最大特点也是，在一份菜谱中覆盖着多个 entity，还有对于这些 entity 的操作。也因此，这篇工作 [5] 设计了<span style="font-size: 15px;text-decoration: underline;">非常显式的记忆机制</span>，专门去建模 entity。</span></p>
<p style="text-align:center;"><img class="" data-ratio="1.0018552875695732" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnShU3SV5cHicdLH5wEt4n2CVCDBz4yPwtMz6S6WnpzNricWtcWZbt18Jw/640?wx_fmt=png" data-type="png" data-w="539" style="width: 400px;height: 401px;"></p>
<p><span style="font-size: 15px;">其主要思想 [5] 可以由上述示意图来表示。给定一个菜谱标题（菜名），和一系列 entity 也就是菜谱的“原材料”，这篇工作 [5] 使用序列模型逐步生成菜谱中的每一个 token。在每一个生成时刻，预测是否该生成一个 entity。如果是，则从原材料列表（agenda）中计算出概率最高的那个，作为此刻的输入。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">到此为止，这个思想还是类似于 CopyNet 或者说 pointer network 的。但是这篇工作 [5] 因为针对菜谱生成，将其进一步细化了。对于菜谱来说，有些 entity 被生成后，还会被再次提及。所以这篇工作 [5] 将每一时刻的门的判断，</span><span style="font-size: 15px;text-decoration: underline;">由两种选项，扩展成三种选项</span><span style="font-size: 15px;">——（1）普通 GRU 生成，也就是非 entity 限定；（2）生成一个新的，还没被提过的 entity；（3）生成一个旧的，已经被提过的 entity。自然而然地，一个新的 entity 一旦被生成，就会变成“旧”的 entity，也就需要记录这种“变化”——由此就有了 checklist 的思想。这篇文章 [5] 也因此将模型命名为 checklist model。</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">上面的 checklist model [5] 的适用范围不止是菜谱生成。如果把 agenda 中存储的 entity 从菜谱原材料变成一些 slot，就可以进行特定领域对话的生成。其根本思想是<span style="text-decoration: underline;">追踪、记录和更新</span>某些特定的某些实体/信息——即用特殊机制建模特定信息。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">从这个角度来说，曾经我们详细介绍过的另一篇工作 [6]，也可以被划分到这里来。</span></p>
<p style="text-align:center;"><span style="font-size: 15px;"><img data-ratio="0.3716651333946642" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNovfg8qxSeea4YOrXcmDVhXzriaYAtliclr0lzFvrx47MbthqebSSYia9hFYgeCYRJ3FjkEZiboOANwaw/640?wx_fmt=png" data-type="png" data-w="1087" style="text-align: center;white-space: normal;"></span></p>
<p><span style="font-size: 15px;">在这篇工作 [6] 中，对话中涉及的 item（朋友列表）是被追踪和更新的对象。这些对象是图中的结点，作者 [6] 希望获得每个结点的表达（vector representation for each node），也就是 Graph embedding。每当新的对话信息被加入，就需要更新这个对话信息中涉及到的相关结点的表达。比如上图中，当出现了“anyone went to Columbia”的时候，首先相关的结点肯定是 columbia。但和 columbia 相关的别的结点，jessica 和 josh 也应该被更新。<a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692766&amp;idx=1&amp;sn=6eace56369297127832978ef1065b437&amp;chksm=812338a7b654b1b1f53e48ab5d586adc7797500e5acad3cafa261891b303ec50dab875c7ef3a&amp;scene=21#wechat_redirect" target="_blank">关于这篇论文的具体解读可以点此查看</a>。</span></p>
<p style="text-align:center;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="text-align: center;vertical-align: middle;box-sizing: border-box;"></p>
<p><span style="font-size: 15px;">上面这篇论文 [6] 的核心思想就是利用动态表达（Dynamic Representation）来建模核心信息（item）。这类思想也在过去几年的关于 general entity 建模中被广泛使用。这方面的一个代表性工作就是由 [7] 提出的 EntityNLM，entity neural language model。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.8387096774193549" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnwH2flMZDN4mK78T579mePVricZskguuCHJ9AdcskQ9ONf53SS6hFIRw/640?wx_fmt=png" data-type="png" data-w="310"></p>
<p><span style="font-size: 15px;">可以看到，作者并不局限在某一个特定场景下的工作，而是使用 language model 建模含有大量 entity 的文本。换句话说，作者的工作使用的是生成模型，而且是<span style="text-decoration: underline;">含有隐变量的条件生成模型</span>。</span></p>
<p><img class="" data-ratio="0.2980251346499102" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn0aTT7MPuSg2LWpvwYG4KGefMRVia3XrG3JBR479mXunvwU6j8apfzFQ/640?wx_fmt=png" data-type="png" data-w="557" style=""></p>
<p><span style="font-size: 15px;">由上图看到，R，E，L 都是要学习的变量。这些变量主要用于辅助预测当下时刻要生成的是否是一个 entity。接下来就是如何去学习这些 entity 的表达。可以想到，entity 的表达受 context 和这些变量影响下的其他 entity 的影响。作者 [7] 也因此使用了 dynamic representation。并且，同上面两篇论文一样，每一词预测出一个 entity 后，这些 dynamic representation 也要更新。关于这两部分的细节公式，请大家查阅原始论文 [7]。</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">限定和指导 Seq2Seq 模型生成结果的另一个重要应用是风格迁移（Domain Adaptation），其潜在的场景就是领域迁移。最早的基于神经网络风格迁移多数是先预训练一个模型，再在新领域的数据上 fine-tuning。这样的过程虽然简单粗暴，但效果往往不尽如人意。尤其是在由离散数据组成的文本迁移上，由于很难通过连续改变一个像素点的数值来训练，导致训练结果更加不可控。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">这篇工作 [8] 则给出了一个新的可行思路。文本的一些属性（attribute），比如文本传达出来的情感信息等，可以体现在文本中的某些特定词汇上。通过改变这些词汇，就可能直接改变整句文本的性质属性（value）。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="1.564516129032258" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKntyYJBUYibdY5JbJZVBlzfmiaOB9qqf3JMa3DygFYhq94qYpMvsjtwcWg/640?wx_fmt=png" data-type="png" data-w="310"></p>
<p><span style="font-size: 15px;">上面的示意图给出了这篇工作 [8] 的思路。作者将那些具有指示功能的特定词称为 attribute marker，通过（主要）改变这些 marker，进行 attribute transfer。文章进一步提出了4种具体的方法来改变 marker。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4241935483870968" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnoLp2NaXNdpELibvYCiaIuZVbtNAKOBqHicAia1XnTh2x7Jyps27nt424gQ/640?wx_fmt=png" data-type="png" data-w="620"></p>
<p><span style="font-size: 15px;">这4种具体的方式也由上图很清晰地展示了出来。值得一提的是 DeleteOnly 和 DeleteAndRetrieve。在删除了具体的 marker 后，作者提出再去根据这个没有情感特定性的句子去从数据中检索出一句最相似但是 attribute 的数值相反（target attribute）的话。检索出的这句话主要用于</span><span style="font-size: 15px;text-decoration: underline;">提取新的 target attribute marker</span><span style="font-size: 15px;">，作为“重写” RNN 的输入之一。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">很明显，这样的工作可能有一定局限性，但在一些简化场景中，却是更可控的。其实验结果也印证了这点 [8]。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.7072" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnVmibed5J4FeUqRBiaWTjTtwuLAaBWRz4vNM2V2YKwGhevibMib38tyuJKw/640?wx_fmt=png" data-type="png" data-w="625"></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">最后来看的这篇论文 [9] 有着异曲同工的思想——也是将 Retrieval 和 Generation 结合在一起——却是用在自动摘要（Summarization）的任务上。自动文摘的研究可以大致分为模板式（template），抽取式（extractive）和生成式（abstractive）。三者各有利弊。这篇论文 [9] 也算是尝试将模板式和生成式结合起来的一种新思路。思想虽然简单，效果依然出乎意料地好。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.2159090909090909" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnNPWMJyK0wPTswsrtzVEQtaNaQ3VNFYFuoiciaibjYQRaRXTgyglSNZRUw/640?wx_fmt=png" data-type="png" data-w="616"></p>
<p><span style="font-size: 15px;">这篇工作 [9] 的思路如上示意图。给定一句话（first sentence in the news），先去检索出一些相似 news，将它们的摘要（headline）作为候选 summary candidates。然后遵照常用方式进行 rerank 找出最合适的那一句 headline <span style="text-decoration: underline;">作为“soft template”</span>。这个 soft template 结合原始的输入，一起传给改写 RNN，生成最终的摘要。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">对比 [8] 和 [9]，这两篇工作都通过 retrieve 的方式从数据库中找了最相似的数据来作为“soft template”。可以猜想，这一步也可以认为是改写出更自然的、更流畅的句子的关键。</span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">这次的两篇总结提到的论文远远覆盖不了 Lexically Constrained Decoding 的全部研究。从更广泛的角度来讲，大家非常熟悉的 CopyNet 或者说 pointer network、placeholder approach 都可以归为这个范畴。相关文章推荐大家阅读：</span></p>
<blockquote>
<p><span style="font-size: 15px;">1. <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6&amp;chksm=81233b7fb654b269794e8246aa3c1512421a406625438a864699337a6c3da7f94afa4497fd22&amp;scene=21#wechat_redirect" target="_blank">干货 | 限定和指导 Seq2Seq 模型的生成结果（一）</a></span></p>
<p><span style="font-size: 15px;">2. <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692766&amp;idx=1&amp;sn=6eace56369297127832978ef1065b437&amp;chksm=812338a7b654b1b1f53e48ab5d586adc7797500e5acad3cafa261891b303ec50dab875c7ef3a&amp;scene=21#wechat_redirect" target="_blank">ACL 2017 | 别再对抗了，通过对话合作找出共同的好朋友吧！</a></span></p>
<p><span style="font-size: 15px;">3. <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692152&amp;idx=1&amp;sn=1ccec5b1d743be5d8a79400e291a1b91&amp;scene=21#wechat_redirect" target="_blank">ACL 2016 | CopyNet 和 Pointer Softmax</a></span></p>
<p><span style="font-size: 15px;">4. <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692579&amp;idx=1&amp;sn=ee3239d8594d75ec7623e0bcd186b959&amp;chksm=812339dab654b0cc40abaaacb0580deccb1160c34354cc93ac43785c67bd515534c40e341fa9&amp;scene=21#wechat_redirect" target="_blank">干货 | Pointer Network 与 CopyNet 之上有哪些新颖任务</a></span></p>
</blockquote>
<p><br></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693201&amp;idx=1&amp;sn=e385650379ede59df28ae504f453dc9a#wechat_redirect</link>
      <pubDate>Sun, 10 Jun 2018 17:21:09 -0700</pubDate>
      <guid isPermaLink="false">2283</guid>
    </item>
    <item>
      <title>干货 | 限定和指导 Seq2Seq 模型的生成结果（一）</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p><span style="font-size: 15px;">Seq2Seq 模型被广泛应用在 NLP 的任务中，包括最被人熟知的机器翻译（Neural Machine Translation，NMT）和越来越火的对话系统（Dialog）。今天我们主要关注 Seq2Seq 模型应用到序列生成的问题。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">序列生成的过程，就是解码器解码的过程。在过去几年的发展中，研究者们逐渐开始关注使用一些先验知识来辅助解码过程，如要求解码出来的序列必须包含某个词汇。这样的研究可以广泛划为“<strong>Lexically Contrained Decoding</strong>”。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">Lexically Contrained Decoding 可以认为是启发自基于统计的 phrase-based MT，但随后其思想被转化应用到许多相似任务中。在一些任务中，含有指定词汇的要求被放宽到了由指定词汇指导（implicitly guided）。本文会涵盖的相关论文，覆盖了 NMT, Dialog Response Generation, Style Transfer 和 Summarization 等。这些论文是：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><span style="font-size: 14px;"><strong>Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation</strong>. NAACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Sequence to Backward and Forward Sequences: A Content-Introducing Approach to Generative Short-Text Conversation</strong>. COLING 2016.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Towards Implicit Content-Introducing for Generative Short-Text Conversation Systems</strong>. EMNLP 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Neural Response Generation with Dynamic Vocabularies</strong>. AAAI 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong style="font-size: 15px;white-space: normal;">Globally Coherent Text Generation with Neural Checklist Models</strong>. EMNLP 2016.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer</strong>. ACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Retrieve, Rerank and Rewrite: Soft Template Based Neural Summarization</strong>. ACL 2018.</span></p></li>
</ol></blockquote>
<p><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">在 NAACL 2018 上，相关学者 [1] 提出了 O(1) 复杂度的 lexically constrained decoding 算法，使得 batch decoding 等技巧也可以同时使用，大大提高了解码效率。[1] 的方法中，首先也保留了像常规 k-size beam search 一样的需要维护的一个 k*|V| 的 scores 矩阵去记录每一时刻的扩展当前 hypothesis 的词汇生成概率。也就是下图的中间部分。如果按照正常的 k-size beam search，下一个时刻保留的 hypothesis 就会是 scores 矩阵中的 top-k 的词了（也就是下图中打★的那些词）。但是，这样的算法无法保证解码到最后，含有要求的词汇，也就是不满足 Lexically Constrained Decoding。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.6648822269807281" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn0RLaJKw0OzU3aXrY64JAAVeU2I38qIzwZLFmZ97ZvxkyKc58jK5tKg/640?wx_fmt=png" data-type="png" data-w="934"></p>
<p><span style="font-size: 15px;">为此，[1] 的工作提出将 beam 按照满足限定词汇的个数的多少来分组。上图中最左侧的不同形状的符号，就表示一个</span><span style="font-size: 15px;text-decoration: underline;">限定词汇（constraint）</span><span style="font-size: 15px;">，而空心则表示还没被满足，实心则表示已经生成过。这些词汇也可以中间的 scores 矩阵最下方找到。基于这样的分组，[1] 提出每一个解码时刻，除了要保留 top-k 以外，还要保留那些还没生成的但有很高概率的词汇去加入到下一个时刻的 hypothesis 中去。这些词也就是上图中的加了→的。又由于，本身有一个 beam bank 是记录已经满足了全部限定词汇的，可以认为这类 hypothesis 是已经 partically-completed 的。对此，[1] 提出保留的第三类 item 就是 single-best token，也就是图中◇的。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">有了以上</span><span style="font-size: 15px;text-decoration: underline;">三类 candidate items（★，→和◇）</span><span style="font-size: 15px;">，就剩下如何分配这些 item 到相应的 beam bank 了。这篇文章 [1] 提出了比较 ad-hoc 的办法去动态分配这些 item。因为整体只维护着一个 k-size beam，最终的算法效率还是 O(Nk)，其中 N 是解码的句子长度，k 是 beam size。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">因为这篇论文[1] 主要研究的是不限定个数的、不限定位置的 target-side lexically constrained decoding，可以看到其适用范围是非常广泛的。一个比较直观的展示结果的方式就是去看看解码出来的句子，是否能找到比较自然的、合理的囊括限定词汇的位置，从而不破坏整体句子的可读性。以下是作者的一些实验案例：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4772234273318872" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnic9j2NZ9phoH8vGFts51WQJ9LwZCGatJt0DWn0ick08pk957sicuZFIjA/640?wx_fmt=png" data-type="png" data-w="922"></p>
<p><span style="font-size: 15px;">这个展示结果也可以从侧面证明，BLEU scores 的提高并不仅仅是因为涵盖了限定词汇，而是正确运用、放置了这些词汇。</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;"></span><br></p>
<p><span style="font-size: 15px;">在 NMT 中使用限定词汇可以帮助一些初级语言学习者。word-by-word 的翻译虽然有时候比较生硬呆板，但是对于语言学习者来说比较容易在翻译结果中学到新的词汇。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">接下来要介绍的几篇论文是要指导和限定对话回复生成的结果。因为对话中的你来我往的一对聊天（utterance pair）的语义关联不像 NMT 翻译任务中的那样显著和紧密，直接使用经典 Seq2Seq 模型常常会学出过于普适的万能回复（generic response）。缓解这个问题的办法有很多，其中一个办法是引导模型生成一些特定主题的词汇。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">在这个方向上，先来看 COLING 2016 的一篇工作 [2]。[2] 的工作提出的 Seq2BF 模型在生成回复时，采用</span><span style="font-size: 15px;text-decoration: underline;">逆向和正向结合的生成方式去引入指定词汇</span><span style="font-size: 15px;">。作者希望指定的词汇是对于某一句最相关的，而不是最常见的。所以在这一步，他们 [2] 用 PMI 相关指标来预测出某一句话的指定词汇。随后，为了让这个指定词汇能出现在回复中的“任意”位置（不仅仅是在开头），作者将生成过程分解成了两个子过程，从而提出了下面的模型：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.683944374209861" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnB9BAnupHmGTFjbfCOgv6AZU70mdgiaxNv5EKBzznYiboflueCjIj9cLg/640?wx_fmt=png" data-type="png" data-w="791"></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><br></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">上面的工作有一个比较明显的问题就是，他们要求生成的回复必须包含这个预测出来的指定词汇。首先这是一个过于强的假设，其次，当预测的词汇不是很相关时，会非常严重的影响生成的回复质量。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">为了解决这个问题，[3] 提出了 implicitly content-introducing 的方法。一提到 implicit，比较容易想到的办法就是利用 gate mechanism。[3] 的工作基于 gate，提出了 hierarchical gated fusion unit (HGFU) ，让预测出来的关键词能</span><span style="font-size: 15px;text-decoration: underline;">指导</span><span style="font-size: 15px;">回复的生成，而不是限定。具体地，[3] 提出 fusion 需要两步，local information initialization 和 global information inception。local 的部分就是用于初始化解码状态，而 global 的部分则是重中之重。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="1.2731277533039647" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKndrdtvNHEB86AWtsJTzj1yRAyfPmGGJeTFnu3iclkZN4Mn0neMicn0k1g/640?wx_fmt=png" data-type="png" data-w="454" style="width: 317px;height: 404px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">上图就是 [3] 提出的用于 global 的 HGFU。其创新点在于引入了一个额外的 cue word GRU，用于独立地捕捉指导词汇的信息，而不至于让 cue word information 弥散在 decoder 的状态中。最后，一个 high-level 的 fusion unit 用于整合两个底层 GRU 的信息，从而更综合地生成回复。</span><span style="font-size: 15px;">如果把 fusion unit 中的 gate 成为 k，k 的开关就决定了信息到底是来自于两个底层 GRU 中的哪一个。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;"><span style="font-size: 15px;">虽然看起来方法比较简单，但效果还是不错的。可以先来看看 k 的可视化：</span></span></p>
<p><span style="font-size: 15px;"><span style="font-size: 15px;"></span></span></p>
<p style="text-align:center;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnhicFFz7QobMza7vtnBORXd5Gia1EVOTTbhib67UWOr0IGGaxnHXNmZ8Tw/0?wx_fmt=png" data-cropx1="5.372781065088757" data-cropx2="454" data-cropy1="0" data-cropy2="460.71597633136093" data-ratio="1.0267260579064588" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnLlUUmUo715WEWYB4TJ7J8ZE4OFFPJ3HWauUlJyuQTALVjibibib3EH7kQ/640?wx_fmt=jpeg" data-type="jpeg" data-w="449" style="width: 334px;height: 343px;"></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">再来看看更多的生成回复。从这些回复中可以看到，有些回复是包含了预测的词汇的，有些回复则没有显性包含。这说明 [3] 提出的方法确实可以做到 implicitly content-introducing。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnTtTE7m4ic3dal2DDUJ5f7jIqYWf8y8peTbbufG25H5UfupKvIY1l6mg/640?wx_fmt=png" data-type="png" data-w="915"></p>
<p><img class="" data-ratio="0.4444444444444444" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnCQuOlsT9dWSt2MdCGzRA9ibfTaa2l4BfWOtS6jiaUB1ofX6tpp3mN07Q/640?wx_fmt=png" data-type="png" data-w="927" style=""></p>
<p><span style="font-size: 15px;"></span><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><br></p>
<p><span style="font-size: 15px;">下面要介绍的这篇工作有一点点不同。在限定词汇方面，之前的工作都是找到一个或多个特定的词汇，强加给解码器。而这篇工作 [4] 则从另一个角度来帮忙解码器生成更好的回复——从解码词表中，动态剔除完全不相关的词汇，让解码器不再受无关词汇的扰乱，同时加快解码过程。一图胜千言：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.5203045685279187" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnkxKb6LJC0oSruRl0zOEfpCpYngVwIO3UUaH8BbFQa8mADQfDQrD2GQ/640?wx_fmt=png" data-type="png" data-w="788"></p>
<p><span style="font-size: 15px;">这样的方法最直观的效果就是解码的搜索空间小了，所以先来看看效率提升：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKndesaS7Vibt0FnoNJmOLucD6rx7u8o9UKUvlD11ROvKMh2KQtLlCxl6g/640?wx_fmt=png" data-cropx1="0" data-cropx2="471" data-cropy1="0" data-cropy2="208" data-ratio="0.4394904458598726" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn4ZiaAiciaYnq6RrPdicKKku5yyZJLCOeJ3d5VT3BoPbFFTKicibZ0EaSia9HA/640?wx_fmt=jpeg" data-type="jpeg" data-w="471" style="width: 471px;height: 208px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">再来看一些生成的回复：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.3244635193133047" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnPFkAxgbvZgWXqqpnoz7UiblI6g53q6SGU3WgxyV0IicP2euYCVhQ9zFA/640?wx_fmt=png" data-type="png" data-w="1165"></p>
<p><span style="font-size: 15px;">上图中的 keyword 就是动态词表构建过程中根据当前输入来找到的语义相关的词。图中给出的是概率最高的三个 keyword。这些 keyword 也可以认为是一种大范围的指定词汇。</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><br><span style="font-size: 15px;"></span></p>
<p><span style="font-size: 15px;">由于篇幅限制，剩下三篇工作留待明天分享。剩下三篇工作中，会提到一类非常重要的限定词汇：entity。希望大家多多支持！</span><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080" style="text-align: center;"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6#wechat_redirect</link>
      <pubDate>Sat, 09 Jun 2018 19:48:05 -0700</pubDate>
      <guid isPermaLink="false">2282</guid>
    </item>
    <item>
      <title>日常 | GHC 18 参会感想</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p><span style="font-size: 15px;">大家好，我是 Yangyang，目前在 Microsoft Vancouver 工作了一年多。今年是我第一次参加 GHC，</span><em style="font-size: 15px;">很受鼓舞和启发</em><span style="font-size: 15px;">，在此与大家分享我的所得：GHC速览 &amp; GHC 18 感想 &amp; 参会Tips。 </span><br></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 270px;height: 90px;vertical-align: top;overflow: hidden;background-position: 50% 50%;background-repeat: no-repeat;background-size: cover;background-attachment: scroll;background-image: url("https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyyIBpNXGAgYNvjUiaFE0GfrSdPTIo4cdN32nXfYcS2Dom5bgDFFQBzuQ/640?wx_fmt=gif");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 45px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="font-size: 18px;color: rgb(98, 82, 63);box-sizing: border-box;"><p style="box-sizing: border-box;"><strong style="box-sizing: border-box;">GHC 速览</strong></p></section></section></section></section></section></section></section><p><strong><span style="font-size: 15px;">GHC</span></strong><span style="font-size: 15px;"> 的全称为 <em>GraceHopper Celebration of Women in Computing</em>，官网为https://ghc.anitab.org/。GHC 由 AnitaB.org 和 ACM 主办，从1994年至今已经举办了18届，2006年开始每年举办一次，会期多在九月和十月。<br>GHC 主要包括 <strong>technical 和 career 两大类 session</strong>，同时也有 <em>career fair, awards ceremony, competitions </em>等多个环节，聚集了来自全球各地的女性科技工作者以及支持 GHC 社区的同盟（GHC也非常欢迎非技术类妹子参加）。<br><strong>GHC 18 于9月26-28日在 Houston, Texas 举办，参会人数为 20,000+。</strong></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.66484375" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy4Jp2NYbCibarPakIeTNz0PE5mlqxHqm4KDFibj5Yc5EKCHIVKdIKiaicFg/640?wx_fmt=jpeg" data-type="jpeg" data-w="1280" style=""></p></section></section><p style="text-align: center;"><span style="font-size: 12px;color: rgb(178, 178, 178);">主会场 - GRB 外观</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 270px;height: 90px;vertical-align: top;overflow: hidden;background-position: 50% 50%;background-repeat: no-repeat;background-size: cover;background-attachment: scroll;background-image: url("https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyTSnJxbdWZAdDan0Ash9GlhBnDLwjWuEs9H6icY1qiaPKiaodGyMPZAibNg/640?wx_fmt=gif");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 45px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="font-size: 18px;color: rgb(98, 82, 63);box-sizing: border-box;"><p style="box-sizing: border-box;"><strong style="box-sizing: border-box;">GHC 18 感想</strong></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;"></span></p>
<p><span style="font-size: 15px;">在 GHC 18 的三天里，我每天给自己安排了6-7个 session，<em>主要是关于 career 的，想获得一些 early-career 的发展建议</em>。<br>九月的最后一天，我回想过去的九月和过去的二十几年，<strong>问了自己的很多问题，重新去认识和面对自己</strong>，然后明白了我从 GHC 18 学到的最重要的事是：<strong>要真正地面对自己，相信自己，想说什么就要说出来，想要什么就要开口，主动向前一步。</strong></span></p>
<p style="text-align: center;"><br></p>
<p><span style="font-size: 15px;">那天我有点激动，写了很长的朋友圈，下午又写了英文版（点击阅读原文可查看）以分享给更多人，在这里分享最主要的三个方面给大家：</span><br></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 20px 0% 10px;box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;padding-right: 8px;padding-bottom: 8px;padding-left: 8px;background-color: rgb(245, 213, 66);box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: -10px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="padding: 3px;display: inline-block;border-bottom: 1px solid rgb(62, 62, 62);line-height: 1;letter-spacing: 0.8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 15px;text-decoration: none;">Confront Yourself Honestly and Trust Yourself from Your Deep Heart</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;"> 在 GHC18 的 Session:Nail Your Promotion 里有一个环节：和你身边的人互相练习要用“I”开头的句子来描述你最近工作里的 achievements，</span><strong style="font-size: 15px;">一定是“I”而不是“We”</strong><span style="font-size: 15px;">。我跟同事 H 坐在一起，一开始我们不知道说什么，后来一句一句越说越多，到最后我们都有点想哭。因为我们是同组的同事，互相了解对方的工作，所以后面都帮对方补充了一些“I”句子。</span><em style="font-size: 15px;">那一瞬间，我意识到，我原来这么不相信自己，原来我做了这么多，原来别人也认可我的工作，只有自己没有给自己认可。</em></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">在这个 session 过后的几天里，我一直在问自己：什么让我觉得自豪？做什么让我真正地从内心感到快乐？我在害怕什么？<strong>我为什么不相信自己？我想成为什么样的人？</strong>我过去是怎么做的？有什么是我想保持的和想改变的？后来我写下了非常多的“我”句子，这里分享一部分吧： </span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: 64.8852%;background-attachment: scroll;letter-spacing: 0px;padding: 15px;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyjk8sdm4lhbanj6jb0lzruVT6GWXdbOfKJMhrzrAAPN11GicOr7hiacRQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">我是个好姑娘，完全是。我热爱生活，努力工作，想要多挣钱，带爸妈去看世界，给家庭带来更好的未来。我唱歌很好听，特别是唱邓丽君的歌。我拍照也不错，我有发现美的眼睛，我把那些美好的片刻分享出来，让快乐和感动翻倍翻倍再翻倍。我是个生活和工作都比较 organized 的人，我把家里收拾得窗明几净，我也很会打包行李。我有点强迫症，有时其实这给我带来了一些麻烦，但我现在意识到了，也乐于去改进。我积累着自己的生活和工作经验，方方面面，而且我非常乐于分享给别人，帮助别人，这也让我感到和快乐。我是个helpful的人，未来我想做个 knowledgeable 的人，然后依然 helpful，给他人支持和帮助！</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;"> </span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">我有些胖，但不是那么胖，我尝试去健身，但我确实毅力上还有些问题。我害怕所有和高速有关的冒险，比如滑雪比如开卡丁车。我也很害怕未知，现在游泳还不敢去深水区，因为我总在脑子里想象一万遍自己呛水然后被救生员捞起来的场景。我发现克服心理障碍真的不容易，不要因为别人说这很容易啊然后觉得自己不行，不要为自己此刻做不到而且苛责自己。我害怕开车，未来我当然要学会在高速开车，因为这是我生活的一部分，我会找个靠谱的教练，多上几节课，问问我的女性朋友们怎么做到的，寻求她们的帮助。我也挺开心我尝试过一些对我来说有点难的冒险，然后跪求自己不要再逼自己尝试了，因为我真的不喜欢，比如卡丁车。有个朋友在看我开过卡丁车后对我说：这对你来说是折磨，我完全同意，也不想再尝试，let it go。我会继续学滑雪，即使我今年一月摔裂了膝盖右侧韧带，过了半年才好，因为我觉滑雪场风景太好了，我也有点享受滑雪的感觉。我会学着安全地摔倒，学着冷静一些，学着 trust my body，然后不再阻止我老公去蓝道探索，而是对他说：Have fun and take care of yourself。我也会继续学自由泳，学踩水，即使可能需要很久才能学会，要过很久才能在深水区游个来回。</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">在这样几番的自我对话之后，我开始相信自己是很棒的人，是值得被爱的人，也是认真去爱的人。如果你是一个不够自信的人，那么就尝试向自己提问吧，一定要让内心深处的自己来回答。<strong>Confidence is the stuff that turns our ideas into actions. What we believe matters.</strong> 诚实地面对自己，相信自己，接受自己，是自我成长的重要一步，有时甚至是第一步。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 20px 0% 10px;box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;padding-right: 8px;padding-bottom: 8px;padding-left: 8px;background-color: rgb(245, 213, 66);box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: -10px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="padding: 3px;display: inline-block;border-bottom: 1px solid rgb(62, 62, 62);line-height: 1;letter-spacing: 0.8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 15px;">Be Active and Lean In</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">这是我在 GHC18 的多个 session 听到的共同点，听起来有点鸡汤，但从前我都是“道理我都懂，生活还是过成了这样”。<em>向前一步，曾经真的很难，尤其对于女生。</em>多少个时刻，我在犹豫，我在胆怯，我欲言又止，然后什么都没有改变，自己也不开心。在 GHC18 的经历告诉我：</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;text-align: left;"><section style="display: inline-block;vertical-align: top;width: 29%;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-left: 0%;text-align: right;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"><section style="display: inline-block;vertical-align: bottom;width: 10px;margin-bottom: 5px;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: top;margin-left: 2px;color: rgb(132, 132, 132);font-size: 14px;text-align: left;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 12px;">1<br style="box-sizing: border-box;"></span></p></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 71%;border-left: 5px solid rgb(245, 213, 66);border-bottom-left-radius: 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 15px 0% 8px;box-sizing: border-box;"><section style="border-width: 0px 0px 1px;border-style: solid;border-bottom-color: rgb(132, 132, 132);color: rgb(132, 132, 132);text-align: justify;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="font-size: 12px;">Ask For What You Want</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">第一次听到这句话是在Session: Nail Your Promotion。回想自己工作满一年的时候有个升职节点，我想知道自己有没有机会，但我总等着 manager 来跟我聊，而我同组的两个哥们一直在催我主动跟 manager 聊。我犹犹豫豫好几次，我怕尴尬，我怕失望。终于在一次1/1上跟 manager 问了，manager 给了我反馈和认可，也解释了为什么这次没升，也聊到了其他人都做怎么样。听过这个 Session，我回想起来，这件事最后并不尴尬，反而从中获得了很多有用的信息。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">所以，当你想退缩时，问问自己：我在害怕什么呢？事情真的这么复杂吗？最坏的结果是什么呢？我能做什么让结果不那么坏？然后你会发现，有时你只需要 Say Hello，只需要 Ask Questions。所以啊，<strong>Don’t wait for the door to open. Open doors and go through them. </strong></span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;width: 29%;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-left: 0%;text-align: right;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"><section style="display: inline-block;vertical-align: bottom;width: 10px;margin-bottom: 5px;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"></section></section><section style="display: inline-block;vertical-align: top;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;margin-left: 2px;color: rgb(132, 132, 132);font-size: 14px;text-align: left;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 12px;">2</span></p></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 71%;border-left: 5px solid rgb(245, 213, 66);border-bottom-left-radius: 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 15px 0% 8px;box-sizing: border-box;"><section style="border-width: 0px 0px 1px;border-style: solid;border-bottom-color: rgb(132, 132, 132);color: rgb(132, 132, 132);text-align: justify;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;"><span style="font-size: 12px;">Measure Your Achivements</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">在 Keynote 里 Padmasree Warrior 提到：Eyes on the star, feet on the ground。当时我就想，这不就是“仰望星空，脚踏实地”！可是还缺点什么呢，<em>怎么才能更接近自己的星星，怎么更好地脚踏实地？</em>后来我从其他的 session 得到了一种答案：Measure you achievements。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.665625" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyoXRAnaXLKmWruibp12adLy5LWl2HegvqoYL2vNR9BsHbyNaPR5I6IGA/640?wx_fmt=jpeg" data-type="jpeg" data-w="1280" style=""></p>
<p style="text-align: center;"><span style="font-size: 12px;color: rgb(178, 178, 178);">Keynote 现场</span></p></section></section><p><span style="font-size: 15px;">这是一条很有实践性的建议。比如工作中，你会不会觉得自己做了很多事情，但是好像太琐碎，别人也看不到？这时量化你的工作成果就很重要。你可以问自己，这个月我做了几次 Demo？我手中的项目有几个阶段，我做了哪几项工作，推进到了哪里？这个项目的结果有什么指标可以衡量，我期望以及实际提升了多少？比如生活中，你会不会觉得每年都过得好快啊，我怎么好像什么也没做？你同样可以尝试量化生活中的 achievements。比如问自己，今年去了哪几个地方，拜访了哪些人，读了哪些书，有哪些是我今年的大事件？<strong>Measure you achievements, 其实是在沉淀和分析你的所得，也帮你去设定新的目标，审视自己，也展示给他人。</strong></span><br></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;width: 29%;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-left: 0%;text-align: right;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"><section style="display: inline-block;vertical-align: bottom;width: 10px;margin-bottom: 5px;border-bottom: 1px solid rgb(132, 132, 132);box-sizing: border-box;"></section><section style="display: inline-block;vertical-align: top;margin-left: 2px;color: rgb(132, 132, 132);font-size: 14px;text-align: left;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="box-sizing: border-box;">3</p></section></section></section></section></section><section style="display: inline-block;vertical-align: top;width: 71%;border-left: 5px solid rgb(245, 213, 66);border-bottom-left-radius: 0px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 15px 0% 8px;box-sizing: border-box;"><section style="border-width: 0px 0px 1px;border-style: solid;border-bottom-color: rgb(132, 132, 132);font-size: 12px;color: rgb(132, 132, 132);text-align: justify;padding-right: 8px;padding-left: 8px;box-sizing: border-box;"><p style="white-space: normal;box-sizing: border-box;">Communicate And Connect: GIVE &amp; GAIN</p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">在 GHC 18 的第三天上午，我取消了之前注册的 session，我花了一上午在 Career Fair 的各个展台之间穿梭，我问不同公司的工程师：How do you feel about yourcompany? Do you like your job? What’s your most interesting project recently? 朋友开玩笑说，你不是去参加面试，你是去面试 ta 们呀！再回想整个九月，我跟很多人交流了很多事情，比如，和不同的朋友聊工作聊家庭聊自己，和小伙伴们在 KTV 聊唱歌的经验，和当时的 Airbnb 房东聊中西文化差异，在 GHC 和同事聊参会感想，和爸妈聊对过去的和现在的自己的想法。这些都让我觉得很开心。</span><br><span style="font-size: 15px;"></span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><strong><span style="font-size: 15px;">我想，交流的意义在于连接，在于给予和获得的正向循环，在于建立属于自己的Sponsor Network。</span></strong><span style="font-size: 15px;">我越来越愿意说出自己的想法，获得他人的反馈，无论是支持还是反对；我也倾听和了解 ta 们，向 ta 们分享我的经验，给出我的反馈。<em>曾经我觉得“Impact”是一个很大很远的词，现在我觉得只要有一个人因为我分享的东西而受到启发，这就很有意义。</em>我也感谢他人给我带来 ta 们的 impact。这些都是你我所处世界的碰撞、拓展和延伸。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 20px 0% 10px;box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;padding-right: 8px;padding-bottom: 8px;padding-left: 8px;background-color: rgb(245, 213, 66);box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: -10px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="padding: 3px;display: inline-block;border-bottom: 1px solid rgb(62, 62, 62);line-height: 1;letter-spacing: 0.8px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 15px;">Embrace Vulnerability</span></p></section></section></section></section></section></section></section><p><span style="font-size: 15px;">很多个 Session 都提到，要 walk across boundaries，要 push yourself to your boundaries，也提到 Growth is inversely proportional to comfort。我在从 GHC 18 回来之后，跟当时的 Airbnb 房东聊起这个话题，她给我推荐了 TED 演讲：<em>The power of vulnerability | Brené Brown</em>。我看完之后的体会是，<strong>如果我想要突破自己的极限，push 自己一把，首先要做的是接受自己的不完美。</strong></span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">就像我在上面那段“我”句子所说的，我有很多不完美的地方。而从前我是那么那么那么在意别人的看法，典型的讨好型人格。我把自己限制在自己想要的“完美”里，然后对自己的弱点深恶痛绝。我总是要让自己做到 well-prepared，然后对那偏离计划的10%而焦躁不安，而责备自己。现在我明白了，因为我不曾相信自己，不曾接受自己，我总想100%的去 control，去 predict，其实完全不可能做到。我要做的是去调研、去交流、去发现、去解决，去化解100%的压力，去把0%的信心变成60%的把握，然后认真地努力地向着100%行动。即使这件事成功之后，再回头看，你会发现过程本来就会 messy，再重头来过，也没法做到每个细节100%。Jesscia O. Matthews 在 Keynote 里说：<em>Just because it’s not your plan doesn’t mean it’s not your destiny。</em>所以我与自己和解，接受过去和现在的自己，接受不完美的自己，不苛求，有期待，有行动。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.66328125" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyZyicsibKg6zntaZ43ffwMZicGQY1QTMSs6FLePqeEutUF1pzlA6GF6BrA/640?wx_fmt=jpeg" data-type="jpeg" data-w="1280" style=""></p>
<p style="text-align: center;"><span style="font-size: 12px;color: rgb(178, 178, 178);">主会场 - GRB 涂鸦</span></p></section></section><p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">还想再说一些琐碎的感想：</span><br></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: 48.2185%;background-attachment: scroll;letter-spacing: 0px;padding: 15px;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUythBJicdAZFf5HbpstE9I0mGho7LKYdpKYUtLQzbkiaSg7DwJFtJjEvRg/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">1. 社区真的很重要。在 GHC 18，我时常感受到“啊，原来我不是一个人，原来我不孤单”，这从内心为我带来了勇气和支持。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">2. 我观察了 Keynote 和 Featured Session 的这些speaker，她们很明白自己想要给这个世界带来什么改变，也相信自己可以。自信自强自立的人真的会发光。(<em style="box-sizing: border-box;">Keynote的视频可以在YouTube上观看。</em>)</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">3. 关于 Career Fair：现场真的很热闹，逛不同公司的展台，观察他们的风格差异，是很有趣的事情。也感受到找工作竞争越来越大，佩服那些在展台和招聘人员非常流利地讲解自己简历的妹子们，我当年找工作时做不到这么自信从容。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">4. 在 Career Fair 的 Indeed 和 LinkedIn 展台可以拍那种职业头像，我在参会的第二天下午也去拍了一张。照片出来后我很惊讶，当时我问自己：这个看起来还挺自信的笑着的妹子，是我？我觉得这是 GHC 给我带来的即时改变。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">5. 我还需要从文化角度去提升英语交流能力。虽然我在温哥华工作了一年，基本的交流没有问题，对 keynote 和 session 中大部分梗的理解可以和大家同步，但有时还会有延迟。以前在国内读书，看了美剧英剧；来温哥华的一年，反而主要看的是国内剧集和综艺。借口包括“我要和国内朋友们维持话题”和“工作已经都是英语交流了，回家我想放松”，其实是自己不愿走出舒适区。这是我需要改变的地方。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">6. 我对很多印度妹子在 session 里的积极和主动感到很惊讶，听到了一些惊艳的想法和观点，体会到了职场中印度 vs 中国的竞争感。</span></p>
<p style="box-sizing: border-box;"><span style="font-size: 14px;">7. 在我参加的 session 里，台上的亚裔（我指的中日韩）面孔略少，但在颁奖环节又有好几个亚裔妹子获奖，让人感到惊喜。期待未来能够在世界舞台感受到更多亚裔女性工作者的影响力。</span></p></section></section></section></section></section></section></section><section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;width: 270px;height: 90px;vertical-align: top;overflow: hidden;background-position: 50% 50%;background-repeat: no-repeat;background-size: cover;background-attachment: scroll;background-image: url("https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUyyIBpNXGAgYNvjUiaFE0GfrSdPTIo4cdN32nXfYcS2Dom5bgDFFQBzuQ/640?wx_fmt=gif");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-top: 45px;margin-right: 0%;margin-left: 0%;box-sizing: border-box;"><section style="font-size: 18px;color: rgb(98, 82, 63);box-sizing: border-box;"><p style="box-sizing: border-box;"><strong style="box-sizing: border-box;">GHC 参会 Tips</strong><span style="font-size: 15px;"> </span></p></section></section></section></section></section></section></section><section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">注册 &amp; 抢票</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">今年没有 guaranteed tickets，需要自己守在电脑前注册抢票。注册开放时间为 9AM PST。主办方今年更新了注册算法，记录每人最后一次打开的浏览器窗口中的最后一次点击为最后排队顺序。所以找一个网络靠谱的地方，刷出注册按钮后，一次点击即可，大概1-5分钟后会出现下一步。整个注册全部完成大概花了10分钟。每一步的响应都比较慢，要耐心。最好给最终显示注册成功的页面截个图，以防万一。最后收到 confirmation 邮件，才算是完全确保注册成功。</span><br></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">GHC 18 注册类型包括 general, academic, student三种，general 价格为1,150USD，后两者价格会便宜很多，听说学生注册是 300 USD，但我没有注册过，猜测会需要相应的身份证明。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">酒店</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">今年 GHC 的会场主要在 George R.Brown Convention Center (GRB), Marriott Marquis (MMQ) 和 Hilton Americas-Houston (HIL)，GRB 的两端有天桥分别连接 MMQ 和 HIL。GHC 官网有推荐的酒店列表，也有 shuttle 连接各个酒店和会场。我跟同事妹子 H 很幸运抢到了 HIL 的房间，省去了每天 shuttle 来回的时间。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.66640625" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUybwCQ13Vk01e3PHl1BzhuQHXNyiaNQtbfByNLuFahexcJhicodC7VYygg/640?wx_fmt=jpeg" data-type="jpeg" data-w="1280" style=""></p>
<p style="text-align: center;"><span style="font-size: 12px;color: rgb(178, 178, 178);">休斯顿 - 有轨电车</span></p></section></section><p><span style="font-size: 15px;">我们实际订酒店的过程有点慌乱。由于我们公司报销的 approver 多次强调确认拿到票了才可以订酒店，我们收到 confirmation 之后再去看酒店列表时，发现 HIL 每人一间会超额度。于是马上跟 approver 联系，确认两人一间没问题，便赶紧下手，最后在注册后的半小时抢到了一个 HIL 的房间。有国内小伙伴告诉我，当时她抢完票国内已经半夜了，想着第二天一早订酒店，结果第二天只能订到 shuttle 最远处的酒店了。</span><br></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">建议想参加 GHC 的小伙伴，先提前查看官网推荐的酒店列表信息，确认公司/学校报销限额，列出可选酒店；注册成功就马上订酒店。如果无需考虑报销限额，或不怕注册失败而酒店可能无法取消的问题，可以提前订酒店。</span><span style="font-size: 15px;"> </span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">Pre-Registration</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">GHC 18 可以通过网站或 APP 提前注册感兴趣的 session，获得 reserved spot。Keynotes 无需注册，featured sessions 的 reserved spot 很快会被约满。入场时有两个队列：reserverd, walk-up。Reserved 队列先入场，walk-up 队列后入场，一般也都有位置。</span><br></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">我是在 GHC 18 iOS APP 注册 reserved spot，Schedule 页面点击 session 右侧的+号即可。重叠时段只能注册一个，之后可以在 My Schedule 看到自己选择的 session。建议慎选20min session，个人感觉 career 类的20min session 比较水，内容基本和相似的1h session 重叠。<em>（另外同事反映，今年她在 Web 端的 reserved spot全部丢失，Andriod APP 经常卡死闪退。希望今年会有人对此给出反馈，GHC 19 能做得好些。）</em></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">航班 &amp; 时差</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">尽量选择直飞航班，不得不转机的话就尽量多留一些时间，飞机还是很容易延误的；如果没有买买买需求，带个小的登机箱会比较方便，转机也不用担心行李跟不上。</span><br></p>
<p><br></p>
<p><span style="font-size: 15px;">国内过来的小伙伴，中美时差很长，而会期很短，所以可能刚倒过来时差，回去又得倒回来。建议借助褪黑素，尽量撑到晚上睡。比如上飞机之前吃1-2粒褪黑素，尽量在飞机上多睡；落地后困了也尽量撑到晚上再睡；睡前1小时可以吃1粒褪黑素。</span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">行装准备</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">建议关注出发地和目的地天气，考虑会场有空调，必备薄外套和雨伞。比如：Houston 当时气温大概是23-30℃，偶尔有雷阵雨。室外很湿热，会场里面很冷，所以我当时是：夏装+外套+雨伞。<em>（我跟好几个妹子都吐槽了，GHC 作为面向女性的大会，空调温度上的关怀还不够啊。）</em></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">吃饭</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">第一天没有考虑人很多的问题，比较狼狈。酒店本身的自助早餐和附近的星巴克都排着很长的对，转了好几个地方，我们才在 GRB 二楼找到队伍稍短些的小餐吧买到吃的，一路奔去丰田中心听 Keynote。午饭时间依然到处排长队，因为我和同事都有 pre-registration，想听的 session 间隔很短，之后放弃了午饭。下午主办方供应了一些甜点、面包、蔬菜沙拉，可以拿来充饥。晚饭时间比较宽裕，可以寻找合适的餐馆，如果去参加 party 也有食物。建议参加 GHC 的妹子，每晚买好第二天的早饭或者订好酒店的早餐客房服务，同时也买点零食面包，免得因为来不及排队而吃不上午饭。</span><span style="font-size: 15px;"> </span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: left;margin: 10px 0%;box-sizing: border-box;"><section style='display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-position: 0% 0%;background-repeat: repeat;background-size: auto;background-attachment: scroll;border-style: solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;background-image: url("https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqEanM1JFoicXia0dYwPNyvUy5CtURiafrIegBuxuDYfTkxxUMmgk0xUbdtN0u4xvzVuCFchBQLKib6pQ/640?wx_fmt=png");box-sizing: border-box;'><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="margin-right: 0%;margin-bottom: -2px;margin-left: 0%;box-sizing: border-box;"><section style="display: inline-block;background-color: rgb(255, 255, 255);padding: 5px;border-style: none none solid;border-width: 1px;border-radius: 6px;border-color: rgb(62, 62, 62);overflow: hidden;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="display: inline-block;min-width: 10%;max-width: 100%;vertical-align: top;background-color: rgb(216, 98, 34);border-width: 0px;border-radius: 6px;border-style: none;border-color: rgb(62, 62, 62);overflow: hidden;padding: 5px 10px;box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="box-sizing: border-box;"><section style="text-align: left;color: rgb(252, 242, 241);padding-right: 10px;padding-left: 10px;box-sizing: border-box;"><p style="box-sizing: border-box;"><span style="font-size: 14px;">After-hour Party</span></p></section></section></section></section></section></section></section></section></section></section></section></section></section><p><span style="font-size: 15px;">GHC 三天会期，前两天晚上有些公司会安排自己的 party, 第三晚是 GHC 官方的 party。第一天我去了 Microsoft party，公司也发了注册的 link 给 FTE，鼓励大家转发&amp;邀请其他参会者，位置离主会场步行20min可达。第二天有 Google 主办的 WomenTech Worker 的 party，需要先排队登记，然后排队乘大巴去 party 场地。第三天GHC 官方 party 就安排在主会场，凭借注册时发的票入场。也听说了 Twitter 等其他公司有 party，有兴趣参加的小伙伴可以尝试去 career fair 的各公司展台直接提问。</span><br></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section style="max-width: 100%;vertical-align: middle;display: inline-block;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 640 18.8" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" width="640"><g style="box-sizing: border-box;"><path style="box-sizing: border-box;" d="M1.8,15c9.6,0,19.2,0.1,28.8,0.2l0.2-3c-9.7-0.1-19.3-0.2-29-0.2C-0.6,12-0.6,15,1.8,15z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M75.7,13.1l-25.4-0.5l-0.2,3l25.6,0.5c3.7,0.1,7.5,0.1,11.2,0.1l0.4-3l-5.8,0L75.7,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M138,13.1l-5.8,0.1l-25.4,0.1h-0.1l-0.3,2.9c10.5,0,21,0,31.6-0.2l3.4-0.1l0-3L138,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M161,15.6c10.9-0.2,21.8-0.4,32.7-0.5c0.1-1,0.2-2,0.2-3c-11,0.1-22,0.3-33,0.5L161,15.6z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M291.5,13.2l-5.8-0.1l-13.2-0.3l0.3,3l12.9,0.3c8.2,0.2,16.3,0.2,24.5,0.2l-0.5-3L291.5,13.2z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M348.1,13.1l-5.8,0.1l-12.8,0.1c0.2,1,0.3,2,0.5,3c6,0,12-0.1,18.1-0.2l20.7-0.4l-0.2-3L348.1,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M422,12c-11.4,0-22.7,0.1-34.1,0.3l0.2,3c11.3-0.2,22.6-0.3,33.9-0.3C424.4,15,424.4,12,422,12z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M213.3,12c0,1-0.1,2-0.2,3c13.4,0,26.7,0.2,40.1,0.4c-0.1-1-0.2-2-0.3-3   C239.7,12.2,226.5,12.1,213.3,12z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M1.8,5.5c9.8,0,19.7,0.1,29.5,0.2l0.2-3C21.6,2.6,11.7,2.5,1.8,2.5C-0.6,2.5-0.6,5.5,1.8,5.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M75.7,3.5L51,3l-0.2,3l24.9,0.5c4,0.1,8.1,0.1,12.1,0.2c0.1-1,0.1-2,0.2-3l-6.5,0L75.7,3.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M138,6.5l3.3-0.1l0-3L138,3.5l-5.8,0.1l-24.7,0.1l-0.2,2.9C117.6,6.7,127.8,6.7,138,6.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M160.8,3.1l0,3c11-0.2,22-0.4,33.1-0.5c0-1-0.1-2-0.2-3C182.8,2.7,171.8,2.8,160.8,3.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M251.9,5.9c-0.2-1-0.4-2-0.6-3c-12.7-0.2-25.5-0.3-38.2-0.3c0.1,1,0.2,2,0.2,3   C226.2,5.5,239.1,5.7,251.9,5.9z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M308.9,6.7c-0.1-1-0.2-2-0.3-3l-17.1-0.1l-5.8-0.1l-14.6-0.3l0.5,3l14.1,0.3   C293.5,6.7,301.2,6.7,308.9,6.7z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M367.9,6.1c-0.2-1-0.3-2-0.5-3l-19.2,0.4l-5.8,0.1l-14.1,0.1c0.1,1,0.2,2,0.3,3   c6.5,0,13-0.1,19.5-0.2L367.9,6.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M422,2.5c-11.7,0-23.4,0.1-35.1,0.3c0.2,1,0.3,2,0.4,3c11.6-0.2,23.1-0.3,34.7-0.3   C424.4,5.5,424.4,2.5,422,2.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M447.8,15c9.6,0,19.2,0.1,28.8,0.2l0.2-3c-9.7-0.1-19.3-0.2-29-0.2C445.5,12,445.5,15,447.8,15z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M521.8,13.1l-25.4-0.5l-0.2,3l25.6,0.5c3.7,0.1,7.5,0.1,11.2,0.1l0.4-3l-5.8,0L521.8,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M584.1,13.1l-5.8,0.1l-25.4,0.1h-0.1l-0.3,2.9c10.5,0,21,0,31.6-0.2l3.4-0.1l0-3L584.1,13.1z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M607.1,15.6c10.9-0.2,21.8-0.4,32.7-0.5c0.1-1,0.2-2,0.2-3c-11,0.1-22,0.3-33,0.5L607.1,15.6z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M447.8,5.5c9.8,0,19.7,0.1,29.5,0.2l0.2-3c-9.9-0.1-19.8-0.2-29.7-0.2   C445.5,2.5,445.5,5.5,447.8,5.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M521.8,3.5L497.1,3l-0.2,3l24.9,0.5c4,0.1,8.1,0.1,12.1,0.2c0.1-1,0.1-2,0.2-3l-6.5,0L521.8,3.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M584.1,6.5l3.3-0.1l0-3l-3.2,0.1l-5.8,0.1l-24.7,0.1l-0.2,2.9C563.7,6.7,573.9,6.7,584.1,6.5z" fill="rgb(193,154,107)"></path><path style="box-sizing: border-box;" d="M639.8,2.6c-11,0.1-21.9,0.3-32.9,0.5l0,3c11-0.2,22-0.4,33.1-0.5C640,4.6,639.9,3.6,639.8,2.6z" fill="rgb(193,154,107)"></path></g></svg></section></section></section></section><p><span style="font-size: 15px;">以上就是 Yangyang 此次参加 GHC 的主要感想啦。</span></p>
<p><span style="font-size: 15px;">当我们收到 Yangyang 的投稿时，真的非常感动。这样长的感受，这样深刻的自我认识，这样无私的分享，无不传递出大会给人带来的积极影响，以及源自她内心深处的能量。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">如果大家看后有所收获，也希望大家积极把这篇文章分享给周围的朋友~</span></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240" style=""></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693282&amp;idx=1&amp;sn=0f19cd6c63c7ea874ead46a78ac92eba</link>
      <pubDate>Tue, 09 Oct 2018 18:18:35 -0700</pubDate>
      <guid isPermaLink="false">899</guid>
    </item>
    <item>
      <title>职场 | 我的十年</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740" style=""></p>
<p style="margin-right: 8px;margin-bottom: 15px;margin-left: 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="color: rgb(136, 136, 136);letter-spacing: 1px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;">08年，</span><span style="color: rgb(136, 136, 136);letter-spacing: 1px;">我刚加入公司的时候，被告知在公司工作十年以上就有资格享受12周的年资假。当时并没有觉得这个福利和自己有什么关系，毕竟太过遥远。</span><br></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">可似乎转眼间，十年就这样悄无声息走到了面前。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6DfOLQMW7bppOPCbqcWu0RZnB9Sibia6HholhXpW0aKoJPa3fo5m3D6QQ/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="940" data-cropy1="0" data-cropy2="528.9667896678966" data-ratio="0.5627659574468085" data-s="300,640" data-type="jpeg" data-w="940" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6Y54P5paia3QnG8iaRYicVb5dUTEuaia4kHvxVlMmicQbzpxOUzK8fn6JLmw/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">十年中，我的角色发生过很多次变化，刚入职的时候是QA（质量保证分析师） ，后来做项目管理，之后因为意想不到的机会转职做People Lead（人力资源总监），从此一发不可收拾，我在转型的路上越走越远。十年来，我尝试过的角色不少于五个，技术、人力、运营、软件交付、咨询，还不包括兼职做下来了社区创立者、全球创新项目负责人以及培训师。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">回首过去的十年，每一次转型都有很大的跨度，也正是因为这些跨度让我这十年很丰富。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有一位前辈曾经和我说过：这家公司给不了你一个看得到的未来，但过程中的未知和不确定会让你终身难忘。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我用十年的时间验证了这句话的正确性，作为一个喜欢做梦，热爱思考，经常发呆也乐于自省的职场女性，有一些关于职场和成长的感悟，想掏心掏肺和大家分享。</span></p>
<h3 style="margin: 20px 8px 15px;font-weight: bold;font-size: 18px;max-width: 100%;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;-webkit-font-smoothing: antialiased;cursor: text;font-family: Helvetica, arial, sans-serif;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;letter-spacing: 1px;color: rgb(0, 0, 0);font-size: 16px;box-sizing: border-box !important;word-wrap: break-word !important;">1 从软件测试到人力资源总监，经历了什么</span></strong></h3>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">当初来ThoughtWorks是因为猎头的推荐，真正让我下定决心加入是因为ThoughtWorks践行的自动化测试，当时我预见自己会在测试领域摸爬滚打好多年，甚至想着测试将会是自己专注一生的领域，因为业内层出不穷的新技术、新工具、新观点让我一直很兴奋，不断涌现的市场需求也给我带来安全感，当时身边也有一些牛人指路，让我看到了上升空间，所以那时候我的志向差不多就是做测试界的百科全书。 </span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但是在我加入三年后，突然有一天，我的Sponsor（导师）过来找我，说有一个机会，问我愿不愿意尝试中国区 People Lead（人力资源总监）这个职位。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">当时这个职位已经空缺几个月了，前任离职后，中国区花了很长时间去招聘，但一直没有找到合适的人。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6q4ia3Qa2ibvIQNVn9XrdSxadZZQwRYILxGick2m1t6PsIaHhTbzcdM2sw/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="1280.0000000000002" data-cropy1="205.46125461254613" data-cropy2="935.2029520295202" data-ratio="0.5703125" data-s="300,640" data-type="jpeg" data-w="1280" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6Kic3ibgibKHba5Qlzn2yqlF9P4xZiaz42nYcW9rDubf3w1TMkyiaxwMFFNg/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">ThoughtWorks特殊的文化决定了对这个角色的要求和普通的人力资源总监不一样，这是一家专业技术公司，绝大多数的员工都是技术人员，作为人力资源总监，只有充分了解技术人员，了解他们的工作，能听懂并且理解他们讲的故事，才能真正发挥同理心，更好的关注到他们，给他们提供帮助并且支持他们的发展。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">中国区决定在内部寻找，而我就是在这个时候被识别出的候选人。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我当时很惊讶，完全没有做好准备，这个转型是我从来没有考虑过的事情。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但是有很多同事过来找我，告诉我他们的想法。</span></p>
<section class="" data-tools="135编辑器" data-id="86005" style='margin-left: 8px;max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.544px;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);border-width: 0px;border-style: none;border-color: initial;box-sizing: border-box !important;word-wrap: break-word !important;'><section class="" style="margin-bottom: 16px;padding: 16px;max-width: 100%;word-wrap: normal;overflow: auto;font-size: 15px;line-height: 1.45;border-radius: 3px;color: rgb(51, 51, 51);background-color: rgb(247, 247, 247);box-sizing: border-box !important;"><p style="margin: 5px 8px;max-width: 100%;min-height: 1em;-webkit-print-color-adjust: exact;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-family: Helvetica, arial, sans-serif;font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有人和我说，机会不错，能学到很多东西；</span></p>
<p style="margin: 5px 8px;max-width: 100%;min-height: 1em;-webkit-print-color-adjust: exact;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-family: Helvetica, arial, sans-serif;font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有人和我说，在这个职位上能和更多视野高的人合作，成长快；</span></p>
<p style="margin: 5px 8px;max-width: 100%;min-height: 1em;-webkit-print-color-adjust: exact;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-family: Helvetica, arial, sans-serif;font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有人和我说，试一下吧，失败了可以重来，年轻有什么可怕的；</span></p>
<p style="margin: 5px 8px;max-width: 100%;min-height: 1em;-webkit-print-color-adjust: exact;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);font-family: Helvetica, arial, sans-serif;font-size: 14px;letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">还有人和我说，这是ThoughtWorks的风格，鼓励每个人挑战意料之外的事情。</span></p></section></section><p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">真正影响我做决定的是这句话：多年以后，你可能不会后悔自己做过什么，但是你可能会后悔自己没有做过什么。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">总之，没有一个人告诉我放弃。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">直到现在，我还是很喜欢ThoughtWorks保留下来的这种文化，有时看似是一个人的事情，但会有很多人过来提供建议，他们不会干预你的决定，但是他们会毫无保留的和你分享他们的想法。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">他们所有的人都鼓励你勇敢，尝试新的事情，他们容忍失败，认可每个人突破自己。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">很多年以后，我发现自己成长最快的时候都是从做大胆决定那一刻开始的。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6gUuZEFLwJ8CmV1OPYoCLKoibg2YIApq5kvKYzqcVgk9BaskcjyrVCqg/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="1050" data-cropy1="0" data-cropy2="534.4086021505376" data-ratio="0.5095238095238095" data-s="300,640" data-type="jpeg" data-w="1050" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6eiamo2xc5HSCLECM2LTGDV7l1ZKibSuaMFPllVcKmlEkDbLnOias3n1NA/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">在新的岗位上，我接手的第一件事情就是给员工销假，我怎么也没有想到，本来是怀着为人民服务的理想踏上岗位的，但上岗之后做的第一件事情竟然是做“恶人”，赤裸裸“剥夺”员工的利益，“剥夺”以前和我紧密合作的同事的利益。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那时候公司的规模开始增长，但是各种制度还不健全，有的制度还成了公司发展的硬伤，比如休假制度，以前的规定是员工可以无限期延长自己的年假，有的老员工累积的年假已经超过50天，这在财务上是一种负债，对于一家专业服务公司来讲，这种负债越大，公司运营的风险就越高。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">当时中国区还不盈利，一直是全球的投资在支持，所以即使销掉员工的一部分假期，也没有办法按照百分百的比例给予补偿。对于假期累计很多天的同事来讲，这是难以接受的事情，包括我自己。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">在职场上，甚至在人生的某些个阶段，你可能都有机会遇到这样的事情，理性告诉你这样去做，但是情感上你说服不了自己，或者站在一个立场上你是在做正确的事情，但换个角度你就罪大恶极，这种问题没有答案，考验你的是如何在获得他人理解的基础上解决问题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">骨头非常硬，但还是要啃。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我和所有受影响的同事一对一沟通，解释上下文，解释为什么做这个决定，倾听他们的抱怨，接纳他们的情绪，有一位老外同事在我面前拍了桌子，指着我的鼻子发泄他的愤怒。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那段时间，每个晚上我都失眠，感觉自己站到了喜欢的那些人的对立面。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那时候，我想到了放弃或者离职，但我上岗不过才一个星期。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-copyright="0" data-ratio="0.553125" data-s="300,640" data-type="jpeg" data-w="1280" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6qAQibH6ZlADckvoazWngPDVH6RvIc70yZohrQFxqfaic4GQG3g8gk83w/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">后来想想，那其实就是不舒适区，突然跳进去，没有半点过渡，我由一个关注事情的人突然变成要关注人，有太多的技能我需要学习，沟通能力，影响力，自我管理等都需要提升。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">没有办法的话，死扛着也会有转机。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我完成了所有的谈话，也完成了设定的目标，这中间有些人让我很感动，他们理解这个决定也非常配合我推进这件事情，我知道他们是基于对我的信任。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">后来，那个老外过来和我道歉，在他理解了之后，还帮我推动了一些事情。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那件事情结束之后，我就开启了疯狂补课的模式，学习人力资源的系统知识，做各种主题阅读，参加研讨会学习行业经验，考取了人力资源从业资格证，我还给在校大学生做了一百个小时以上的职业咨询，在几个人力资源平台做签约作者。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">真正到了一年以后，我才敢和别人说，我是人力资源从业者，尽管那个时候我仍然在不舒适区，因为我还有很多东西需要学习，我也还有很多解决不了的问题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但时间是不等人的，新的机会总是会在你还没有准备好的时候就出现了。</span></p>
<h3 style="margin: 20px 8px 15px;font-weight: bold;font-size: 18px;max-width: 100%;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;-webkit-font-smoothing: antialiased;cursor: text;font-family: Helvetica, arial, sans-serif;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;letter-spacing: 1px;color: rgb(0, 0, 0);font-size: 16px;box-sizing: border-box !important;word-wrap: break-word !important;">2 再一次归零，失败了会怎样</span></strong></h3>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">2013 年，公司在全球的领导团队发生结构调整，时任中国区MD（中国区总经理）的郭晓调任全球CEO，中国区的其他运营角色也随之发生变动，我由People Lead 岗位转为PSM（专业服务资源管理）。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这是一个为所有项目匹配资源的角色，前提是要对中国区所有的专业人员和客户项目有充分的了解。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">项目匹配的复杂度之高是难以想象的，因为项目的需求各不一样，客户领域不一样，工作地点不一样，周期和起始时间不仅不一样还不断动态调整，相应的，在人员的分配上，不但要考虑每个人的技能，还要考虑机会是否符合个人发展意愿，是否需要出差协调家庭。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">归根结底，这份工作协调的是供求关系，困难的是供和求都在动态调整。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">不仅如此，这个工作的边界也很难定义，从密切关注商机预测，到把供求关系转化为培训计划，招聘优先级和启动全球支持，都需要参与。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那个时候中国区已经进入了一个快速发展的通道，规模的扩大已经带来了一些问题，以前是一个人处理的事情，现在需要一个团队协作来完成。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">在新的岗位上，我第一件事情就是组建团队。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-copyright="0" data-ratio="0.6666666666666666" data-s="300,640" data-type="jpeg" data-w="1050" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6WhlUyhzLn0AqCXNN168ws35RM4Ocuq9J8NKTW198ViaC1ibszNPccibUQ/640?wx_fmt=jpeg" style="background-color: rgb(238, 237, 235);border-width: 1px;border-style: solid;border-color: rgb(238, 237, 235);background-size: 22px;background-position: center center;background-repeat: no-repeat;height: 440.667px !important;box-sizing: border-box !important;word-wrap: break-word !important;width: 661px !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">团队里面的第一位RM（资源协调经理），是我们从业务部门识别出来的，我们俩在没有前人经验的基础上，摸着石头过河，与人力资源比起来，这个角色不属于任何领域，市场上也找不到相关的行业经验，所以在这个岗位上，发展空间难以预测，职业安全感相对也较低。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">三个月后，我唯一的团队成员提出了辞职，我像一个创业失败的人重新回到起点，那段时间，我自己也曾想着放弃，觉得自己不擅长从零做到一，那是在那个职位上最迷茫的一个阶段。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">后来在中国区总经理胡凯的帮助下，我去了印度，当时印度是唯一一个有RM团队的国家。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我加入了印度的RM团队，花了三个月的时间和他们一起处理印度的资源调配，和每个RM以及业务部门密切合作，和他们的专业人员沟通项目详情。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">他们也没有成熟的经验，但是我还是学到了很多东西，团队是如何协作的，有什么流程，在规模扩大的时候可能会面临一些什么样的坑。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">现在想想，我之所以坚持了下来，是看到了这家组织对失败的宽容，要想改变，就不可能一帆风顺。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">而且当时整个中国区的管理团队也都比较年轻，在没有人指路的情况下，大家是在一起战胜困境，探讨可能的。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我从印度回来后，立即开始组建团队，半年后，每个大的办公室都有了专职的RM，小的办公室也有兼职协调资源的人。在这个过程中，我也有了一些思考，因为工作内容复杂多变，所以很难给从事这份工作的人做严格的职责定义，但这份工作实践起来却非常锻炼人的抗压能力，应变能力，适应性和敏感性。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">更为重要的是，这份工作能让人看清专业服务公司是如何运作的。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">若干年后，我再次进入业务部门，和一家战略咨询公司合作，在一次吃饭的闲谈中，我理清了他们团队背后的配置逻辑，也帮他们解答了专业服务公司如何配置资源的问题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6zyXZvGnLcqWG13loC1dao97gRWU8BiaEot0tvoLeuBpjfkxrpUodEdg/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="1280" data-cropy1="0" data-cropy2="781.6974169741698" data-ratio="0.61015625" data-s="300,640" data-type="jpeg" data-w="1280" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6H7aWGg9T59wnicFG5RjpJFhTQEJygiavo9Go76Se1XyvEnKxX3UqjSxQ/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">那时候想到的是《法华经》里的一个词“功不唐捐”。我们学到的知识，做过的事情，都会在未来的某一天以经验的方式回馈给我们。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">现在，Staffing（资源协调）这个团队仍在公司担任着非常重要的运营职责，这个团队也走出来一些很出色的leader（带头人和管理者），想起来是件很欣慰的事情。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">我在这个岗位做了三年，直到后来休产假。</span></p>
<h3 style="margin: 20px 8px 15px;font-weight: bold;font-size: 18px;max-width: 100%;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;-webkit-font-smoothing: antialiased;cursor: text;font-family: Helvetica, arial, sans-serif;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;letter-spacing: 1px;color: rgb(0, 0, 0);font-size: 16px;box-sizing: border-box !important;word-wrap: break-word !important;">3 做不到家庭和工作平衡的话，就做一个无所畏惧的女性</span></strong></h3>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">还在做People Lead的时候，有一天，有位女同事找我倾诉，她的孩子那时不满一周岁，她说：</span></p>
<section style='max-width: 100%;font-family: -apple-system-font, system-ui, "Helvetica Neue", "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;box-sizing: border-box !important;word-wrap: break-word !important;'><blockquote label="Copyright Jason Ng of Kenengba and knbknb" style="margin-right: 30px;margin-left: 32px;max-width: 100%;border-left-width: 5px;border-left-color: rgb(38, 189, 207);line-height: 25.6px;box-sizing: border-box !important;word-wrap: break-word !important;"><p style="margin-right: 5px;margin-bottom: 15px;margin-left: 5px;max-width: 100%;min-height: 1em;white-space: pre-wrap;line-height: 1.65em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;font-family: Helvetica, arial, sans-serif;font-size: 14px;box-sizing: border-box !important;word-wrap: break-word !important;">我加班的时候觉得对不起娃，在家里陪娃又觉得对不起我的工作，做女性为什么这么难，做什么都feel guilty（带着负罪感）。</span></p></blockquote></section><p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这就是我当时还理解不了的处境。在几年之后我理解了，因为我也成了母亲。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">当了母亲之后，就特别关注工作生活平衡的话题，这可能也是绝大多数职场女性回避不了的话题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">《奇葩大会》有一期，女强人刘楠和 5 位男性CEO一起被采访，男性CEO被问到的问题，都是关于用户，关于利润的，到了刘楠，她被问到的问题是，你是如何平衡工作和生活的？</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这揭示了当下社会和职场的一个普遍认知，男性只要管好自己就行了，女性则需要平衡事业和家庭。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">产假回来，我再一次转型，调到内部IT 部门做项目管理，共有5 个团队，服务全球。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这个转型的决定其实也是考虑了我产后回归，更好的平稳过渡而做出的。这个工作在内容上我已经很熟悉，但实际中仍然需要克服一些问题。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">最大的问题就是时差，在一个全球的团队里，很多的会议不是发生在清晨就是发生在晚上，这也就预示着在家里开会是一个常态，而我家的小朋友很小，还是听不进任何道理的年纪。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">开会的时候我通常把自己关在屋里，任凭他在外面捶打着门撕心裂肺哭，我也不能分身，但这会极大的影响我的注意力和效率，我觉得私人时间里陪伴不了他，对不起小娃，开会的时候做不到全神贯注，又愧疚于信任我的同事。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">生活就是个矛盾的综合体。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6gkGxMyoU9zDTRIaah8PJVTDuViaicvcS3n7qTziaU8K0vnLjtjYvok4ibQ/0?wx_fmt=jpeg" data-cropx1="0" data-cropx2="1280" data-cropy1="40.147601476014756" data-cropy2="805.3136531365313" data-ratio="0.596875" data-s="300,640" data-type="jpeg" data-w="1280" src="https://mmbiz.qpic.cn/mmbiz_jpg/aaVJqS7LaMIMJYXWkAEyrMJUlnDL6Fib6g7ypE3Ec27YoYEItcIBLeLFdsoKOL1Vnj6vV9mElyIiaw2KbWE9DBsw/640?wx_fmt=jpeg" style="box-sizing: border-box !important;word-wrap: break-word !important;width: 677px !important;visibility: visible !important;"></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">第二个问题就是个人发展，因为孩子小，在养育中必然会分心，所以我其实是做了一个放缓自己发展的决定。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这是一个取舍的过程。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">Facebook的首席运营官谢丽尔.桑德伯的畅销书《向前一步》里介绍了很多方法，鼓励女性冲破性别障碍，做好工作生活平衡。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有些女性觉得这本书确实给她们带来了改变，但也有一些女性做不到，在兼顾工作和家庭的同时，她们既忙碌又焦虑，因为大部分的女性之所以家庭和事业都美满，是因为自己的另一半或者家里的老人们在为她们负重前行。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">也就是说，所谓的平衡，不是自己做出妥协，就是家人做出牺牲。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">就连桑德伯自己也在书里提到，她能在职场打拼，没有后顾之忧，是因为她的丈夫分担了很多家庭责任。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">这其实是一种方法，或许是最能支持到女性在职场发展的方法了。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">前面提到我所在的团队的特殊性，晚上开会的时候如果不是特别晚，七八点钟的会议我大都选择在公司开，这样的话我就可以在五六点钟的时候去健身房跑跑步练练器械，回到公司正好赶上开会，回家的路上我会觉得精神焕发、动力十足，这段时间是我的家里人在承担责任。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但周末的时候，我会全心全意陪伴孩子，陪伴家人。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">所以，在家庭工作平衡这件事情上，我是比较幸运的，既放飞了自我，也没有远离家庭，因为我有非常支持我的家里人。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但对于一些女性来讲，可能并不会如此幸运，网络上有很多建议，我比较认可的做法是这样的：</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">如果真的无法平衡，那就选一个自己最想成为的角色，起码是现阶段最想成为的角色。</span></strong><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">但是不管你选了哪一个，都意味着你要放弃另一个，不要怕，也不要担心，做一个无所畏惧的女性。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">有时想想，我们的人生就像坐标系，横轴已经订好了，区别在于纵轴上如何发挥，十年可以很短，转瞬即逝，十年也可以很长，因为成长永无止境。</span></p>
<p style="margin: 15px 8px;max-width: 100%;min-height: 1em;color: rgb(51, 51, 51);letter-spacing: 0.544px;text-align: justify;white-space: normal;-webkit-print-color-adjust: exact;font-family: Helvetica, arial, sans-serif;font-size: 14px;text-size-adjust: auto;background-color: rgb(255, 255, 255);line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(136, 136, 136);letter-spacing: 1px;box-sizing: border-box !important;word-wrap: break-word !important;">最后，18年伊始，用送给自己的话来结束吧：千帆过尽，愿自己仍是那个明眸善睐的女子，永葆一颗少年心！</span></p>
<p style="margin: 0em 8px 20px;max-width: 100%;min-height: 1em;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;font-family: 微软雅黑;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">本文转载自ThoughtWorks公司。<br></span></strong><span style="color: rgb(0, 0, 0);"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">作为「全球最佳女性科技雇主」ThoughtWorks向来不拘泥于常规而致力于打破框架，突破极限，寻求机遇。程序媛的日常在去年也曾经跟ThoughtWorks合作过</span></strong></span></strong></span><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692638&amp;idx=1&amp;sn=27ab27d01c75acb08683a1a91fc03ed1&amp;chksm=81233927b654b0311fe2d060ddd9fc8808157115986ffbde22ca5989425189908a487ccf6487&amp;scene=21#wechat_redirect" target="_blank">线下活动 | 全球最佳女性科技雇主 ThoughtWorks 的秘密</a></span></strong></span></strong><span style="color: rgb(0, 0, 0);"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">。</span></strong></span></strong></span></p>
<p style="margin: 0em 8px 20px;max-width: 100%;min-height: 1em;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;font-family: 微软雅黑;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(0, 0, 0);max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">ThoughtWorks 希望验证女性能在技术领域承担一些非常关键的岗位。但是现在行业里并没有许多这样的女性技术人员，所以需要自己培养，所以 ThoughtWorks 开展这些活动和方式来自己培养。这是远大的 ambitious vision。</span></strong></span></strong></p>
<p style="margin: 0em 8px 20px;max-width: 100%;min-height: 1em;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;font-family: 微软雅黑;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(0, 0, 0);max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">说起短期目标，就是因为 ThoughtWorks 要求每年校招进来的都是男女比1：1，这很难，只能自己先尽早培养。而个人认知——靠个人，在这一点上，ThoughtWorks 和我们 GirlsWhoCode 是非常相似的。我们都希望让大家认识到女性程序员也是多姿多彩的，不是千篇一律的，不是形象差、目光呆滞的。</span></strong></span></strong></p>
<p style="margin: 0em 8px 15px;max-width: 100%;min-height: 1em;text-align: justify;white-space: normal;color: rgb(62, 62, 62);font-size: 16px;background-color: rgb(255, 255, 255);letter-spacing: 1.5px;font-family: 微软雅黑;line-height: 1.75em;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="color: rgb(0, 0, 0);"><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">想要进一步了解吗？</span></strong></span></strong></span><strong style="max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;"><strong style="color: rgb(62, 62, 62);font-size: 16px;max-width: 100%;box-sizing: border-box !important;word-wrap: break-word !important;"><span style="max-width: 100%;color: rgb(171, 25, 66);font-size: 12px;box-sizing: border-box !important;word-wrap: break-word !important;">ThoughtWorks校招火热进行中，详情戳“阅读原文”，欢迎大家投简历～推荐人写“大S”据说有惊喜哟！</span></strong></span></strong></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.4514877102199224" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNpKic4DiblDA75KduXD0EeqoZ8McBkcR5DUfib3SQ5IiaNYpzLW98Valvic1XIQmFEIbibgFwZUPNb69vcg/640?wx_fmt=png" data-type="png" data-w="773" style=""></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693273&amp;idx=1&amp;sn=7cbc76c4277172e260b00d1269d75ad1</link>
      <pubDate>Wed, 12 Sep 2018 02:47:03 -0700</pubDate>
      <guid isPermaLink="false">898</guid>
    </item>
    <item>
      <title>不用 KL annealing，将注意力机制建模为隐变量</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p><em><span style="font-size: 15px;">作者：Harvard NLP group (harvardnlp)</span></em></p>
<p><em><span style="font-size: 15px;">论文：</span></em><em><strong><span style="font-size: 15px;">Latent Alignment and Variational Attention</span></strong></em></p>
<p><em><span style="font-size: 15px;"><span style="font-size: 15px;">地址：</span></span></em><span style="font-size: 15px;"><span style="font-size: 15px;"><span style="font-size: 15px;">https://arxiv.org/pdf/1807.03756.pdf</span></span></span></p>
<p><em><span style="font-size: 15px;"><span style="font-size: 15px;">代码：</span></span></em><span style="font-size: 15px;"><span style="font-size: 15px;">https://github.com/harvardnlp/var-attn </span></span></p>
<p><em><span style="font-size: 15px;">摘要：</span></em><span style="font-size: 15px;">attention 注意力模型在神经网络中被广泛应用。在已有的工作中，attention 机制一般是决定性的而非随机变量。我们提出了将 attention 建模成隐变量，并应用 VAE 和 policy gradient 训练模型。在不使用 KL annealing 等 trick 的情况下训练，在 IWSLT 14 German-English 上建立了新的 <span style="font-size: 15px;">state-of-the-art</span>。</span></p>
<p><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 10px;margin-bottom: 10px;text-align: center;box-sizing: border-box;"><section class="" style="border-bottom: 1px dotted rgb(160, 160, 160);padding-top: 3px;width: 100%;margin-bottom: -3px;box-sizing: border-box;"></section><section class="" style="height: 5px;line-height: 5px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;background-color: rgb(255, 255, 255);box-sizing: border-box;"><span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> </section><section style="clear: both;box-sizing: border-box;"></section></section></section></section></section><p><span style="font-size: 15px;">近年来很多文章将 VAE 应用到文本生成上，通过引入隐变量对文本中的一些不确定性（diversity，如文章风格主题、蕴含情感等）进行建模。<strong>这样做往往会遇到一个常见的问题—— KL collapsing。</strong>这个问题最早在16年时由 Bowman 指出 [1]，其描述的现象是直接训练 VAE 得到的 KL 接近0——这也就意味着近似后验和先验一样，使得隐变量被模型忽略 [5]。Bowman的解决办法 [1] 是使用 KL annealing（KL 项的权重从0开始逐渐增加到1）或者 word dropout（不常用在此略过）。随后，17年 Yang 等人对 KL collapsing 的问题进行了更细致的分析 [2]，并提出降低 decoder 的 contextual capacity 改善这个现象。此外 Zhao 等人提出 bag-of-word loss 去解决这个问题 [3]。18年 Graves 等人也对 KL collapsing 进行了分析 [4]。（我们在过去也介绍过许多解决 KL collapsing 的工作，<a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692981&amp;idx=1&amp;sn=37ed31cc6ec1ded76a53a26503ba4f93&amp;chksm=8123384cb654b15a6582beb3efb3bae1f350a12077456aded80ddb3d1993be4354bde3fc16a3&amp;scene=21#wechat_redirect" target="_blank">点此查看</a>）。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">在我们的工作中，attention 被建模成隐变量。</span><strong><span style="font-size: 15px;">值得注意的是，我们将 attention 建模成隐变量并不是为了单纯应用 VAE 这个工具，而是因为我们认为将 attention 建模成隐变量可以为 decoder 提供更 clean 的 feature，从而在不增加模型参数的情况下提高模型的表达能力</span></strong><span style="font-size: 15px;">（注意 inference network在测试阶段不被使用因此不计入模型参数）。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><strong><span style="font-size: 15px;">以下是一个简单的直觉。</span></strong><span style="font-size: 15px;">下图蓝色部分展示的是传统 attention，<span style="font-size: 15px;">下图红色部分展示的我们提出的隐变量 attention</span>。传统的 attention 机制仅能通过之前生成的单词确定当前即将生成单词的 attention，而因为存在多种翻译方式，所以会出现 attend 到和实际翻译的单词并不对应的位置的问题。而在红色部分展示的我们提出的隐变量 attention 中，我们可以通过全部的源文本和目标文本去得到更准确的后验 attention，因此 attention 和实际翻译应该 attend 的源单词对应得更好。并且，这样得到的更好的后验 attention 可以提供给 decoder，从而使 decoder 拿到更 clean 的 feature，藉此可以得到更好的模型。</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.5475285171102662" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3nsr4p7W1ibTTk2Qn5F3HI9HyiahcvT9OzUWyrsS9Xf6PvLsRjic9r1Jvg/640?wx_fmt=png" data-type="png" data-w="526" style="width: 411px;height: 225px;"></p>
<p><span style="font-size: 15px;"> </span></p>
<p><strong><span style="font-size: 15px;">基于这个直觉，我们将注意力 attention 建模成隐变量。</span></strong><span style="font-size: 15px;">假定 x 是源文本，y 是目标文本，z 是 attention，根据标准的 VAE 方法，我们引入 inference network q(z | x, y)去近似后验，那么 ELBO 可以表达为（为了简单我们考虑目标文本只有一个单词的情况）：</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.05363528009535161" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3N6qmf6mmcfHJVYf16x6vTcdaZfdDBVQTXS9rcOGkQyqnoia4icS1mqIw/640?wx_fmt=png" data-type="png" data-w="839" style="width: 435px;height: 23px;"></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">上面不等式的右侧是 ELBO，其中第一项是从 q(z | x, y) 中采样出 attention，使用采样出的 attention 作为 decoder 的输入计算 cross entropy loss，第二项是确保后验分布接近先验分布。这里值得注意的是，我们此处的先验和一般的 VAE 不同，我们的先验是和模型一起学习的。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><strong><span style="font-size: 15px;">因为我们的 p(z | x) 和 q(z | x, y) 都是 categorical 分布，所以我们使用 policy gradient 去优化上面的目标函数。</span></strong><span style="font-size: 15px;">由于 decoder 和 encoder 之间的主要信息传输通道是通过 attention，如果忽略了这个隐变量，就会无法得到源文本的信息而得到很大的 penalty。这与之前的许多工作中直接把隐变量加入到每个 decoding step 不同，因为那样即使 decoder 忽略了隐变量，也可以达到很好的模型表现 [5]。因此通过直接优化目标函数这个隐变量也不易被忽略，我们的实验完全验证了这一点。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 10px;margin-bottom: 10px;text-align: center;box-sizing: border-box;"><section class="" style="border-bottom: 1px dotted rgb(160, 160, 160);padding-top: 3px;width: 100%;margin-bottom: -3px;box-sizing: border-box;"></section><section class="" style="height: 5px;line-height: 5px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;background-color: rgb(255, 255, 255);box-sizing: border-box;"><span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"><span style="font-size: 15px;"> </span></span> </section></section></section></section></section><p><strong><span style="font-size: 15px;">由于我们的后验 q 能看到全部的 x 和 y，因此后验中采样的 attention 可以比先验 p(z | x)好。</span></strong><span style="font-size: 15px;">比如以下的例子：</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.9883527454242929" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3ic0T01mM6eYlRtIh1IFcDOIUXadFhbYl8ia8vVXovmdL1pO8QnQKHqcg/640?wx_fmt=png" data-type="png" data-w="601" style="width: 466px;height: 461px;" data-backw="233" data-backh="230" data-before-oversubscription-url="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3ic0T01mM6eYlRtIh1IFcDOIUXadFhbYl8ia8vVXovmdL1pO8QnQKHqcg/640?wx_fmt=png"></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">这里我们把德语（纵向）翻译成英语（横向）。红色代表先验，即只观测到 x 而不观测到 y 的 p(z | x)，蓝色代表后验，即观测到全部信息的 p(z | x, y)。注意到在第二个单词 <em>actually</em> 处，红色的先验试图 attend 到 <em>nun</em> 后面的逗号<em>“,”</em>，从而试图生成一个 <em>“well, ” </em>的翻译结果。然而实际的英语翻译中并没有逗号，反而直接是 <em>well actually</em>，由于后验 q(z | x, y) 可以看到实际的翻译，因此蓝色的后验正确 attend 到了 <em>tatsachlich </em>上。注意到训练目标 ELBO 中我们从 q 中采样 attention 给 decoder，因此通过使用 VAE 的方法，decoder 得到了更准确的 attention 信号，从而可能提高模型的表达能力。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 10px;margin-bottom: 10px;text-align: center;box-sizing: border-box;"><section class="" style="border-bottom: 1px dotted rgb(160, 160, 160);padding-top: 3px;width: 100%;margin-bottom: -3px;box-sizing: border-box;"></section><section class="" style="height: 5px;line-height: 5px;box-sizing: border-box;"><section style="display: inline-block;vertical-align: top;background-color: rgb(255, 255, 255);box-sizing: border-box;"><span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"></span> <span style="width: 5px;height: 5px;float: left;border-radius: 50%;border-width: 1px;border-style: solid;border-color: rgb(160, 160, 160);margin-left: 5px;box-sizing: border-box;"><span style="font-size: 15px;"> </span></span> </section></section></section></section></section><p><strong><span style="font-size: 15px;">实验上，我们在 IWSLT 14 German-English 上得到了新的 state-of-the-art。</span></strong><span style="font-size: 15px;">其中 KL大约在0.5，cross entropy loss大约在1.26，而且我们人工比较了很多后验和先验也很符合我们的建模直觉。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4455909943714822" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3PtSJQPTvxjMkpcibgzTqDALv1gspl8eXiahZkrAUmwdURObzvvefOyHA/640?wx_fmt=png" data-type="png" data-w="1066"></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">欢迎尝试我们的代码，项目地址在 https://github.com/harvardnlp/var-attn。我们提供了能复现我们 state-of-the-art 效果的 preprocessing、training、evaluation的command，以及我们报告的模型。</span><br></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><strong><span style="font-size: 15px;">相比过去的大部分工作是<span style="font-size: 15px;">从 attention 计算出来的固定维度的 context vector，</span><span style="font-size: 15px;">我们提出了将 attention 建模成隐变量，即在 simplex 上的 attention 本身。</span></span></strong><span style="font-size: 15px;"><span style="font-size: 15px;">由于</span></span><span style="font-size: 15px;"><span style="font-size: 15px;"><span style="font-size: 15px;">我们的工作是对 attention 注意力机制的改进，</span></span>因此理论上可以被应用到一切包含 attention 的 task 中。文章里除了机器翻译外我们也做了个视觉问答系统的实验。我们的具体模型和 inference network 的结构请参见我们的论文和代码。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">限于作者的水平，本文中有错误和纰漏在所难免，望读者朋友多多包涵！也欢迎发邮件给我 dengyuntian@seas.harvard.edu 交流。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><em><strong><span style="font-size: 15px;">References </span></strong></em></p>
<p><span style="font-size: 15px;">[1] Bowman et al, "Generating Sentences from a Continuous Space".</span></p>
<p><span style="font-size: 15px;">[2] Yang et al, "Improved Variational Autoencoders for Text Modeling using Dilated Convolutions".</span></p>
<p><span style="font-size: 15px;">[3] Zhao et al, "Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders".</span></p>
<p><span style="font-size: 15px;">[4] Graves et al, "Associative Compression Networks for Representation Learning".</span></p>
<p><span style="font-size: 15px;">[5] Zhang et al, "Variational Neural Machine Translation".</span></p>
<p><br></p>
<p><strong><span style="font-size: 15px;">更多相关文章阅读</span></strong></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><span style="font-size: 15px;"> <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692981&amp;idx=1&amp;sn=37ed31cc6ec1ded76a53a26503ba4f93&amp;chksm=8123384cb654b15a6582beb3efb3bae1f350a12077456aded80ddb3d1993be4354bde3fc16a3&amp;scene=21#wechat_redirect" target="_blank">干货 | 你的 KL 散度 vanish 了吗</a></span></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692433&amp;idx=1&amp;sn=b880134068f670cfa59a81b59d432045&amp;chksm=81230668b6548f7eb5bf73a696b5e918d0486e96dca8df55740a947df827d2263e951bcb2136&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">ACL 2017 | VAE for Dialog Response Generation</span></a></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692200&amp;idx=1&amp;sn=5211ddee37bc3a3cad0da93641a7081f&amp;chksm=81230751b6548e476ec2e2c91d023dfd1c7213e34dd085fa57a3f2aa4bd3253920a57aeb5281&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">ICLR 2017 | Attention 和 Memory Networks</span></a></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=212457976&amp;idx=1&amp;sn=ab503fb91c6a126fbe767f0dd37fb48a&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">干货 | Applied Attention-based Models in NLP（一）</span></a></p></li>
<li><p><span style="font-size: 15px;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=212530824&amp;idx=1&amp;sn=55ca55ddb7ce7d2cd9b316c598b6da03&amp;scene=21#wechat_redirect" target="_blank">干货 | Applied Attention-based Models in NLP（二）</a><br></span></p></li>
</ol></blockquote>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693269&amp;idx=1&amp;sn=c281f9ad65a38ce43ce88c578c12bc74</link>
      <pubDate>Fri, 20 Jul 2018 18:05:14 -0700</pubDate>
      <guid isPermaLink="false">897</guid>
    </item>
    <item>
      <title>CVPR 2018 | 加速模型收敛的新思路（控制理论+深度学习）</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p><span style="font-size: 15px;"></span>今天要分享的论文是一篇 <span style="font-size: 15px;">CVPR 2018 Spotlight，</span><strong><em><span style="font-size: 15px;">A PID Controller Approach for Stochastic Optimization of Deep Networks</span></em></strong><span style="font-size: 15px;">. 荣幸的是，我们请到了这篇论文的第一作者 Wangpeng An 为大家带来第一角度解读：从想法的萌生，到投稿的坎坷——走近一篇 CVPR 论文的诞生。 </span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"></section></section></section></section></section></section><p><span style="font-size: 15px;">具有近百年历史的“比例-积分-微分”控制方法（简称 PID 控制器）是一种通过控制系统的误差来调整机器的反馈控制方法。它是迄今为止自动控制中应用最为广泛的控制方法。这个算法被改进了将近200年无人撼动，几乎所有自动化的设备都是使用 PID 算法进行控制。那么它是怎么来的呢，请看下图：</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1.334140435835351" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGsEaLM0CT1o95dCfia0YJZtf8Oibb9ulglhEUDKIzzeHibqGU3bBic8qvJw/640?wx_fmt=png" data-type="png" data-w="413" style="width: 239px;height: 319px;"></p>
<p><span style="font-size: 15px;">这个东西就叫离心机，控制给引擎送多少油，刚开始被用在怀特兄弟造的飞机 [6]。从研究离心机如何供油，控制理论慢慢地被运用到各种各样的机器控制中，<strong>并且引导了第三次工业革命：自动化——如何控制机器让其离开人自动的运行（这个特性好像跟人工智能是一样的）。</strong>如何控制机器的算法就叫控制理论，而 PID 一直在工业控制里占统治地位，它被广泛的应用在工厂 [1]、无人机 [2] 以及无人车 [3] 中。到目前为止，90%的控制器都是 PID。<br></span></p>
<p><br></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.5093085106382979" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaG8gcswWdN92oLDARBKHUAaeIAonibezXyibtZSFec4V3MhwIny6CZsQSg/640?wx_fmt=png" data-type="png" data-w="752"></p>
<p style="text-align: center;"><span style="font-size: 15px;color: rgb(136, 136, 136);">大家熟知的大疆无人机也不例外</span></p>
<p><span style="font-size: 15px;"> </span><br></p>
<p><strong><span style="font-size: 15px;">下面就来说说控制理论和深度学习优化的关系。</span></strong><span style="font-size: 15px;">在控制理论里，<span style="font-size: 15px;">“误差”即</span>受控机器的输出与期望值之间的距离。如何根据“误差”让机器做出下一步更新的算法就是控制器。<br></span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">对于深度学习优化，将模型输出和期望输出的误差来定义为损失函数。损失函数通过执行“误差的反向传播” [1] 得到梯度来影响权重。就像 Hinton 老先生在其发表在 Nature [1] 的文章里那样描述的，梯度是每个权值对总的损失函数贡献，其与损失函数为线性关系。</span></p>
<p><br></p>
<p><strong><span style="font-size: 15px;">把深度学习里的”梯度”看成是当前权重与最优权重的“误差”，这样 SGD-Momentum就与 PID 优化器通过”梯度”与“误差”的联系而联系到了一起。</span></strong><span style="font-size: 15px;">如下图，可以看到控制过程与深度学习的优化过程的联系，可以发现控制系统里的反馈、误差、PID控制器与深度学习里的反响传播、梯度、优化器有着很高的相似度。</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.563076923076923" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGVOu2CzyTIjicEa8HsAicnEJ3DeeFUBYpJb68iczwiaFY8WNdOJPmyu8BPQ/640?wx_fmt=png" data-type="png" data-w="975"></p>
<p style="text-align:center;"><br></p>
<p><span style="font-size: 15px;">如果把控制里的误差用梯度替换掉的话，那么深度学习优化器 SGD-Momentum 和反馈控制器 PID 就有如下的关系：</span><br></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.9282945736434108" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGFwH8EqOMTdRUo8ibcXntvS2KT2wicc5RPH47QNOT7KA30MWwrqhJTTLQ/640?wx_fmt=png" data-type="png" data-w="516" style="width: 379px;height: 352px;"></p>
<p><span style="font-size: 15px;"> </span><br></p>
<p><strong><span style="font-size: 15px;">这样 SGD-Momentum 可以等效为 PI 控制器。</span></strong><span style="font-size: 15px;">而在控制理论中，PI 控制有超调的问题，也就是说 SGD-Momentum 有超调问题，这一点其实很容易理解，因为 I（Integral）是历史梯度的积累。也就是说如果历史梯度朝着同一个方向变化的话，那么朝着该方向的变化就会不断地加速，但是如果要改变方向，那么由于历史梯度还是朝着相反的方向变化，就会产生滞后反应或者也叫超调（Overshoot）的问题。</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.52" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGc7kmIkTMZISWbRALlLNsCLvkSPEebd5nauic3cqzoDEiaJ56bMZuzFEQ/640?wx_fmt=png" data-type="png" data-w="600"></p>
<p><span style="font-size: 15px;">超调也可以简单的描述为在该改变方向的时候还是沿着之前错误的方向继续行进，也就是执着于“过去”。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><strong><span style="font-size: 15px;">而 PID 控制算法中的 D 的这一项就专门是用来较小超调这个问题，因为 D 是梯度的变化趋势，其放眼“未来”。</span></strong><span style="font-size: 15px;">下面的 gif 是我见过解释 PID 这三个参数意义最好的一个动画。从中可以看出 PID 控制器中 PID 前的三个超参数如何根据误差从初始值（0）收敛到最优值（1）的一个动画。</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.75" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNpq7mMy3oeqdavibfYVXj8l3vdJkdoNWSBzUe8ky5nyibNUmmoiccenuTkgaKzB9fWzXuHcAX5V9xVRw/640?wx_fmt=gif" data-type="gif" data-w="400"></p>
<p><span style="font-size: 15px;"></span></p>
<p><span style="font-size: 15px;">根据这个观察我们就顺其自然地提出了 PID 优化器，也就是过去、现在、未来同时存在。过去、现在、未来，究竟哪个比较重要我们不知道，但是我们认为“未来”是必须要有的！</span></p>
<p><span style="font-size: 15px;"> </span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.34150076569678406" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGAq3jwR2a9ayCbRz2Sg85pNFtP1licibtbniaMPib5d71DwEz0OYhmQbVew/640?wx_fmt=png" data-type="png" data-w="653" style="width: 405px;height: 138px;"></p>
<p><span style="font-size: 15px;"> </span><br></p>
<p><span style="font-size: 15px;">由于我们证明了 SGD-Momentum 其实是 PI 控制器，我们新的算法只在 SGD-Momentum 上加入了梯度的变化这一项。我们也并没有彻底地探索如何更好的设置 P I D这三个参数，为以后的创新也留有广阔空间。：）</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">为了验证我算法能在真实数据上起作用，我们在2维数据上做了可视化的研究。</span></p>
<p><span style="font-size: 15px;">优化一个包含 x 和 y 的实验，目标函数以及结果如下：</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.3815413891531874" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGkqZPiaeRoYUkIVYFHNqezIalveO3sGFCvnomZELQQEVJPic3ta2osFrg/640?wx_fmt=png" data-type="png" data-w="1051"></p>
<p><span style="font-size: 15px;">其中背景的颜色代表f的大小，越红的地方其值越大。黄色的轨迹是 SGD-Momentum的收敛路线，D 前的超参数为 K_d，红色是在不同 K_d 下 PID 优化器的收敛路线。可以从图中看出在同样步长的条件下，K_d 这个超参数可以减小甚至消除超调问题。</span><br></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;">然后是在 MNSIT 数据集上的结果:</span></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.7538265306122449" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqW047ebTyNklIgicFp7T4iaGEt2F2TIhZYFSzG2icRibAXNLInicVIxic7k9zPeSKtZflrbGqNwRlD8cTQ/640?wx_fmt=png" data-type="png" data-w="784" style="width: 471px;height: 355px;"></p>
<p><span style="font-size: 15px;">可以看到 PID 优化器明显地好于 SGD-Momentum。</span><br></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"><span style="font-size: 15px;"> </span><br></section></section></section></section></section></section><p><span style="font-size: 15px;">觉得在控制理论里下面的工作中也可以把其误差用梯度替代掉从而提出新的深度学习优化算法：</span></p>
<p><span style="font-size: 15px;">1. 现代控制理论。</span></p>
<p><span style="font-size: 15px;">2. 模糊控制器。</span></p>
<p><span style="font-size: 15px;">3. 模型预测控制。</span></p>
<p><span style="font-size: 15px;"> </span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"><br></section></section></section></section></section></section><p><strong><span style="font-size: 15px;">下面说说做这个工作是如何产生以及最后投稿被接收的：</span></strong></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">其起初就是在折腾 SGD-Momentum 的公式的时候，用了一个高中就会学到的数列求和的技巧化简了一下，发现有一项是历史梯度的求和……在这个过程中不断地把东西往一块凑，不断地丰富实验，丰富理论。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">但截稿前其实没有写完论文，投出去之后发现虽然做了两个实验，然而其中一个结果还忘了贴上去，论文中提到的一个图还并不存在。当时心情复杂，很多师兄师姐都告诉我毕竟是第一次投还只写了7页可以原谅，老师建议我再好好改一下准备投 ECCV。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;"></span></p>
<p class="p1"><span style="font-size: 15px;">在跑 CIFAR 10 实验的时候发现加入梯度的变化那一项并没有 SGD-Momentum 好，当时心情沮丧觉得弄不好做了这么长时间的 idea 可能就是不 work。</span></p>
<p class="p1"><br></p>
<p class="p1"><span style="font-size: 15px;">之后就开始了很长一段时间每天起床，出门，实验室坐下打开 Youtube 和 Github 找事情做的过程。有一次在 Youtube 看到刘媛媛的励志演讲，看完充满了力量异常的有动力。联想起自己做的工作，脑中一闪发现其实并没有人限制我是不是要给梯度的变化这一项前面加的那个超参数必须是设定好不可调节的……于是我就开始了漫长的调参之路。</span></p>
<p class="p1"><br></p>
<p class="p1"><span style="font-size: 15px;"><strong>其实现在也觉得如何面对可能不 work 的 idea 是一门艺术，那种被偶然发现，纯天然无污染的 idea 太少， 大部分可能都是人造 idea。</strong>只要意志坚强，总能做完投出去。</span></p>
<p class="p1"><br></p>
<p class="p1"><span style="font-size: 15px;">最后把她说的一句话送给大家：</span></p>
<p class="p1"><span style="font-size: 15px;"><em>命里有时终须有，命里无时须强求。努力就有 CVPR！</em></span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"><br></section></section></section></section></section></section><p><span style="font-size: 15px;">概括一下，这篇论文（<span style="font-size: 15px;">CVPR 2018 Spotlight，</span><strong style="white-space: normal;"><em><span style="font-size: 15px;">A PID Controller Approach for Stochastic Optimization of Deep Networks</span></em></strong><span style="font-size: 15px;">）</span>介绍了一个简单的<span style="font-size: 15px;">优化方法：</span>修改 SGD + Momentum 使模型更快收敛。其 Motivation 主要来自自动化学科的比例积分微分（PID）控制算法。作者证明了 SGD -Momentum 是一个 PI 控制器（缺乏一个D，与 PID 更新相比方程）。通过加入 D 对 SGD-Momentum 进行改进从而实现加速收敛，而 D 就对应梯度的变化。</span></p>
<p style="white-space: normal;"><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"></section></section></section></section></section></section><p style="white-space: normal;"><span style="font-size: 15px;">这篇论文的 Reviewers 们是这么说的：</span></p>
<blockquote>
<p style="white-space: normal;"><span style="font-size: 15px;">1. I think the paper makes a very interesting connection between two classical approaches, respectively from the field of feedback control and the field of optimization.</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"> </span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">2. The link between PID and SGD+Momentum opens the door to further innovations in optimization of deep networks. I wish the ML community would pick this up and go much further.</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"> </span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">3. I actually believe the authors would do better by performing additional MLP and LSTM experiments, and then submitting this paper to ICLR, NIPS. I don't think CVPR is the right venue and I fear it will not give theauthors the exposure the work deserves.</span></p>
</blockquote>
<p style="white-space: normal;"><span style="font-size: 15px;"> </span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">被多次问过，为什么研究改进 SGD-Momentum 而不是 Adam 呢。因为目前卷积神经网络里最好的模型 VGG、Resnet、DenseNet、Mask-RCNN 等都是由其训练。最近也有人指出 Adam 这种自动化的学习速率方法其实并没有 SGD-Momentum 训练的模型效果好 [2]。简单来说就是 SGD-Momentum 就是一台单反，自己调节一些参数可以达到很好的效果，而 Adam 就像是一台傻瓜相机，虽然大多数时候都行，但追求极致还是得用“单反”。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"> </span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">因为 PID 一直是控制理论中的黑箱，其参数的调节也如炼丹一样的 CNN 至今无人完全参透，所以知乎上也有这样的评论：</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">PID，CNN，双壁合一，但凡练成者，江湖人送外号：“调参圣手”。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 0.5em;margin-bottom: 0.5em;box-sizing: border-box;"><section class="" style="display: inline-block;float: left;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="display: inline-block;float: right;width: 45%;margin-top: 0.2em;border-width: 1px;border-style: dotted;border-color: rgb(160, 160, 160);box-sizing: border-box;"></section><section class="" style="width: 0.5em;height: 0.5em;margin: auto;box-sizing: border-box;"><section style="transform: rotate(45deg);-webkit-transform: rotate(45deg);-moz-transform: rotate(45deg);-o-transform: rotate(45deg);box-sizing: border-box;"><section style="width: 0.5em;height: 0.5em;background-color: rgb(160, 160, 160);box-sizing: border-box;"></section></section></section></section></section></section><p style="white-space: normal;"><strong><em><span style="font-size: 15px;">References: </span></em></strong></p>
<p><span style="font-size: 15px;">[1] D. E. Rumelhart, G. E. Hinton,and R. J. Williams, “Learning representations by back-propagating errors.”  Nature, 1986.</span></p>
<p><span style="font-size: 15px;">[2] The Marginal Value of AdaptiveGradient Methods in Machine Learning- arXiv</span></p>
<p><span style="font-size: 15px;">[3] Knospe, Carl. "PID control." IEEEControl Systems 26.1 (2006): 30-31.</span></p>
<p><span style="font-size: 15px;">[4] A. L. Salih, M. Moghavvemi,H. A. F. Mohamed, and K. S. Gaeid, “Modelling and pid controller design for aquadrotor unmanned air vehicle,” International Conference on Automation.</span></p>
<p><span style="font-size: 15px;">[5] P. Zhao, J. Chen, Y. Song, X.Tao, T. Xu, and T. Mei, “Design of a control system for an autonomous vehiclebased on adaptive-pid,”International Journal of Advanced Robotic Systems.</span></p>
<p><span style="font-size: 15px;">[6] https://en.wikipedia.org/wiki/Control_theory.  </span></p>
<p><span style="font-size: 15px;"> </span></p>
<p><span style="font-size: 15px;"> </span></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693262&amp;idx=1&amp;sn=a2f0dfc40664b377d1ba9b7735629cfa</link>
      <pubDate>Thu, 19 Jul 2018 19:30:24 -0700</pubDate>
      <guid isPermaLink="false">896</guid>
    </item>
    <item>
      <title>干货 | 人脑与语言——语言模型解码与语义空间映射</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p><span style="font-size: 15px;">人类大脑认知和语言功能之间的关联奥秘，一直是研究者们关注的焦点。在本科的脑认知课程上，小S就听到老师说，人类如何处理语言、形成语言，已经是二十一世纪几乎最受关注也最“一筹莫展”的脑认知研究领域。<br></span></p>
<p><span style="font-size: 15px;"><br></span></p>
<blockquote><p><em><span style="font-size: 14px;">Computational psycholinguistics has “always been...the thing that computational linguistics stood the greatest chance of providing to humanity” (Kay, 2005).</span></em></p></blockquote>
<p><br></p>
<p><span style="font-size: 15px;">随着统计语言模型，神经网络语言模型等的出现，脑认知和自然语言处理这两个学科的交叉研究——<strong>计算神经科学</strong>——也不断涌现出来。曾经的爆款 Science 论文《Human-level concept learning through probabilistic program induction》也属于这个范畴。今天又看到三篇这个领域的研究工作，想分享给大家，包括：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><span style="font-size: 15px;"> <strong>Finding Syntax in Human Encephalography with Beam Search</strong>. ACL 2018.</span></p></li>
<li><p><strong><span style="font-size: 15px;">Navigating with Graph Representations for Fast and Scalable Decoding of Neural Language Models</span></strong><span style="font-size: 15px;">. from Microsoft.</span></p></li>
<li><p><strong><span style="font-size: 15px;">Does the brain represent words? An evaluation of brain decoding studies of language understanding</span></strong><span style="font-size: 15px;">. from MIT.</span></p></li>
</ol></blockquote>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 100%;overflow: hidden !important;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 311.5 25.1" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" width="311.5"><g style="box-sizing: border-box;"><g style="box-sizing: border-box;"><rect x="165.1" y="17.1" transform="matrix(0.7073 -0.7069 0.7069 0.7073 34.8572 124.992)" width="6.6" height="6.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><path d="M56.7,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C57.9,14,56.7,12.8,56.7,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M59.4,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C64.6,14.2,62.3,16.6,59.4,16.6z     M59.4,7.2c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C63.5,9.1,61.6,7.2,59.4,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M262.4,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C263.6,14,262.4,12.8,262.4,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M265,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C270.3,14.2,267.9,16.6,265,16.6z M265,7.2    c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C269.1,9.1,267.3,7.2,265,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><rect x="144.7" y="9.1" width="6.1" height="4.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><polygon points="16.7,18.2 0,18.2 0,17.1 16.3,17.1 22.6,10.8 145.8,10.8 145.8,11.9 23.1,11.9" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="311.5,19.5 288.7,19.5 281.1,11.9 149.8,11.9 149.8,10.8 281.6,10.8 289.1,18.3 311.5,18.3" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="282.5,8.9 271.7,8.9 271.7,7.8 282,7.8 289.5,0.4 311.3,0.4 311.3,1.5 289.9,1.5" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="47.9,7.9 22.7,7.9 15.9,1.1 0,1.1 0,0 16.4,0 23.2,6.8 47.9,6.8" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="201.6,21 167.5,21 162.2,15.7 125.8,15.7 125.8,14.6 162.7,14.6 168,19.8 201.6,19.8" fill="#000000" style="box-sizing: border-box;"></polygon></g></g></svg></section></section></section></section><p><span style="font-size: 15px;">发表在 ACL 2018 上的第一篇工作 [1] 主要关注的是语言模型解码中的语法信息，。比起我们非常熟悉的 LSTM，</span><span style="font-size: 15px;text-decoration: underline;">这篇工作 [1] 使用的基本语言模型是 RNNG，G for Grammers.</span><span style="font-size: 15px;"> 由下面的示意图可以看到，RNNG 在生成过程中，不仅生成具体的词，还生成语法结构。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.5537017726798749" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpRKEsE5e0zIBWG1991tuLCCIDMYlCVZxvFBhAyT3csjSuqkhmmdb7Yg/640?wx_fmt=png" data-type="png" data-w="959" style="width: 387px;height: 215px;"></p>
<p style="text-align: left;"><span style="font-size: 15px;">这样的结构非常适合增量过程建模（incremental processing）。可以想到，当利用这样的生成过程做语法建模时主要包含三种语法分析决策（parser action）：是否开始一个新的成分（左括号），是否结束最近的一个成分（右括号），以及是否生成一个具体的词。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="1.222707423580786" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpFFlbpAPgorXsQ4xawFh7WHyzEoYAq2JC8yPTic1YkiaTxkMJO91A06kg/640?wx_fmt=png" data-type="png" data-w="458" style="width: 373px;height: 455px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">上方给出的示意图的左上角就是三个彩色的方块就代表着三种决策动作。然而，经典的 beam search 算法用于这样结合语法信息的语言模型时，会有些不合适。毕竟语法信息和具体的词，不是一个层次上的信息，把它们放在一起去预测得到的概率分布会严重的不均衡。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.2976076555023923" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp7mZTSwbsErGsnvLDq5zKicVplExHdzFtqPySJADPichdRvuB2YIicq1QA/640?wx_fmt=png" data-type="png" data-w="1045" style="width: 428px;height: 127px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">为此，作者 [1] 使用的是 word-synchronous beam search。其主要思路是在解码时，不断搜索扩展语法结构，直到有足够多的结构都生成了具体的词，然后同步这些词，再开始下一次的结构解码。示意图见下，具体的算法描述请大家参考论文 [1]。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.42799597180261834" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp5vupXiaLPL0an1t8dvX157Y1MGnypaFdib51B6ohNtibdWEmfRUZGcp9g/640?wx_fmt=png" data-type="png" data-w="993" style="width: 389px;height: 166px;"></p>
<p style="text-align: left;"><span style="font-size: 15px;text-decoration: underline;">作者 [1] 希望探究 RNNG+word-synchronous beam search 是不是一个合适的语言认知模型（cognitive model of language processing）</span><span style="font-size: 15px;">。为此，他们比较了人脑听到一段语言时的 EEG 信号和语言认知模型计算过程中的复杂度指标（具体请见论文 Section 4）。从下图的部分实验结果示意来看，回归模型还是比较契合的。</span><br></p>
<p style="text-align:center;"><img class="" data-ratio="0.41259500542888167" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpR06gEzgiaicjLglja0mFOoITG7aYteiajwJWAPHibAYdjtraEXWkHpuvzw/640?wx_fmt=png" data-type="png" data-w="921" style="width: 465px;height: 192px;"></p>
<p><span style="font-size: 15px;">这篇工作 [1] 里还有很多有意思的 insight analysis，比如实验结果就发现，LSTM 的回归效果就比较差，说明不适合作为一个语言认知模型使用。<span style="text-decoration: underline;">而且这篇论文 [1] 给我的最大的冲击还是他们并不是使用 RNNG 直接建模 EEG 数据，而是用 RNNG 中的计算语言复杂度来建模。</span>这可能也是纯自然语言处理和计算语言处理的一个比较大的区别。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 100%;overflow: hidden !important;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 311.5 25.1" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" width="311.5"><g style="box-sizing: border-box;"><g style="box-sizing: border-box;"><rect x="165.1" y="17.1" transform="matrix(0.7073 -0.7069 0.7069 0.7073 34.8572 124.992)" width="6.6" height="6.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><path d="M56.7,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C57.9,14,56.7,12.8,56.7,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M59.4,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C64.6,14.2,62.3,16.6,59.4,16.6z     M59.4,7.2c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C63.5,9.1,61.6,7.2,59.4,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M262.4,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C263.6,14,262.4,12.8,262.4,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M265,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C270.3,14.2,267.9,16.6,265,16.6z M265,7.2    c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C269.1,9.1,267.3,7.2,265,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><rect x="144.7" y="9.1" width="6.1" height="4.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><polygon points="16.7,18.2 0,18.2 0,17.1 16.3,17.1 22.6,10.8 145.8,10.8 145.8,11.9 23.1,11.9" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="311.5,19.5 288.7,19.5 281.1,11.9 149.8,11.9 149.8,10.8 281.6,10.8 289.1,18.3 311.5,18.3" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="282.5,8.9 271.7,8.9 271.7,7.8 282,7.8 289.5,0.4 311.3,0.4 311.3,1.5 289.9,1.5" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="47.9,7.9 22.7,7.9 15.9,1.1 0,1.1 0,0 16.4,0 23.2,6.8 47.9,6.8" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="201.6,21 167.5,21 162.2,15.7 125.8,15.7 125.8,14.6 162.7,14.6 168,19.8 201.6,19.8" fill="#000000" style="box-sizing: border-box;"></polygon></g></g></svg></section></section></section></section><p><span style="font-size: 15px;">第二篇想分享的论文 [2] 则是希望通过借鉴人脑中的语义映射或者说词汇关联，来加速语法模型的解码过程（beam search decoding）。</span><span style="font-size: 15px;text-decoration: underline;">因为过去的认知科学研究多数认为，人脑中的词汇是复杂网络结构</span><span style="font-size: 15px;">，所以当给定一个语境时，人脑能联想到的词汇会锁定在网络中的某一小部分，而不是像 softmax 一样遍历一遍所有词。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;text-decoration: underline;">基于这样的启发，作者 [2] 提出了基于图的快速解码算法 Fast Graph Decoder (FGD)。</span><span style="font-size: 15px;text-decoration: none;">算法的过程也很直接，先根据词向量的语义距离构建一个小世界网络来“模拟”人脑中的语言网络。当解码时，将语境编码（隐变量）和图中的结点进行比较，从而找到最相似的结点作为图解码的起点，并启发式地进行图推理。</span><span style="font-size: 15px;text-decoration: underline;">值得一提的是，在小世界网络上的图搜索有比较成熟的快速算法，这也是这篇论文的算法能大幅提速传统 beam search 的主要原因</span><span style="font-size: 15px;text-decoration: none;">。算法的示意图如下。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpY4bOpSPvTrYhOx7ibzXshyeFCqKdyBwy8iaRTZ9WzPQGbVibcia8iaEA7ibw/640?wx_fmt=png" data-cropx1="0" data-cropx2="941" data-cropy1="0" data-cropy2="300.17562724014334" data-ratio="0.3188097768331562" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJponJT853NIewAayOkHgkz01kxC2y86polD60zcopWkqbUn0RdE15x8A/640?wx_fmt=jpeg" data-type="jpeg" data-w="941" style="width: 558px;height: 178px;"></p>
<p><span style="font-size: 15px;"></span><br></p>
<p><span style="font-size: 15px;">这样简单的思路在实验中效果还不错。在不损失准确度的情况下，其解码计算速度大幅提高。这是因为经典 beam search 主要受限于 softmax 计算，其复杂度是 O( D * |V| )，而这篇中的快速图解码 FGD 则是 O( D * log|V| )。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.34763948497854075" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp5ODkriamDrYNJcRBHS5kZSK0Me8IgKyc6jLicibshnysmxiaN0VtsdPrvA/640?wx_fmt=png" data-type="png" data-w="932"></p>
<p style="text-align:center;"><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 100%;overflow: hidden !important;box-sizing: border-box;"><svg xmlns="http://www.w3.org/2000/svg" x="0px" y="0px" viewbox="0 0 311.5 25.1" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" width="311.5"><g style="box-sizing: border-box;"><g style="box-sizing: border-box;"><rect x="165.1" y="17.1" transform="matrix(0.7073 -0.7069 0.7069 0.7073 34.8572 124.992)" width="6.6" height="6.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><path d="M56.7,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C57.9,14,56.7,12.8,56.7,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M59.4,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C64.6,14.2,62.3,16.6,59.4,16.6z     M59.4,7.2c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C63.5,9.1,61.6,7.2,59.4,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M262.4,11.3c0-1.5,1.2-2.7,2.7-2.7c1.5,0,2.7,1.2,2.7,2.7c0,1.5-1.2,2.7-2.7,2.7C263.6,14,262.4,12.8,262.4,11.3z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><path d="M265,16.6c-2.9,0-5.2-2.4-5.2-5.2c0-2.9,2.4-5.2,5.2-5.2c2.9,0,5.2,2.4,5.2,5.2C270.3,14.2,267.9,16.6,265,16.6z M265,7.2    c-2.3,0-4.1,1.8-4.1,4.1c0,2.3,1.8,4.1,4.1,4.1c2.3,0,4.1-1.8,4.1-4.1C269.1,9.1,267.3,7.2,265,7.2z" fill="#000000" style="box-sizing: border-box;"></path></g><g style="box-sizing: border-box;"><rect x="144.7" y="9.1" width="6.1" height="4.6" fill="#000000" style="box-sizing: border-box;"></rect></g><g style="box-sizing: border-box;"><polygon points="16.7,18.2 0,18.2 0,17.1 16.3,17.1 22.6,10.8 145.8,10.8 145.8,11.9 23.1,11.9" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="311.5,19.5 288.7,19.5 281.1,11.9 149.8,11.9 149.8,10.8 281.6,10.8 289.1,18.3 311.5,18.3" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="282.5,8.9 271.7,8.9 271.7,7.8 282,7.8 289.5,0.4 311.3,0.4 311.3,1.5 289.9,1.5" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="47.9,7.9 22.7,7.9 15.9,1.1 0,1.1 0,0 16.4,0 23.2,6.8 47.9,6.8" fill="#000000" style="box-sizing: border-box;"></polygon></g><g style="box-sizing: border-box;"><polygon points="201.6,21 167.5,21 162.2,15.7 125.8,15.7 125.8,14.6 162.7,14.6 168,19.8 201.6,19.8" fill="#000000" style="box-sizing: border-box;"></polygon></g></g></svg></section></section></section></section><p><span style="font-size: 15px;">今天想分享的最后一篇论文 [3] 则有些与众不同。其更多地是从脑认知的角度来探讨计算语言相关研究的价值。如上面提到过的，计算语言学中经常把脑认知的一些数据（EEG, fMRI等）和神经网络处理的文本数据联系起来，希望从中找到特征的相关性甚至一致性。这样的研究框架可以用下面这样的示意图来展示：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4841549295774648" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJpB6UGoFqMtcZHbhSR38T6XzsqsiaFbOxHjlFdoKNlfFMmx9C7AmWRvzQ/640?wx_fmt=png" data-type="png" data-w="568"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">如果将 fMRI 数据转换后的特征和经过语言模型建模后 m_k 的文本特征分别称为 r(x_i, s_j) 和 r(x_i, m_k), 那么这两者的差距越小，说明这个模型拟合某个被试（s_j）拟合得越好。为此，作者比较了下面多种模型：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.31841652323580033" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp1LFwos6YL9L3YriaRUwRGzM3gIjxf4wtbSBYu6qJic1NGSq93Mhbr42Q/640?wx_fmt=png" data-type="png" data-w="581" style="width: 449px;height: 144px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">可以看到，这些模型主要是针对不同自然语言处理相关的任务提出的。从某种程度来说，作者希望看到的是，针对不同任务的语言模型和人类阅读某一句话时的反应是否一致。<span style="text-decoration: underline;">换句话说，人类阅读某一句话是，具体在进行着哪些语言处理（不同的语言任务）。</span></span></p>
<p><br></p>
<p style="text-align:center;"><img class="" data-ratio="0.5940246045694201" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNpliacLf9Kjd2zsia2EC27dJp8xpickMUGa7IpykXhoxt9SaDfUdicmZdGA0pw6Eba1L4p60q6WlVCTMA/640?wx_fmt=png" data-type="png" data-w="569" style="width: 400px;height: 237px;"></p>
<p><span style="font-size: 15px;">实验结果来看，大部分模型学到的特征都可以一定程度上契合人脑活动数据中提取的特征。而且，表现最好的模型是 GloVe、SkipThought 等前四个模型，这些模型也可以认为是最“包罗万象”的模型。因为它们主要建模的是对下一个词的预测，并不涉及一个具体的自然语言任务（task-transferrable）。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">然而，这样的结果只能说明 GloVe 等模型学到了一些和人脑活动相关的特征，却并不能解释人脑在处理语言到底做了哪些任务。</span><span style="font-size: 15px;">如果对这方面的研究感兴趣，推荐大家看一篇综述：John Hale. 2017. </span><strong style="font-size: 15px;">Models of human sentence comprehension in computational psycholinguistics</strong><span style="font-size: 15px;">. In Mark Aronoff, editor, Oxford Research Encyclopedia of Linguistics. Oxford University Press.</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">诚然，这些工作可能很多不够研究，只是借鉴一些思路灵感。就像 CNN 的设计与视觉细胞，神经网络的设计与大脑网络，这些关联可能都不是真实准确的。但是，人工智能并不一定是复原或者说模仿人类的智能。研究工作带来的灵感可能是更宝贵的。在这方面，大家又怎么看呢？</span></p>
<p><br></p>
<p style="white-space: normal;"><span style="font-size: 15px;">更多相关阅读：<br></span></p>
<blockquote style="white-space: normal;">
<p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6&amp;chksm=81233b7fb654b269794e8246aa3c1512421a406625438a864699337a6c3da7f94afa4497fd22&amp;scene=21#wechat_redirect" target="_blank" style="text-decoration: underline;font-size: 15px;">干货 | 限定和指导 Seq2Seq 模型的生成结果（一）</a></p>
<p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693201&amp;idx=1&amp;sn=e385650379ede59df28ae504f453dc9a&amp;chksm=81233b68b654b27ee2f6c4833493e7aec053c6b14bb6afa8c1f2920df9a933e53ce95db4430a&amp;scene=21#wechat_redirect" target="_blank" style="text-decoration: underline;font-size: 15px;">干货 | 限定和指导 Seq2Seq 模型的生成结果（二）</a></p>
</blockquote>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693249&amp;idx=1&amp;sn=038815e711d5ea91fba308a6429e1785</link>
      <pubDate>Mon, 18 Jun 2018 23:03:02 -0700</pubDate>
      <guid isPermaLink="false">895</guid>
    </item>
    <item>
      <title>日常 | 端午安康，来跟你分享大S的生日寄语</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <section style="line-height: 1.5;letter-spacing: 1.5px;padding: 0px 1px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740" style=""></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-size: 14px;">转眼距离上次午夜电台又是4个月，大S回来报道！祝福大家端午安康～</span></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin-top: 10px;margin-bottom: 10px;text-align: center;box-sizing: border-box;"><section class="" style="display: inline-block;vertical-align: middle;box-sizing: border-box;"><section class="" style="max-width: 100%;display: inline-block;vertical-align: bottom;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.7948718" data-w="351" src="https://mmbiz.qpic.cn/mmbiz_gif/tXT40RmPWYZP7ibOyLsmyyueOy3yhuiaIJus3iatZfNsxYsScBzZBh6afqFZuGSsB8XDuyoJ0RQiaMIAkXuMQibaR6w/640?wx_fmt=gif" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="gif"></section><section class="tn-yzk-fuid-text-14814-1528344263398" style="display: inline-block;vertical-align: bottom;padding-left: 5px;padding-right: 5px;line-height: 1.2em;margin-bottom: 2px;font-size: 18px;color: rgb(60, 102, 77);box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-family: PingFangTC-light;box-sizing: border-box;">来吃端午粽吧</span></p></section><section class="" style="max-width: 100%;display: inline-block;vertical-align: bottom;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.8303571" data-w="336" src="https://mmbiz.qpic.cn/mmbiz_gif/tXT40RmPWYZP7ibOyLsmyyueOy3yhuiaIJbBSiarpUJ662Qt2kqWF6toKr0tV1gn2HBicmWr9yAGoaYlibyoLbUbXWg/640?wx_fmt=gif" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="gif"></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-size: 14px;">记忆中，每年炎热的夏天，都是考试的季节，而大S的生日，也往往是在准备考试/考试/等考试成绩的过程中度过。毕业工作后，好像终于迎来了第一个不再充斥着考试的生日，但生活中其实又不是何处都有“考验”呢。</span></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.1203125" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_png/t8fqR4fCPn852ALia8OgxEe3yvLia1Nx0JdBCzqLbIGBuQVe1MT8W0IFeHgHBCbsCehWEDvGOWE2xRQPZWjBFnrg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;font-size: 13px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1.0147059" data-w="68" src="https://mmbiz.qpic.cn/mmbiz_png/9LfNnibXYmxqaPnOTmCZ9vibWFA9elPibic8hEhuI62pMDmoaiaiaKQibtEGCXE5mb1mu58waOys6sRL5exQTtFNib7yWg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px solid rgb(71, 193, 168);border-bottom-left-radius: 0px;padding: 0px 0px 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">随着年纪增长，记忆力下降、熬夜修复困难等等初老症状来袭【误】，大S和身边的朋友好像也都越来越意识到身体健康的重要性。从小不擅长运动的大S，本来是非常不喜欢运动的（也可能是因为下意识不愿意面对自己的短板or可能的出丑？）所以基本能不动就不用动…而2018年至今，最大的收获大概就是终于把健身融为了自己生活的一部分，虽然我还不能拍拍胸脯说我已经爱上了运动，但是至少逐渐习惯到了日子就该去健身房找教练报道，也渐渐有了探索未知运动的好奇心，就已经是巨大的进步了。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">为了让自己拥有<strong style="box-sizing: border-box;">健康的身体</strong>，大S今年立下了一个减肥的小目标，几个月过去了，最令人惊喜的大概就是生活方式的改变了吧。仍然热爱美食，却可以吃的更加健康。瘦的7公斤里面5公斤都是脂肪，大概就是最大的礼物啦！</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: right;margin: 20px 0% 10px;transform: translate3d(-20px, 0px, 0px);-webkit-transform: translate3d(-20px, 0px, 0px);-moz-transform: translate3d(-20px, 0px, 0px);-o-transform: translate3d(-20px, 0px, 0px);box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.959375" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8o0XXvfFKWOLqma4ia4cZ55TNwkB2k3bAtiaDZEOPM0Ou5HTAmJUtf5ow/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="png"></section></section></section></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.9677419" data-w="62" src="https://mmbiz.qpic.cn/mmbiz_png/ceU3PHgscVLmeJK16r1B7ApFWuCgTMrkyXdwRcJ1KRTLEdHfy4I6Ty3MSaAwwJtu9NryZ40DxIMJCyY5ibcfhPg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px none rgb(160, 160, 160);border-bottom-left-radius: 0px;padding: 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;border-right: 1px solid rgb(71, 193, 168);border-top-right-radius: 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">生活和工作中难免要经历各种起伏与变化，如何用平常心面对这些变化，大概是大S最近正在努力学习的功课了。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">很多时候，如果自己决定了要花时间好好去做，就不用想太多的去努力。既然是自己决定要去做的，也就不要额外的去期待什么回报，在这个过程中的收获，无论是好的坏的，都将是人生中重要的财富。只要自己付出了百分之百的努力，问心无愧地对得起自己，那么很多外部的噪音也就可以不用在意了。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="font-size: 14px;text-align: justify;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">大S其实是个偏理想主义的人，重感情，不喜欢跟别人争抢，总觉得把事情做成了比什么都重要。也在渐渐锻炼自己<strong style="box-sizing: border-box;">强大的内心</strong>，学习辨别到底什么样的人值得真心的付出，而什么样的时候也要为自己多争取一点。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="line-height: 1.8;font-size: 14px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">有时候觉得自己骨子里其实是一个内向的人，虽然生活好像把我锻造成了现在这个活泼的样子，但是闲暇的时候，还是希望自己一个人窝在咖啡馆的小角落码码字看看书。【当然了天天听我在家里废话的室友估计是不同意这段的哈哈哈</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;box-sizing: border-box;"><section class="" style="display: inline-block;vertical-align: top;width: 45%;padding: 0px 10px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: 10px 0% 20px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);margin: 5px 0% -5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);margin: 5px 0% -5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 0px 0%;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;box-shadow: rgb(0, 0, 0) 0px 0px 0px;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8F6qq2ZHvju863gDGYR9Q3lBAS7JsRHTfAvAd4yAwia347BnCdU7xicGQ/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section></section></section></section></section></section></section></section></section></section></section></section><section class="" style="display: inline-block;vertical-align: top;width: 45%;padding: 0px 10px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="transform: rotateZ(8deg);-webkit-transform: rotateZ(8deg);-moz-transform: rotateZ(8deg);-o-transform: rotateZ(8deg);box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: 20px 0% 10px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(-5px, 0px, 0px);-webkit-transform: translate3d(-5px, 0px, 0px);-moz-transform: translate3d(-5px, 0px, 0px);-o-transform: translate3d(-5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 0px 0%;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;box-shadow: rgb(0, 0, 0) 0px 0px 0px;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8sL7hPK7TSoMZCxuM2Y902JPzCPdiaUoI2tHQF385AS8Yqm2xX7cgkQQ/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section></section></section></section></section></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin: 10px 0%;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1.0147059" data-w="68" src="https://mmbiz.qpic.cn/mmbiz_png/9LfNnibXYmxqaPnOTmCZ9vibWFA9elPibic8hEhuI62pMDmoaiaiaKQibtEGCXE5mb1mu58waOys6sRL5exQTtFNib7yWg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px solid rgb(71, 193, 168);border-bottom-left-radius: 0px;padding: 0px 0px 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">      “平稳的人生，一定闷。我受不了闷，我决定活得有趣。”<br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="text-align: right;font-size: 14px;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">——蔡澜</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">一直羡慕那些拥有<strong style="box-sizing: border-box;">有趣的灵魂</strong>的人，那种从骨子里散发出来的自在和“幽默”，让他们在我眼中拥有人群也遮不住的耀眼光芒。希望自己也能做个有趣的人，拥有自己的爱好与特点。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="font-size: 14px;line-height: 1.8;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">最近捡起了做手帐的习惯。重新翻出大学的时候做的黑卡，真的包含着非常多美好的回忆。这次去台湾旅行，特地找心灵手巧的同事定制了一本Junk Journal，准备记录一下这趟台湾行。有个10秒的视频给大家秀一下最终的作品：</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><iframe class="video_iframe" data-vidtype="2" allowfullscreen="" frameborder="0" data-ratio="1.7647058823529411" data-w="480" src="https://v.qq.com/iframe/preview.html?vid=t1342v9rwkl&amp;width=500&amp;height=375&amp;auto=0"></iframe><br></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="font-size: 14px;line-height: 1.8;box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;">累计加起来大概做了要有快20个小时吧…本来计划是旅行过程中每天晚上做，结果每天我除了记得盖章和买和纸胶带外，其他时间都累瘫，完全顾不上…【现在一切都电子化了真的不利于我们这些喜欢收集票根的人呀&gt;&lt;<br style="box-sizing: border-box;"></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;">回来制作的过程中真的太累了…但是事后看着那些照片都会忍不住微笑，真的是美好的回忆呀！</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.9677419" data-w="62" src="https://mmbiz.qpic.cn/mmbiz_png/ceU3PHgscVLmeJK16r1B7ApFWuCgTMrkyXdwRcJ1KRTLEdHfy4I6Ty3MSaAwwJtu9NryZ40DxIMJCyY5ibcfhPg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px none rgb(160, 160, 160);border-bottom-left-radius: 0px;padding: 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;border-right: 1px solid rgb(71, 193, 168);border-top-right-radius: 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">梦想这个词，说大不大，说小也不小。相信大家都对这个词有自己的理解，只希望能够拥有自己<strong style="box-sizing: border-box;">美好的梦想</strong>，在“忙”中不“盲”，不要忘记你曾经是怎么样的小孩，不要忘记你曾希望变成怎样的大人。而大S跟大家的30岁之约，也又近了一年，不知道有多少人还记得这个约定呢，希望我们可以一起努力朝这个目标迈进。</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: right;margin: 20px 0% 10px;transform: translate3d(-20px, 0px, 0px);-webkit-transform: translate3d(-20px, 0px, 0px);-moz-transform: translate3d(-20px, 0px, 0px);-o-transform: translate3d(-20px, 0px, 0px);box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.75" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8KBnhbUDnNbQwTDOiaZMhfRP7wa7b82ELkTUCnR99CXAFVSQibqfDU4YA/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section></section></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px none rgb(160, 160, 160);border-bottom-left-radius: 0px;padding: 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;border-right: 1px solid rgb(71, 193, 168);border-top-right-radius: 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: transparent rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">顺路给大家推荐一本书“牧羊少年的奇幻之旅”（The Alchemist），这本书是巴西作家Paulo Coelho的作品，2015年还曾经被改编成电影。这本寓言式小说运用富含哲理和诗意的语言讲述了牧羊少年圣地亚哥追寻宝藏的奇幻冒险故事，其中的很多小故事都可以映射到我们的活动中，非常多的小细节都值得品味。最后一碗浓鸡汤，嗯</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: right;margin: 20px 0% 10px;transform: translate3d(-20px, 0px, 0px);-webkit-transform: translate3d(-20px, 0px, 0px);-moz-transform: translate3d(-20px, 0px, 0px);-o-transform: translate3d(-20px, 0px, 0px);box-sizing: border-box;"><section class="" style="display: inline-block;width: 90%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(254, 255, 255);box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="transform: translate3d(5px, 0px, 0px);-webkit-transform: translate3d(5px, 0px, 0px);-moz-transform: translate3d(5px, 0px, 0px);-o-transform: translate3d(5px, 0px, 0px);margin: -5px 0% 5px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;box-shadow: rgba(81, 81, 81, 0.54) 0px 0px 3px;background-color: rgb(255, 255, 255);letter-spacing: 0px;padding: 8px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.5625" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNotpySZ2MZcHtNic7ehMNyp8byRNovaeAKNrGjOe4XEwa3353PVPHVWdP7OkR92peQfjdzPC4dZAmw/640?wx_fmt=jpeg" style="vertical-align: middle;max-width: 100%;box-sizing: border-box;" data-type="jpeg"></section></section></section></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 10%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="1.0147059" data-w="68" src="https://mmbiz.qpic.cn/mmbiz_png/9LfNnibXYmxqaPnOTmCZ9vibWFA9elPibic8hEhuI62pMDmoaiaiaKQibtEGCXE5mb1mu58waOys6sRL5exQTtFNib7yWg/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-left: 1px solid rgb(71, 193, 168);border-bottom-left-radius: 0px;padding: 0px 0px 0px 10px;border-top: 0px solid rgb(62, 62, 62);border-top-left-radius: 0px;box-shadow: rgb(0, 0, 0) 0px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="display: inline-block;width: 100%;vertical-align: top;border-style: solid;border-width: 1px 0px 0px;border-radius: 0px;border-color: rgb(71, 193, 168) rgb(62, 62, 62) rgb(62, 62, 62);padding: 10px 0px 0px;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">一直都觉得很幸福，因为自己拥有<strong style="box-sizing: border-box;">可爱的家人朋友</strong>。</p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="margin: 10px 0% 0px;box-sizing: border-box;"><section class="" style="text-align: justify;font-size: 14px;color: rgb(35, 34, 34);line-height: 1.8;box-sizing: border-box;"><p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">每次面临选择的时候，爸妈总是跟我说，你开心最重要，想选什么就什么，我们在你背后支持你。</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">每次心情不好的时候，朋友们总是陪在我身边，虽然有时候他们觉得那些小事儿没啥大不了，但是还是耐着性子听我念叨完…有时候在电话里难过得大哭一场，也会让他们跨越整个北京来陪我，跟我说“有我们在，再难也要熬过去”。</p>
<p style="white-space: normal;margin: 0px;padding: 0px;box-sizing: border-box;">在一段旅行的最后，一个天天嬉皮笑脸，打打闹闹的朋友，突然一脸真诚的对你说“你真的已经很棒了！不要有压力，要开心的做自己。”那一刻的感动好像无法用语言形容。于是在他说他pick菊姐是因为有些地方让他想到了我的时候我也只能保持微笑了。</p></section></section></section></section></section></section></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;width: 60%;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.159375" data-w="640" src="https://mmbiz.qpic.cn/mmbiz_png/PlhySLk3mOGTKbibTBRr39rOiakvBRn90Ka8ssKkMaaeicq1KJoMcxm1ED48V9kva4YHRr1k3vZdvr52002L5OuvQ/640?wx_fmt=png" style="vertical-align: middle;max-width: 100%;width: 100%;box-sizing: border-box;" width="100%" data-type="png"></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><span style="font-size: 14px;">谢谢大家陪伴着大小S一起成长。愿亲爱的你们，拥有健康的身体，强大的内心，有趣的灵魂，美好的梦想，和可爱的家人朋友。下次见啦～</span></p></section></section></section><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="box-sizing: border-box;"><section class="" style="box-sizing: border-box;"><p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br style="box-sizing: border-box;"></p>
<p><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080" style=""></p>
<p style="margin: 0px;padding: 0px;box-sizing: border-box;"><br></p>
<p><img class="" data-copyright="0" data-ratio="0.4514877102199224" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNpKic4DiblDA75KduXD0EeqoZ8McBkcR5DUfib3SQ5IiaNYpzLW98Valvic1XIQmFEIbibgFwZUPNb69vcg/640?wx_fmt=png" data-type="png" data-w="773" style=""></p></section></section></section></section>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693232&amp;idx=1&amp;sn=9c4e7a96ac7fd09c902ec97c66f6d286</link>
      <pubDate>Mon, 18 Jun 2018 05:37:34 -0700</pubDate>
      <guid isPermaLink="false">894</guid>
    </item>
    <item>
      <title>干货 | 文本改写和风格迁移的一些思路</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p style="text-align: left;"><span style="font-size: 15px;">今天想分享几篇比较新的文本改写（Rewrite）和文本风格迁移（Style Transfer）的工作。会涵盖的论文有：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p style="text-align: left;"><span style="font-size: 14px;"><strong>Unpaired Sentiment-to-Sentiment Translation: A Cycled Reinforcement Learning Approach</strong>. ACL 2018.</span></p></li>
<li><p style="text-align: left;"><span style="font-size: 14px;"><strong>Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer</strong>. NAACL 2018.</span></p></li>
<li><p style="text-align: left;"><span style="font-size: 14px;"><strong>Generating Sentences by Editing Prototypes</strong>. TACL 2018.</span></p></li>
</ol></blockquote>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0578125" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Zv4WU8vayrn26QiaPYYkRiaZlrsxFXeEYDgP2QptHy0O2Rp4VSicg8UT4w/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p style="white-space: normal;"><span style="font-size: 15px;">第一篇工作 [1] 从题目来看就很吸引人，unpaired 对应 cycle，很自然联想到 reconstruction。而 reinforcement learning 主要是为了解决训练过程中离散文本不可求导的问题。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"><br></span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">虽然过去也有一些 text style transfer 的问题，但是因为缺少对于非 stylish 或者说“普通文本”的语义信息的显式保留，导致很容易出现下面这种情况：当把“The food is delicious”这句话从正向情感迁移到负向情感时，会得到“What a bad movie”——虽然情感被正确转化了，但是内容的主体也跟着变了。这显然是不好的。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"><br></span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">为了解决这个问题，这篇工作 [1] 使用了两个 module，一个 neutralization module 用于提取 non-emotional content，另一个 emotionalization module 用于生成或者说融入 sentiment words。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="1.1518691588785046" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Z1IWIVHPTsY1KN91b3pROG2Fnv1VngicJwH9F3wXESmQCl7LolhFb9Bw/640?wx_fmt=png" data-type="png" data-w="428" style="width: 384px;height: 442px;"></p>
<p><span style="font-size: 15px;">如上图所示，下面的 neutralization module 其实是通过<span style="text-decoration: underline;">直接删除</span> emotional words 来实现 non-emotional content 的保留。这一步中，作者 [1] 通过使用 self-attention 得到的注意力权重作为分类器特征，从而自动找出 emotional words。从实验结果来看，这个方法的分类准确率也可以达到 90% 左右（不过 sentiment classification 本身也不是很难，稍微复杂一点的模型都可以达到95%甚至更高）。接下来就是 emotionalization module，可以看到，一个 encoder 和两个 decoder。两个 decoder 对应 sentiment classification 的两个类别（positive or negative），并且分别用于训练时的 reconstruction 和测试时的 generation（emotionalization）。</span></p>
<p style="white-space: normal;"><br></p>
<p style="white-space: normal;"><span style="font-size: 15px;">由于文本的离散性，作者采用了 RL 的方法来训练这两个 module。RL 成功与否严重依赖于奖励函数的设计。在这篇论文中 [1]，作者提出了两个 reward：sentiment confidence 和 BLEU。尤其是 BLEU，主要是针对作者想解决的问题——non-emotional content perserving。最终的奖励函数如下：</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.2662192393736018" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZfgWhrGoTmpGWliaDhuiaVdgnLvoAWyWeM3wvzYqvelTfUIUKTEt4ribiaw/640?wx_fmt=png" data-type="png" data-w="447" style="width: 322px;height: 86px;"></p>
<p style="text-align: left;"><span style="font-size: 15px;">从实验结果来看，过去的工作主要都只考虑了 style transfer 的成功与否（sentiment accuracy），而没有显性地考虑 content perserving，所以确实在 BLEU 指标下表现很差。</span><br></p>
<p style="text-align:center;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZOz8RUOyhHF4A5EfT548lHZExzTuHHaxcJuT08MrGCYvz68c0WqLVSQ/0?wx_fmt=png" data-cropx1="0" data-cropx2="459" data-cropy1="0" data-cropy2="191" data-ratio="0.41612200435729846" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZcoPDZ0RMPXskSdAsI0icfpBkWaS1MsgBFJ07D8WdccPrHb5ticaHeVNw/640?wx_fmt=jpeg" data-type="jpeg" data-w="459" style="width: 373px;height: 155px;"></p>
<p style="text-align: left;"><span style="font-size: 15px;">但是也可以看到，这篇工作提出的方法 [1] 在大幅提高 BLEU 的情况下，ACC 也有一定损失。可以从这个结果看到，</span><span style="font-size: 15px;text-decoration: underline;">想要非常自然地融入 sentiment words 并且不破坏语义和语法，是很困难的事情</span><span style="font-size: 15px;">。看一些直观的转换结果：</span></p>
<p style="text-align:center;"><img class="" data-ratio="1.0369515011547343" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Z59HZ4OwfImiao3NMaHQCcoXBsjGzCzzDAY3GphmPQ98yJB7Wq0QupGQ/640?wx_fmt=png" data-type="png" data-w="433" style="width: 377px;height: 391px;"></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0578125" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Zv4WU8vayrn26QiaPYYkRiaZlrsxFXeEYDgP2QptHy0O2Rp4VSicg8UT4w/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p style="white-space: normal;"><span style="font-size: 15px;">要分享的第二篇论文 [2] 几乎也是一样的思路。但是在具体实现上要更直接一点。第二篇论文 [2] 在进行 style transfer 的时候，基于的是这样的假设：文本的一些属性（attribute），比如文本传达出来的情感信息等，可以体现在文本中的某些特定词汇上。通过改变这些词汇，就可能直接改变整句文本的性质属性（value）。</span><br></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<p style="white-space: normal;text-align: center;"><img class="" data-ratio="1.564516129032258" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKntyYJBUYibdY5JbJZVBlzfmiaOB9qqf3JMa3DygFYhq94qYpMvsjtwcWg/640?wx_fmt=png" data-type="png" data-w="310"></p>
<p style="white-space: normal;"><span style="font-size: 15px;">由上面的示意图可以看出，作者将那些具有指示功能的特定词称为 attribute marker，通过（主要）改变这些 marker，进行 attribute transfer。文章进一步提出了4种具体的方法来改变 marker。</span></p>
<p style="white-space: normal;text-align: center;"><img class="" data-ratio="0.4241935483870968" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnoLp2NaXNdpELibvYCiaIuZVbtNAKOBqHicAia1XnTh2x7Jyps27nt424gQ/640?wx_fmt=png" data-type="png" data-w="620"></p>
<p style="white-space: normal;"><span style="font-size: 15px;">这4种具体的方式也由上图很清晰地展示了出来。值得一提的是 DeleteOnly 和 DeleteAndRetrieve。在删除了具体的 marker 后，作者提出再去根据这个没有情感特定性的句子去从数据中检索出一句最相似但是 attribute 的数值相反（target attribute）的话。检索出的这句话主要用于</span><span style="font-size: 15px;text-decoration: underline;">提取新的 target attribute marker</span><span style="font-size: 15px;">，作为“重写” RNN 的输入之一。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"><br></span></p>
<p style="white-space: normal;"><span style="font-size: 15px;">很明显，这样的工作可能有一定局限性，但在一些简化场景中，却是更可控的。其实验结果也印证了这点 [2]。</span></p>
<p style="white-space: normal;"><span style="font-size: 15px;"></span></p>
<p style="white-space: normal;text-align: center;"><img class="" data-ratio="0.7072" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnVmibed5J4FeUqRBiaWTjTtwuLAaBWRz4vNM2V2YKwGhevibMib38tyuJKw/640?wx_fmt=png" data-type="png" data-w="625"></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0578125" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Zv4WU8vayrn26QiaPYYkRiaZlrsxFXeEYDgP2QptHy0O2Rp4VSicg8UT4w/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">上面两篇工作都是进行风格转换，而在进行文本改写的时候，还有一类常见的场景就是进行非情感类改写，比如扩写句子，改写语法错误等。在进行这样的工作时，其实也可以采用类似的思路。</span><span style="font-size: 15px;text-decoration: underline;">先找到一个“中立”的、“简单”的“模板”，再在这个基础上加入希望增加的信息，如情感、一些复杂修饰词等等</span><span style="font-size: 15px;">。今天要分享的最后一篇工作 [3] 就是这样一个思路。</span></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.9395248380129589" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZkB3GxduPiatiau1xlTzBTGosNZl77HmjWFNsSxD2IvZ03xHSmQWymkoA/640?wx_fmt=png" data-type="png" data-w="463" style="width: 409px;height: 384px;"></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">从上面的示意图中可以看到，修改后的句子增加了“modiocre”和“ridiculous”这两个形容词，变得更复杂了。而重中之重就是在学习 edit vector 上。可以看到，此时的 edit vector 不再是 sentiment words，而是一种 semantic operation。如果让 edit vector 作为一种隐变量，也遵循某种分布，那么同样的 edit vector 应该符合同一种 edit operation，并且对于句子的改写是一种微小的可控的操作。</span></p>
<p style="white-space: normal;text-align: left;"><br></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">基于这样的假设和期望，作者用 VAE 来建模 z，目标函数如下：</span></p>
<p><img class="" data-ratio="0.2673913043478261" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZN9lgIJIafVnTeUhNo3fiaFco9pQGM4DqpQhLCHBVscEg4wGj4HUTb9g/640?wx_fmt=png" data-type="png" data-w="460" style="width: 359px;height: 96px;"></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">关键就是如何保证学到好的 p 和 q。这篇论文使用的方式和过去很多 VAE 的文本应用都不太一样，有兴趣的同学请一定去查阅原文。</span><span style="font-size: 15px;text-decoration: underline;">值得一提的是，这篇工作中的 edit vector z 是直接拼接在给 decoder 的每一个时刻的输入上的</span><span style="font-size: 15px;">，并没有额外的 gate 或者 transform。这也是和上面两篇工作的区别之一。</span></p>
<p style="white-space: normal;text-align: left;"><br></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">最后从结果来看，作者提出的方法确实能更自然地改写句子。这点从 case studies 还有句子的平滑度方面可以看出：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.9108695652173913" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZpmNCqkHYibTlFB3WSiaVibibV1E9q2a92zgtymiaxkV7NDibsdNMUhEzxV5Q/640?wx_fmt=png" data-type="png" data-w="460"></p>
<p style="text-align:center;"><img class="" data-ratio="0.23229166666666667" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5ZYoqelGU8kPpoW8umpnduAH7oCia2zzocrFjOnADnkowzy8O1gLDiaj0Q/640?wx_fmt=png" data-type="png" data-w="960"></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0578125" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqESB7eicUDxnawrZMuzyN5Zv4WU8vayrn26QiaPYYkRiaZlrsxFXeEYDgP2QptHy0O2Rp4VSicg8UT4w/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p style="white-space: normal;text-align: left;"><br></p>
<p style="white-space: normal;text-align: left;"><span style="font-size: 15px;">更多相关阅读：<br></span></p>
<blockquote>
<p style="white-space: normal;text-align: left;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6&amp;chksm=81233b7fb654b269794e8246aa3c1512421a406625438a864699337a6c3da7f94afa4497fd22&amp;scene=21#wechat_redirect" target="_blank" style="font-size: 15px;text-decoration: underline;"><span style="font-size: 15px;">干货 | 限定和指导 Seq2Seq 模型的生成结果（一）</span></a></p>
<p style="white-space: normal;text-align: left;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693201&amp;idx=1&amp;sn=e385650379ede59df28ae504f453dc9a&amp;chksm=81233b68b654b27ee2f6c4833493e7aec053c6b14bb6afa8c1f2920df9a933e53ce95db4430a&amp;scene=21#wechat_redirect" target="_blank" style="font-size: 15px;text-decoration: underline;"><span style="font-size: 15px;">干货 | 限定和指导 Seq2Seq 模型的生成结果（二）</span></a></p>
<p style="white-space: normal;text-align: left;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692981&amp;idx=1&amp;sn=37ed31cc6ec1ded76a53a26503ba4f93&amp;chksm=8123384cb654b15a6582beb3efb3bae1f350a12077456aded80ddb3d1993be4354bde3fc16a3&amp;scene=21#wechat_redirect" target="_blank" style="font-size: 15px;text-decoration: underline;"><span style="font-size: 15px;">干货 | 你的 KL 散度 vanish 了吗？</span></a></p>
<p style="white-space: normal;text-align: left;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692433&amp;idx=1&amp;sn=b880134068f670cfa59a81b59d432045&amp;chksm=81230668b6548f7eb5bf73a696b5e918d0486e96dca8df55740a947df827d2263e951bcb2136&amp;scene=21#wechat_redirect" target="_blank" style="font-size: 15px;text-decoration: underline;"><span style="font-size: 15px;">ACL 2017 | VAE in Dialog Response Generation</span></a></p>
</blockquote>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693224&amp;idx=1&amp;sn=96e1e8b74bd9cbda4ca3896a6a95befd</link>
      <pubDate>Wed, 13 Jun 2018 02:00:35 -0700</pubDate>
      <guid isPermaLink="false">893</guid>
    </item>
    <item>
      <title>干货 | 改进注意力机制：分解 LSTM 隐状态</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p style="text-align: left;"><span style="font-size: 15px;">注意力机制想必大家都已经很熟悉了。虽然注意力机制最初是被用在机器翻译上，但随后其在各种任务上都占有了一席之地，比如 multi-modal 相关的 image caption generation, video description，以及现在大热的机器阅读理解（machine reading comprehension）。</span></p>
<p style="text-align: left;"><br></p>
<p style="text-align: left;"><span style="font-size: 15px;">然而，注意力机制也有天然的弱点。最初提出的注意力机制，主要是利用当前状态解码器的状态 h_t 和过去多个时刻的 h_{t-i} 比较的结果，来计算其对应的权重——最终将这些信息和当前状态的信息一起综合利用，去判断当前时刻的输出。这样的机制归根结底还是将信息整合到一个向量中，不可避免会造成信息损失和弥散。</span></p>
<p style="text-align: left;"><br></p>
<p style="text-align: left;"><span style="font-size: 15px;">针对这个问题，研究者们也提出了不少解决办法。今天就想和大家分享其中几篇代表作。今天会涵盖的论文有：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><span style="font-size: 14px;"><strong>Frustratingly Short Attention Spans in Neural Language Modeling</strong>. ICLR 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Recurrent Attention Network on Memory for Aspect Sentiment Analysis</strong>. EMNLP 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Dialog Generation Using Multi-turn Reasoning Neural Networks</strong>. NAACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Natural Answer Generation with Heterogeneous Memory</strong>. NAACL 2018.</span></p></li>
</ol></blockquote>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">首先要介绍的工作 [1] 可以说是第一篇专注分析 NLM+attention 的工作。作者首先将</span><span style="font-size: 15px;text-decoration: underline;">注意力机制拆分为三部分</span><span style="font-size: 15px;">：（1）keys: 对比当前时刻的状态和过去时刻的状态；（2）values: 比较结果用于建模上下文；（3）prediction: 综合当前时刻和上下文的信息来进行预测。然而作者认为，现有的方式并不能很好地实现这三件事。所以作者 [1] 提出拆分解码器的状态。</span></p>
<p><img class="" data-ratio="0.4096989966555184" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdMq7tdibd7RcWUZmrQ2KiciaImYE1yzCZZib7gYFQc9L4Laliby01Q6iaVShg/640?wx_fmt=png" data-type="png" data-w="598"><br></p>
<p><span style="font-size: 15px;">对比上图中的 (a) 和 (b)，可以看到，作者将解码器状态 h_i 拆分为 k_i 和 v_i，并用 k_i 比较每一个时刻的解码器状态，用得到的 v_i 建模上下文和预测下一个词。这就是作者提出的第一个改进模型，key-value attention model。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">显然，这个模型中，v_i 仍然</span><span style="font-size: 15px;text-decoration: underline;">一人分饰两角</span><span style="font-size: 15px;">。所以作者 [1] 进一步提出了第二个改进模型，key-value-predict attention model：将解码器状态拆分为三个向量，其中新引入的 p_i 用于做预测。这个模型示意图请参考下方左侧 (c)。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4833887043189369" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdxR1RlaAbCibZmFpRWd76bB3JrtMEWMABolYElMk4VUp9OhoIrPso5gg/640?wx_fmt=png" data-type="png" data-w="602"></p>
<p><span style="font-size: 15px;">在实验中，作者提出的两个改进模型</span><span style="font-size: 15px;text-decoration: underline;">在大部分数字指标下</span><span style="font-size: 15px;">的表现，都超过了传统注意力机制模型，并且符合预期地：key-value-predict attention &gt; key-value attention &gt; traditional attention。然而，从</span><span style="font-size: 15px;text-decoration: underline;">可视化的结果</span><span style="font-size: 15px;">来看，这些模型的差异却并没有那么大：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.47019867549668876" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdbsiaOAZia03zeuZHBAIJFoNVAAcODetgOpeKQZfaSK4n8Rib2Cfuhn0Cw/640?wx_fmt=png" data-type="png" data-w="604"></p>
<p><span style="font-size: 15px;">作者发现，这些基于注意力机制的模型的注意力基本都局限在最近的5个状态的窗口里，也就是题目中所谓的 short span。这也直接导致了，当作者用基于 N-gram 的 RNN 代替基于注意力机制的 RNN 时，也取得了非常惊人的表现。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">作者也通过改变注意力窗口大小，再次印证了，注意力机制的局限性。当窗口增大到一定程度，也就是注意力范围扩大到一定程度后，实验结果并不会继续提高了。换句话说，</span><span style="font-size: 15px;text-decoration: underline;">注意力机制对于捕捉 long-term 范围内的信息的帮助很小</span><span style="font-size: 15px;">。实验结果可以见：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.475" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdDqANuM1yL4kvqndKUK01b90vMbTMhfBCv5C9Am2XSlodjyLoCPqepw/640?wx_fmt=png" data-type="png" data-w="360"></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;"></span><span style="font-size: 15px;">注意力机制的不足在某些任务上会显得尤为严重，尤其是那些对于句子的细粒度语义需要仔细把握的任务。而 Aspect Sentiment Analysis，也叫 Fine-grained Sentiment Analysis， 就是这样一个代表性任务。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">Aspect Sentiment Analysis 的任务是真根据某一个 product/item 的某一个 aspect 来进行情感分类（negative，neural，positive）。常见的例子是电子产品，比如手机。关于手机的一个商品评价可能是从手机的多个方面来撰写的：照相效果很不错，但是电池使用时间较短。显然，这样的任务对于</span><span style="font-size: 15px;text-decoration: underline;">建模上下文语义</span><span style="font-size: 15px;">（尤其是一些复杂句型的语义）的要求非常高。而仅仅依靠简单的注意力机制，或者说单一的注意力机制，就很难做到这点。举一个例子来说，"Except Patrick, all other actors don't play well" 这样的句子，就需要综合“except”和“don't play well”这两个信息，从而得出针对“Patrick”的情感是正向的。<br></span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">为了具有更复杂的综合信息的能力，今天要分享的第二篇论文 [2] 首先将解码器状态拆分和整合成特殊的</span><span style="font-size: 15px;text-decoration: underline;">记忆单元</span><span style="font-size: 15px;">（memory），然后<span style="font-size: 15px;">提出了一种综合多种注意力机制的</span><span style="text-decoration: underline;font-size: 15px;">非线性</span><span style="font-size: 15px;">方法。</span></span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.5818713450292398" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdZ135PDvstvXXHaE3xJh4UCibZib1OxVH3wDppjdjssniaqm2JYtyur5Hg/640?wx_fmt=png" data-type="png" data-w="684"></p>
<p><span style="font-size: 15px;">由上面的模型图可以看到，LSTM 的状态被分解成了 memory cell，并且进一步整合成了 location weighted memory（左上角）。这是因为，对于 Aspect Sentiment Analysis 来说，一句话中可能有多个 aspect。而代表 aspect 的 target word 在一句话中的位置信息有助于建模 target-based context。换句话说，memory 对于每一个 target 是一样的；</span><span style="font-size: 15px;text-decoration: underline;">而 location weighted memory 对于每一个 target 就不一样了</span><span style="font-size: 15px;">。在这篇论文 [2] 中， location weighting 的计算是启发式的（具体见论文公式7）：越靠近 target 的 sentiment words 的权重或者说影响力越大。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">有了这些基础后，就到了这篇论文 [2] 的重头戏，recurrent (multiple) attention on memory。从名字来看就知道，</span><span style="font-size: 15px;text-decoration: underline;">首先 multiple attentions 用于抽取一句话中可能存在的多个片段的有用信息；而 recurrent 用于整合这些信息</span><span style="font-size: 15px;">。更具体地，作者用 i_t 来代表每一个时刻从 memory 中提取到的（attended）有用信息，而每一次整合的结果存储在 e_t 中，不仅影响下一个时刻的 e_{t+1} 也影响着下一个时刻的 i_t。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">这样的注意力机制不仅弥补了单次单个的注意力机制的不足，还增加了非线性带来的高层语义的捕捉（只是希望）。在实验结果上也有一定的印证：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.5484351713859911" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdQSGefGBYs1RAYCOaGiaKQonNFiauFxcA41CeXGIBMJpfZl1JKHlZoHgw/640?wx_fmt=png" data-type="png" data-w="671"></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="text-align: center;vertical-align: middle;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"></section></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"></section><br></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">上面这两篇的任务中，主要还都是只有单源头的信息。接下来要分享的一篇工作 [3] 改进的是对话系统中的注意力机制。因为对话数据一般是两个说话者的，所以信息是双源头的，这也为注意力机制带来了新的困难。</span></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">和上一篇工作比较相似的地方在于，这篇工作 [3] 也是将各种上下文信息拆分到多个记忆单元（memory）中去。只不过这一次，因为是对话的上下文，或者说语境信息，所以信息的种类比较多。具体地，作者将对话双方的对话历史存在了不同的记忆单元，并且将每个人的对话历史又进一步拆分成了两个粒度的记忆——句子级别的，对话级别的（有时候也会叫 session-level，或者说 conversation-level）。与此同时，因为当前最邻近的一句输入（query）是最重要的，作者提出针对这个特殊 query 建立独立的记忆单元。</span></section><p><img class="" data-ratio="0.8851351351351351" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEduicv7VWvTY9Dzm2eGgcIy8qHADhhSNiaOiaibxKjtdsibomHiaic9f8czmtTg/640?wx_fmt=png" data-type="png" data-w="444" style=""></p>
<section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">从上图可以看到，共有四大种（四大列）记忆单元，Mh, Mc, Mq, Mr。其中 Mh 是句子级别的记忆单元，Mc 是对话级别的，Mq 和 Mr 是针对 query 的（对话双方交替建模）。有了这四大种记忆单元，作者分别（separately）对它们使用了注意力机制（multiple attentions），<span style="text-decoration: underline;">然后将四个 attention vectors 拼接在一起，用于最终的预测</span>。作者将这种模型称为 MULTI。</span></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">基于 MULTI，作者进一步提出了 multi-turn attention reasoning。其主要启发自机器阅读理解中的 multi-hop reasoning。关于这部分的细节请大家查阅原始论文 [3]。</span></section><section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><section class="" style="text-align: left;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><span style="font-size: 15px;">今天最后想分享的论文 [4] 也是希望整合多步注意力机制的结果，但是用了一种比较新颖的、<span style="text-decoration: underline;">和 multi-hop 不同的</span>，整合方法。multi-hop attention 的方法主要是对于同一个、单一的 context vector（可以通过各种方式得到），多次使用注意力机制（multi-turn）。而这篇工作 [4] 则是，对于多个 memory cell 多次使用注意力机制，从而最终得到针对 memory 的经过注意力整合后的 context vector。也就是说，这次的 LSTM 隐状态的拆分，首先表示在数据语言中隐藏的普通语义和外部知识库的拆分；其次针对外部知识库的隐状态，又使用了注意力机制进行知识整合。这两大类的隐状态最终拼接组合成解码器的状态，进行预测。有一些绕，还是请大家看示意图：</span></section><p style="text-align:center;"><img class="" data-ratio="0.36268068331143233" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEd3JIuMxq1Y3RK1mW5ibR11CrHicKO0rNveg3GJQgc3icDg2G5txfPqBaDg/640?wx_fmt=png" data-type="png" data-w="761"></p></section></section><section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.0765625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNodSM1pLdj1F4iaNrfwHibaEdIlSvywSZ3QKNYIzRIw6X1niaYibyvR850m9CzqnaA7upSFQm8MZYa9nw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">以上就是今天想分享的内容。从今天的内容可以发现，过分依赖注意力机制是很危险的。注意力机制的使用最好配合着特殊的信息拆分和表达。如果说曾经一时无两的神经网络大法使得大家都将各种信息一股脑丢进去而等待结果，那么现在越来越多的研究反而开始将信息重新拆分出来，各司其职也许效果更好。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">相关文章推荐阅读：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=212457976&amp;idx=1&amp;sn=ab503fb91c6a126fbe767f0dd37fb48a&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">干货 | Applied Attention-based Models in NLP（一）</span></a></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=212530824&amp;idx=1&amp;sn=55ca55ddb7ce7d2cd9b316c598b6da03&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">干货 | Applied Attention-based Models in NLP（二）</span></a></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692200&amp;idx=1&amp;sn=5211ddee37bc3a3cad0da93641a7081f&amp;chksm=81230751b6548e476ec2e2c91d023dfd1c7213e34dd085fa57a3f2aa4bd3253920a57aeb5281&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">ICLR 2017 | Attention 和 Memory Networks</span></a></p></li>
<li><p><span style="font-size: 15px;"><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=401930842&amp;idx=1&amp;sn=0f23752330fe37166a33dabb8e48e143&amp;scene=21#wechat_redirect" target="_blank">干货 | NIPS 2015 Reasoning, Attention, Memory Workshop</a></span></p></li>
<li><p><a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=400074084&amp;idx=1&amp;sn=598f0b83677a276ec4ba652a0abfff97&amp;scene=21#wechat_redirect" target="_blank"><span style="font-size: 15px;">干货 | 关于 Attention in Cognitive NeuroScience 你需要知道的7件事</span></a><br></p></li>
</ol></blockquote>
<p><br></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693212&amp;idx=1&amp;sn=70b63181946b17eda80f8f35b9dc0b16</link>
      <pubDate>Mon, 11 Jun 2018 23:30:45 -0700</pubDate>
      <guid isPermaLink="false">892</guid>
    </item>
    <item>
      <title>干货 | 限定和指导 Seq2Seq 模型的生成结果（二）</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p style="white-space: normal;"><span style="font-size: 15px;">今天继续关于 Lexically Contrained Decoding 的一些工作。<a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6&amp;chksm=81233b7fb654b269794e8246aa3c1512421a406625438a864699337a6c3da7f94afa4497fd22&amp;scene=21#wechat_redirect" target="_blank">上半部分的内容请点此查看</a>。会涉及的论文有：</span></p>
<blockquote style="white-space: normal;"><ol class=" list-paddingleft-2" style="">
<li><p><span style="font-size: 14px;"><strong>Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation</strong>. NAACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Sequence to Backward and Forward Sequences: A Content-Introducing Approach to Generative Short-Text Conversation</strong>. COLING 2016.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Towards Implicit Content-Introducing for Generative Short-Text Conversation Systems</strong>. EMNLP 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Neural Response Generation with Dynamic Vocabularies</strong>. AAAI 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong style="font-size: 15px;">Globally Coherent Text Generation with Neural Checklist Models</strong>. EMNLP 2016.</span></p></li>
<li><p><strong><span style="font-size: 14px;">Learning symmetric collaborative dialogue agents with dynamic knowledge graph embeddings</span></strong><span style="font-size: 14px;">. ACL 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Dynamic Entity Representations in Neural Language Models</strong>. EMNLP 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer</strong>. NAACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Retrieve, Rerank and Rewrite: Soft Template Based Neural Summarization</strong>. ACL 2018.</span></p></li>
</ol></blockquote>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><section style="white-space: normal;box-sizing: border-box;"><section class="Powered-by-XIUMI V5" powered-by="xiumi.us" style="box-sizing: border-box;"></section></section><p><span style="font-size: 15px;">昨天大部分论文是预测一个主题关键词，来指导和限定回复。很多时候，这样的关键词也可以从上下文中直接抽取。有一类代表词就是实体词，entity。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">关于 entity 的研究在近三四年的时间里获得了许多关注。entity 不仅是知识库的重要组成部分，也是理解很多特定体裁的文本的关键，比如小说和新闻中的各种人物关系。同时，entity 也在特定领域对话中扮演着重要角色。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">今天要分享的第一篇工作 [5] 则集中在一类更特别的文本生成——菜谱生成。菜谱生成的研究和 video description 还有 script knowledge 都有关系。而其最大特点也是，在一份菜谱中覆盖着多个 entity，还有对于这些 entity 的操作。也因此，这篇工作 [5] 设计了<span style="font-size: 15px;text-decoration: underline;">非常显式的记忆机制</span>，专门去建模 entity。</span></p>
<p style="text-align:center;"><img class="" data-ratio="1.0018552875695732" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnShU3SV5cHicdLH5wEt4n2CVCDBz4yPwtMz6S6WnpzNricWtcWZbt18Jw/640?wx_fmt=png" data-type="png" data-w="539" style="width: 400px;height: 401px;"></p>
<p><span style="font-size: 15px;">其主要思想 [5] 可以由上述示意图来表示。给定一个菜谱标题（菜名），和一系列 entity 也就是菜谱的“原材料”，这篇工作 [5] 使用序列模型逐步生成菜谱中的每一个 token。在每一个生成时刻，预测是否该生成一个 entity。如果是，则从原材料列表（agenda）中计算出概率最高的那个，作为此刻的输入。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">到此为止，这个思想还是类似于 CopyNet 或者说 pointer network 的。但是这篇工作 [5] 因为针对菜谱生成，将其进一步细化了。对于菜谱来说，有些 entity 被生成后，还会被再次提及。所以这篇工作 [5] 将每一时刻的门的判断，</span><span style="font-size: 15px;text-decoration: underline;">由两种选项，扩展成三种选项</span><span style="font-size: 15px;">——（1）普通 GRU 生成，也就是非 entity 限定；（2）生成一个新的，还没被提过的 entity；（3）生成一个旧的，已经被提过的 entity。自然而然地，一个新的 entity 一旦被生成，就会变成“旧”的 entity，也就需要记录这种“变化”——由此就有了 checklist 的思想。这篇文章 [5] 也因此将模型命名为 checklist model。</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">上面的 checklist model [5] 的适用范围不止是菜谱生成。如果把 agenda 中存储的 entity 从菜谱原材料变成一些 slot，就可以进行特定领域对话的生成。其根本思想是<span style="text-decoration: underline;">追踪、记录和更新</span>某些特定的某些实体/信息——即用特殊机制建模特定信息。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">从这个角度来说，曾经我们详细介绍过的另一篇工作 [6]，也可以被划分到这里来。</span></p>
<p style="text-align:center;"><span style="font-size: 15px;"><img data-ratio="0.3716651333946642" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNovfg8qxSeea4YOrXcmDVhXzriaYAtliclr0lzFvrx47MbthqebSSYia9hFYgeCYRJ3FjkEZiboOANwaw/640?wx_fmt=png" data-type="png" data-w="1087" style="text-align: center;white-space: normal;"></span></p>
<p><span style="font-size: 15px;">在这篇工作 [6] 中，对话中涉及的 item（朋友列表）是被追踪和更新的对象。这些对象是图中的结点，作者 [6] 希望获得每个结点的表达（vector representation for each node），也就是 Graph embedding。每当新的对话信息被加入，就需要更新这个对话信息中涉及到的相关结点的表达。比如上图中，当出现了“anyone went to Columbia”的时候，首先相关的结点肯定是 columbia。但和 columbia 相关的别的结点，jessica 和 josh 也应该被更新。<a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692766&amp;idx=1&amp;sn=6eace56369297127832978ef1065b437&amp;chksm=812338a7b654b1b1f53e48ab5d586adc7797500e5acad3cafa261891b303ec50dab875c7ef3a&amp;scene=21#wechat_redirect" target="_blank">关于这篇论文的具体解读可以点此查看</a>。</span></p>
<p style="text-align:center;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="text-align: center;vertical-align: middle;box-sizing: border-box;"></p>
<p><span style="font-size: 15px;">上面这篇论文 [6] 的核心思想就是利用动态表达（Dynamic Representation）来建模核心信息（item）。这类思想也在过去几年的关于 general entity 建模中被广泛使用。这方面的一个代表性工作就是由 [7] 提出的 EntityNLM，entity neural language model。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.8387096774193549" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnwH2flMZDN4mK78T579mePVricZskguuCHJ9AdcskQ9ONf53SS6hFIRw/640?wx_fmt=png" data-type="png" data-w="310"></p>
<p><span style="font-size: 15px;">可以看到，作者并不局限在某一个特定场景下的工作，而是使用 language model 建模含有大量 entity 的文本。换句话说，作者的工作使用的是生成模型，而且是<span style="text-decoration: underline;">含有隐变量的条件生成模型</span>。</span></p>
<p><img class="" data-ratio="0.2980251346499102" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn0aTT7MPuSg2LWpvwYG4KGefMRVia3XrG3JBR479mXunvwU6j8apfzFQ/640?wx_fmt=png" data-type="png" data-w="557" style=""></p>
<p><span style="font-size: 15px;">由上图看到，R，E，L 都是要学习的变量。这些变量主要用于辅助预测当下时刻要生成的是否是一个 entity。接下来就是如何去学习这些 entity 的表达。可以想到，entity 的表达受 context 和这些变量影响下的其他 entity 的影响。作者 [7] 也因此使用了 dynamic representation。并且，同上面两篇论文一样，每一词预测出一个 entity 后，这些 dynamic representation 也要更新。关于这两部分的细节公式，请大家查阅原始论文 [7]。</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">限定和指导 Seq2Seq 模型生成结果的另一个重要应用是风格迁移（Domain Adaptation），其潜在的场景就是领域迁移。最早的基于神经网络风格迁移多数是先预训练一个模型，再在新领域的数据上 fine-tuning。这样的过程虽然简单粗暴，但效果往往不尽如人意。尤其是在由离散数据组成的文本迁移上，由于很难通过连续改变一个像素点的数值来训练，导致训练结果更加不可控。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">这篇工作 [8] 则给出了一个新的可行思路。文本的一些属性（attribute），比如文本传达出来的情感信息等，可以体现在文本中的某些特定词汇上。通过改变这些词汇，就可能直接改变整句文本的性质属性（value）。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="1.564516129032258" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKntyYJBUYibdY5JbJZVBlzfmiaOB9qqf3JMa3DygFYhq94qYpMvsjtwcWg/640?wx_fmt=png" data-type="png" data-w="310"></p>
<p><span style="font-size: 15px;">上面的示意图给出了这篇工作 [8] 的思路。作者将那些具有指示功能的特定词称为 attribute marker，通过（主要）改变这些 marker，进行 attribute transfer。文章进一步提出了4种具体的方法来改变 marker。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4241935483870968" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnoLp2NaXNdpELibvYCiaIuZVbtNAKOBqHicAia1XnTh2x7Jyps27nt424gQ/640?wx_fmt=png" data-type="png" data-w="620"></p>
<p><span style="font-size: 15px;">这4种具体的方式也由上图很清晰地展示了出来。值得一提的是 DeleteOnly 和 DeleteAndRetrieve。在删除了具体的 marker 后，作者提出再去根据这个没有情感特定性的句子去从数据中检索出一句最相似但是 attribute 的数值相反（target attribute）的话。检索出的这句话主要用于</span><span style="font-size: 15px;text-decoration: underline;">提取新的 target attribute marker</span><span style="font-size: 15px;">，作为“重写” RNN 的输入之一。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">很明显，这样的工作可能有一定局限性，但在一些简化场景中，却是更可控的。其实验结果也印证了这点 [8]。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.7072" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnVmibed5J4FeUqRBiaWTjTtwuLAaBWRz4vNM2V2YKwGhevibMib38tyuJKw/640?wx_fmt=png" data-type="png" data-w="625"></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">最后来看的这篇论文 [9] 有着异曲同工的思想——也是将 Retrieval 和 Generation 结合在一起——却是用在自动摘要（Summarization）的任务上。自动文摘的研究可以大致分为模板式（template），抽取式（extractive）和生成式（abstractive）。三者各有利弊。这篇论文 [9] 也算是尝试将模板式和生成式结合起来的一种新思路。思想虽然简单，效果依然出乎意料地好。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.2159090909090909" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnNPWMJyK0wPTswsrtzVEQtaNaQ3VNFYFuoiciaibjYQRaRXTgyglSNZRUw/640?wx_fmt=png" data-type="png" data-w="616"></p>
<p><span style="font-size: 15px;">这篇工作 [9] 的思路如上示意图。给定一句话（first sentence in the news），先去检索出一些相似 news，将它们的摘要（headline）作为候选 summary candidates。然后遵照常用方式进行 rerank 找出最合适的那一句 headline <span style="text-decoration: underline;">作为“soft template”</span>。这个 soft template 结合原始的输入，一起传给改写 RNN，生成最终的摘要。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">对比 [8] 和 [9]，这两篇工作都通过 retrieve 的方式从数据库中找了最相似的数据来作为“soft template”。可以猜想，这一步也可以认为是改写出更自然的、更流畅的句子的关键。</span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.10625" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnV8TjDRaPiaibrcO4ia9mQUxxNCPvMRhHhpWjicUUJ56NYsBV6b1KngjZxw/640?wx_fmt=png" data-type="png" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">这次的两篇总结提到的论文远远覆盖不了 Lexically Constrained Decoding 的全部研究。从更广泛的角度来讲，大家非常熟悉的 CopyNet 或者说 pointer network、placeholder approach 都可以归为这个范畴。相关文章推荐大家阅读：</span></p>
<blockquote>
<p><span style="font-size: 15px;">1. <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6&amp;chksm=81233b7fb654b269794e8246aa3c1512421a406625438a864699337a6c3da7f94afa4497fd22&amp;scene=21#wechat_redirect" target="_blank">干货 | 限定和指导 Seq2Seq 模型的生成结果（一）</a></span></p>
<p><span style="font-size: 15px;">2. <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692766&amp;idx=1&amp;sn=6eace56369297127832978ef1065b437&amp;chksm=812338a7b654b1b1f53e48ab5d586adc7797500e5acad3cafa261891b303ec50dab875c7ef3a&amp;scene=21#wechat_redirect" target="_blank">ACL 2017 | 别再对抗了，通过对话合作找出共同的好朋友吧！</a></span></p>
<p><span style="font-size: 15px;">3. <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692152&amp;idx=1&amp;sn=1ccec5b1d743be5d8a79400e291a1b91&amp;scene=21#wechat_redirect" target="_blank">ACL 2016 | CopyNet 和 Pointer Softmax</a></span></p>
<p><span style="font-size: 15px;">4. <a href="http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA==&amp;mid=2652692579&amp;idx=1&amp;sn=ee3239d8594d75ec7623e0bcd186b959&amp;chksm=812339dab654b0cc40abaaacb0580deccb1160c34354cc93ac43785c67bd515534c40e341fa9&amp;scene=21#wechat_redirect" target="_blank">干货 | Pointer Network 与 CopyNet 之上有哪些新颖任务</a></span></p>
</blockquote>
<p><br></p>
<p style="text-align:center;"><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693201&amp;idx=1&amp;sn=e385650379ede59df28ae504f453dc9a</link>
      <pubDate>Sun, 10 Jun 2018 17:21:09 -0700</pubDate>
      <guid isPermaLink="false">891</guid>
    </item>
    <item>
      <title>干货 | 限定和指导 Seq2Seq 模型的生成结果（一）</title>
      <description><![CDATA[<div class="rich_media_content " id="js_content">
                    

                    

                    
                    
                    <p style="text-align:center;"><img class="" data-copyright="0" data-ratio="0.2594594594594595" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz/qrpuZA2scNr5XjZxbO8rQ5Qv1XQjSkxOW7IXQNs70qvdgicBPL2Q3WfLCwyicgoXhGnprW0WnmyxBkw4aDgkPA3g/640" data-type="png" data-w="740"></p>
<p><span style="font-size: 15px;">Seq2Seq 模型被广泛应用在 NLP 的任务中，包括最被人熟知的机器翻译（Neural Machine Translation，NMT）和越来越火的对话系统（Dialog）。今天我们主要关注 Seq2Seq 模型应用到序列生成的问题。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">序列生成的过程，就是解码器解码的过程。在过去几年的发展中，研究者们逐渐开始关注使用一些先验知识来辅助解码过程，如要求解码出来的序列必须包含某个词汇。这样的研究可以广泛划为“<strong>Lexically Contrained Decoding</strong>”。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">Lexically Contrained Decoding 可以认为是启发自基于统计的 phrase-based MT，但随后其思想被转化应用到许多相似任务中。在一些任务中，含有指定词汇的要求被放宽到了由指定词汇指导（implicitly guided）。本文会涵盖的相关论文，覆盖了 NMT, Dialog Response Generation, Style Transfer 和 Summarization 等。这些论文是：</span></p>
<blockquote><ol class=" list-paddingleft-2" style="list-style-type: decimal;">
<li><p><span style="font-size: 14px;"><strong>Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation</strong>. NAACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Sequence to Backward and Forward Sequences: A Content-Introducing Approach to Generative Short-Text Conversation</strong>. COLING 2016.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Towards Implicit Content-Introducing for Generative Short-Text Conversation Systems</strong>. EMNLP 2017.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Neural Response Generation with Dynamic Vocabularies</strong>. AAAI 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong style="font-size: 15px;white-space: normal;">Globally Coherent Text Generation with Neural Checklist Models</strong>. EMNLP 2016.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer</strong>. ACL 2018.</span></p></li>
<li><p><span style="font-size: 14px;"><strong>Retrieve, Rerank and Rewrite: Soft Template Based Neural Summarization</strong>. ACL 2018.</span></p></li>
</ol></blockquote>
<p><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;">在 NAACL 2018 上，相关学者 [1] 提出了 O(1) 复杂度的 lexically constrained decoding 算法，使得 batch decoding 等技巧也可以同时使用，大大提高了解码效率。[1] 的方法中，首先也保留了像常规 k-size beam search 一样的需要维护的一个 k*|V| 的 scores 矩阵去记录每一时刻的扩展当前 hypothesis 的词汇生成概率。也就是下图的中间部分。如果按照正常的 k-size beam search，下一个时刻保留的 hypothesis 就会是 scores 矩阵中的 top-k 的词了（也就是下图中打★的那些词）。但是，这样的算法无法保证解码到最后，含有要求的词汇，也就是不满足 Lexically Constrained Decoding。</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.6648822269807281" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn0RLaJKw0OzU3aXrY64JAAVeU2I38qIzwZLFmZ97ZvxkyKc58jK5tKg/640?wx_fmt=png" data-type="png" data-w="934"></p>
<p><span style="font-size: 15px;">为此，[1] 的工作提出将 beam 按照满足限定词汇的个数的多少来分组。上图中最左侧的不同形状的符号，就表示一个</span><span style="font-size: 15px;text-decoration: underline;">限定词汇（constraint）</span><span style="font-size: 15px;">，而空心则表示还没被满足，实心则表示已经生成过。这些词汇也可以中间的 scores 矩阵最下方找到。基于这样的分组，[1] 提出每一个解码时刻，除了要保留 top-k 以外，还要保留那些还没生成的但有很高概率的词汇去加入到下一个时刻的 hypothesis 中去。这些词也就是上图中的加了→的。又由于，本身有一个 beam bank 是记录已经满足了全部限定词汇的，可以认为这类 hypothesis 是已经 partically-completed 的。对此，[1] 提出保留的第三类 item 就是 single-best token，也就是图中◇的。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">有了以上</span><span style="font-size: 15px;text-decoration: underline;">三类 candidate items（★，→和◇）</span><span style="font-size: 15px;">，就剩下如何分配这些 item 到相应的 beam bank 了。这篇文章 [1] 提出了比较 ad-hoc 的办法去动态分配这些 item。因为整体只维护着一个 k-size beam，最终的算法效率还是 O(Nk)，其中 N 是解码的句子长度，k 是 beam size。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">因为这篇论文[1] 主要研究的是不限定个数的、不限定位置的 target-side lexically constrained decoding，可以看到其适用范围是非常广泛的。一个比较直观的展示结果的方式就是去看看解码出来的句子，是否能找到比较自然的、合理的囊括限定词汇的位置，从而不破坏整体句子的可读性。以下是作者的一些实验案例：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4772234273318872" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnic9j2NZ9phoH8vGFts51WQJ9LwZCGatJt0DWn0ick08pk957sicuZFIjA/640?wx_fmt=png" data-type="png" data-w="922"></p>
<p><span style="font-size: 15px;">这个展示结果也可以从侧面证明，BLEU scores 的提高并不仅仅是因为涵盖了限定词汇，而是正确运用、放置了这些词汇。</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><span style="font-size: 15px;"></span><br></p>
<p><span style="font-size: 15px;">在 NMT 中使用限定词汇可以帮助一些初级语言学习者。word-by-word 的翻译虽然有时候比较生硬呆板，但是对于语言学习者来说比较容易在翻译结果中学到新的词汇。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">接下来要介绍的几篇论文是要指导和限定对话回复生成的结果。因为对话中的你来我往的一对聊天（utterance pair）的语义关联不像 NMT 翻译任务中的那样显著和紧密，直接使用经典 Seq2Seq 模型常常会学出过于普适的万能回复（generic response）。缓解这个问题的办法有很多，其中一个办法是引导模型生成一些特定主题的词汇。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">在这个方向上，先来看 COLING 2016 的一篇工作 [2]。[2] 的工作提出的 Seq2BF 模型在生成回复时，采用</span><span style="font-size: 15px;text-decoration: underline;">逆向和正向结合的生成方式去引入指定词汇</span><span style="font-size: 15px;">。作者希望指定的词汇是对于某一句最相关的，而不是最常见的。所以在这一步，他们 [2] 用 PMI 相关指标来预测出某一句话的指定词汇。随后，为了让这个指定词汇能出现在回复中的“任意”位置（不仅仅是在开头），作者将生成过程分解成了两个子过程，从而提出了下面的模型：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.683944374209861" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnB9BAnupHmGTFjbfCOgv6AZU70mdgiaxNv5EKBzznYiboflueCjIj9cLg/640?wx_fmt=png" data-type="png" data-w="791"></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><br></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">上面的工作有一个比较明显的问题就是，他们要求生成的回复必须包含这个预测出来的指定词汇。首先这是一个过于强的假设，其次，当预测的词汇不是很相关时，会非常严重的影响生成的回复质量。</span></p>
<p><br></p>
<p><span style="font-size: 15px;">为了解决这个问题，[3] 提出了 implicitly content-introducing 的方法。一提到 implicit，比较容易想到的办法就是利用 gate mechanism。[3] 的工作基于 gate，提出了 hierarchical gated fusion unit (HGFU) ，让预测出来的关键词能</span><span style="font-size: 15px;text-decoration: underline;">指导</span><span style="font-size: 15px;">回复的生成，而不是限定。具体地，[3] 提出 fusion 需要两步，local information initialization 和 global information inception。local 的部分就是用于初始化解码状态，而 global 的部分则是重中之重。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="1.2731277533039647" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKndrdtvNHEB86AWtsJTzj1yRAyfPmGGJeTFnu3iclkZN4Mn0neMicn0k1g/640?wx_fmt=png" data-type="png" data-w="454" style="width: 317px;height: 404px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">上图就是 [3] 提出的用于 global 的 HGFU。其创新点在于引入了一个额外的 cue word GRU，用于独立地捕捉指导词汇的信息，而不至于让 cue word information 弥散在 decoder 的状态中。最后，一个 high-level 的 fusion unit 用于整合两个底层 GRU 的信息，从而更综合地生成回复。</span><span style="font-size: 15px;">如果把 fusion unit 中的 gate 成为 k，k 的开关就决定了信息到底是来自于两个底层 GRU 中的哪一个。</span></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;"><span style="font-size: 15px;">虽然看起来方法比较简单，但效果还是不错的。可以先来看看 k 的可视化：</span></span></p>
<p><span style="font-size: 15px;"><span style="font-size: 15px;"></span></span></p>
<p style="text-align:center;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnhicFFz7QobMza7vtnBORXd5Gia1EVOTTbhib67UWOr0IGGaxnHXNmZ8Tw/0?wx_fmt=png" data-cropx1="5.372781065088757" data-cropx2="454" data-cropy1="0" data-cropy2="460.71597633136093" data-ratio="1.0267260579064588" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnLlUUmUo715WEWYB4TJ7J8ZE4OFFPJ3HWauUlJyuQTALVjibibib3EH7kQ/640?wx_fmt=jpeg" data-type="jpeg" data-w="449" style="width: 334px;height: 343px;"></p>
<p><span style="font-size: 15px;"><br></span></p>
<p><span style="font-size: 15px;">再来看看更多的生成回复。从这些回复中可以看到，有些回复是包含了预测的词汇的，有些回复则没有显性包含。这说明 [3] 提出的方法确实可以做到 implicitly content-introducing。</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-ratio="0.4" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnTtTE7m4ic3dal2DDUJ5f7jIqYWf8y8peTbbufG25H5UfupKvIY1l6mg/640?wx_fmt=png" data-type="png" data-w="915"></p>
<p><img class="" data-ratio="0.4444444444444444" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnCQuOlsT9dWSt2MdCGzRA9ibfTaa2l4BfWOtS6jiaUB1ofX6tpp3mN07Q/640?wx_fmt=png" data-type="png" data-w="927" style=""></p>
<p><span style="font-size: 15px;"></span><br></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><br></p>
<p><span style="font-size: 15px;">下面要介绍的这篇工作有一点点不同。在限定词汇方面，之前的工作都是找到一个或多个特定的词汇，强加给解码器。而这篇工作 [4] 则从另一个角度来帮忙解码器生成更好的回复——从解码词表中，动态剔除完全不相关的词汇，让解码器不再受无关词汇的扰乱，同时加快解码过程。一图胜千言：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.5203045685279187" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnkxKb6LJC0oSruRl0zOEfpCpYngVwIO3UUaH8BbFQa8mADQfDQrD2GQ/640?wx_fmt=png" data-type="png" data-w="788"></p>
<p><span style="font-size: 15px;">这样的方法最直观的效果就是解码的搜索空间小了，所以先来看看效率提升：</span></p>
<p><span style="font-size: 15px;"></span></p>
<p style="text-align:center;"><img class="" data-croporisrc="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKndesaS7Vibt0FnoNJmOLucD6rx7u8o9UKUvlD11ROvKMh2KQtLlCxl6g/640?wx_fmt=png" data-cropx1="0" data-cropx2="471" data-cropy1="0" data-cropy2="208" data-ratio="0.4394904458598726" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn4ZiaAiciaYnq6RrPdicKKku5yyZJLCOeJ3d5VT3BoPbFFTKicibZ0EaSia9HA/640?wx_fmt=jpeg" data-type="jpeg" data-w="471" style="width: 471px;height: 208px;"></p>
<p><span style="font-size: 15px;"></span><span style="font-size: 15px;">再来看一些生成的回复：</span></p>
<p style="text-align:center;"><img class="" data-ratio="0.3244635193133047" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_png/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKnPFkAxgbvZgWXqqpnoz7UiblI6g53q6SGU3WgxyV0IicP2euYCVhQ9zFA/640?wx_fmt=png" data-type="png" data-w="1165"></p>
<p><span style="font-size: 15px;">上图中的 keyword 就是动态词表构建过程中根据当前输入来找到的语义相关的词。图中给出的是概率最高的三个 keyword。这些 keyword 也可以认为是一种大范围的指定词汇。</span></p>
<p><span style="font-size: 15px;"></span></p>
<section style="box-sizing: border-box;"><section class="Powered-by-XIUMI V5" style="box-sizing: border-box;" powered-by="xiumi.us"><section class="" style="text-align: center;margin-top: 10px;margin-bottom: 10px;box-sizing: border-box;"><section class="" style="max-width: 100%;vertical-align: middle;display: inline-block;overflow: hidden !important;box-sizing: border-box;"><img data-ratio="0.078125" src="https://mmbiz.qpic.cn/mmbiz_gif/qrpuZA2scNqKiaszib9VCVaDMPsB5yjRKn1HjNic8RHw9TRTpRT7Bl3Zeefvd5iaVtgUOCt4yUVTEANQ5Kp0aibbaibg/640?wx_fmt=gif" data-type="gif" data-w="640" style="vertical-align: middle;box-sizing: border-box;"></section></section></section></section><p><br><span style="font-size: 15px;"></span></p>
<p><span style="font-size: 15px;">由于篇幅限制，剩下三篇工作留待明天分享。剩下三篇工作中，会提到一类非常重要的限定词汇：entity。希望大家多多支持！</span><img class="" data-copyright="0" data-ratio="1" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrtTyIlTh6t7NgpmC3A7mKic3NvqZR2OpguBeaWm0IGmKDlXYXrPmrFibtyZDl6v5Lxf9vuXOj9arKw/640?wx_fmt=jpeg" data-type="jpeg" data-w="1080" style="text-align: center;"></p>
<p style="text-align: center;"><img class="" data-copyright="0" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/mmbiz_jpg/qrpuZA2scNrsphEe4AVpwPMbFgFXp4AdEv82ibslNpJVyuGn3IZwAyB3OFScDiaqJkSTV8rUUxq5mibG6EuG6em3Q/640?wx_fmt=jpeg" data-type="jpeg" data-w="1240"></p>
                </div>]]></description>
      <link>http://mp.weixin.qq.com/s?__biz=MzAwMjM3MTc5OA%3D%3D&amp;mid=2652693190&amp;idx=1&amp;sn=92d478e13527510a23b369746ab0fde6</link>
      <pubDate>Sat, 09 Jun 2018 19:48:05 -0700</pubDate>
      <guid isPermaLink="false">890</guid>
    </item>
  </channel>
</rss>
