<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by 莊鐵頭 on Medium]]></title>
        <description><![CDATA[Stories by 莊鐵頭 on Medium]]></description>
        <link>https://medium.com/@ironhead.chuang?source=rss-be351d550baa------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Du6gJ7QLZFOdADYO.</url>
            <title>Stories by 莊鐵頭 on Medium</title>
            <link>https://medium.com/@ironhead.chuang?source=rss-be351d550baa------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 22:27:13 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ironhead.chuang/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AAAI-19 小記]]></title>
            <link>https://medium.com/blendvision/aaai-19-%E5%B0%8F%E8%A8%98-be899e6c3edb?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/be899e6c3edb</guid>
            <category><![CDATA[events]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Wed, 13 Feb 2019 17:32:47 GMT</pubDate>
            <atom:updated>2019-02-18T02:34:28.425Z</atom:updated>
            <content:encoded><![CDATA[<h3>源起</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gjBRNZrLYbUJqEpVN4Mh3w.jpeg" /><figcaption>第一天下午，人還不多。這個空間之後作為 Poster 與 Job Fair 的場地。</figcaption></figure><p>原本 2018 的計畫是參加 NIPS，但是<a href="https://twitter.com/karpathy/status/1037011082777251841?lang=en">門票在開放註冊後十五分鐘內售罄</a>，收到候補通知時議程也即將開始，跟公司討論後轉換目標到另一個研討會 AAAI-19，於是就在過年前飛往夏威夷了。由於團隊與個人目標，對這個研討會中推薦系統和電腦視覺的題目特別感興趣。議程總共有五天，包括兩天的工作坊跟教學，和三天的主議程。我第二天開始在會議場地進出。</p><h3>Tutorial Forum</h3><p>報名了一個跟推薦系統相關的課程「<a href="https://preferred.ai/aaai19-tutorial/">Recent Advances in Scalable Retrieval of Personalized Recommendations</a>」, 講授的內容是關於 Top-K 推薦時會遇到的效能問題，主要的方向是改變物件的特徵與相似度計算的效能。課程中看到了很多新東西，跟公司現在執行中的專案相印證也有些啟發。當你對一個課程的內容有些接觸卻又瞭解不多時，那上課的效果最好。課程中也列出了一些相關的論文，我想是值得一一瀏覽的。</p><h3>Workshop</h3><p>另外報名了一個跟推薦問題相關的工作坊「<a href="https://recnlp2019.github.io/">Recommender Systems and Natural Language Processing</a>」，形式上就是一個報告論文的場合，但是這一場的基調演講很精采：「Broad Learning via Fusion of Heterogeneous Information for Recommendations」談如何融合各種資料，包含來自不同社群網路的內容來協助推薦系統。目前找不到投影片，不過應該可以從<a href="https://www.cs.uic.edu/~psyu/">講者</a>的著作中拼湊出內容，預計會挑幾篇論文瀏覽。</p><h3>亮點</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F313933438%3Fapp_id%3D122963&amp;dntp=1&amp;url=https%3A%2F%2Fvimeo.com%2F313933438&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F755380818_1280.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=vimeo" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/1b4da5eaff9909bbffb87e1bf44e878d/href">https://medium.com/media/1b4da5eaff9909bbffb87e1bf44e878d/href</a></iframe><p>主持人們報告既有的 AI Roadmap 與已經執行的項目、學術界面臨業界誘惑的困境及參與者的建議等。</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F313941176%3Fapp_id%3D122963&amp;dntp=1&amp;url=https%3A%2F%2Fvimeo.com%2F313941176&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F755387098_1280.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/098388a5a0e97e7b472f71cc3f753539/href">https://medium.com/media/098388a5a0e97e7b472f71cc3f753539/href</a></iframe><p>過去因為對生成式對抗網路的興趣，<a href="https://github.com/imironhead/ml_gans">重製了數篇論文的結果</a>，所以對這個領域還算熟悉，不過對後續的發展認識不多，像是攻擊既有模型的手段等。Ian 在這個演講裡介紹了各種對抗模型產生的新研究方向，比如對抗模型在需要高度安全的應用上扮演的角色。</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8Hlw12dsUef5h2uXJ41sgA.jpeg" /><figcaption>台灣來的 Poster，剛好是最近感興趣的題目。</figcaption></figure><p>第一次參加學術會議，所以也第一次見識到 Poster 的場面。據說這次特別的地方是連口頭報告的論文都有 Poster 的位置。會場提供簡易的食物，讓有興趣的人群可以端著食物四處看海報。我三個晚上都在會場穿梭，可惜對機器學習之外的領域涉略不多，只看得懂一部份的海報。</p><h3>八卦</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lfsZHfwwk2ZWFHiYqfpxww.jpeg" /><figcaption>從 Diamond Head 可以看到會議舉辦的飯店</figcaption></figure><p>所有的議程結束後步行到 Diamond Head，路上聽到兩個中國人的對話：「估計第一作者有 70% 是中國人，但是做的東西還比較淺，重要或有趣的還是米國人在做。」不知道是不是<a href="https://www.nytimes.com/2017/07/20/business/china-artificial-intelligence.html">中國的政策目標</a>開始發酵了。</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=be899e6c3edb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/aaai-19-%E5%B0%8F%E8%A8%98-be899e6c3edb">AAAI-19 小記</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gossip of A.I. — #025]]></title>
            <link>https://medium.com/blendvision/gossip-of-a-i-025-e0dd659f617d?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/e0dd659f617d</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Mon, 08 Jan 2018 08:21:26 GMT</pubDate>
            <atom:updated>2018-01-08T08:21:26.442Z</atom:updated>
            <content:encoded><![CDATA[<h3>Gossip of A.I. — #025</h3><h3>Cost of Tuning Inception-v3</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PEyER6p6I8imMq81nkBFPA.png" /><figcaption>Schematic diagram of Inception-v3</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/380/1*bghIwWfQkoVf6abOWzOrIQ.jpeg" /><figcaption>A newsstand in China</figcaption></figure><p>In China, tuning a convolution neural network layer costs you 5 dollars while tuning a recursive neural network layer costs you 8 dollars.</p><p>Found the picture <a href="https://www.facebook.com/photo.php?fbid=10155844076700833&amp;set=gm.1734557069941415&amp;type=3&amp;theater">here</a>.</p><p><a href="https://research.googleblog.com/2016/03/train-your-own-image-classifier-with.html">Inception-v3</a></p><h3>Locate Birds in Zealandia</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*W4rY-DQkw_hFhAkItvYiDw.png" /></figure><p>Tens of thousands of birdsong recordings were captured at 50 locations. How could environmentalists know what the singing birds were? Machine Learning!</p><p><a href="https://blog.google/topics/machine-learning/machine-learning-gives-environmentalists-something-tweet-about/">Machine learning gives environmentalists something to tweet about</a></p><h3>Discover Invisible Planets</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FS_HRh0ZynjE%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DS_HRh0ZynjE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FS_HRh0ZynjE%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/cd27c57a5157904641057e029c19be19/href">https://medium.com/media/cd27c57a5157904641057e029c19be19/href</a></iframe><p>There is a far far away star, how can we find if there is an invisible planet around it? Machine learning!</p><p><a href="https://www.nasa.gov/press-release/artificial-intelligence-nasa-data-used-to-discover-eighth-planet-circling-distant-star">Artificial Intelligence, NASA Data Used to Discover Eighth Planet Circling Distant Star</a></p><h3>Updating Google Maps with Deep Learning and Street View</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/566/1*89o5cDrFFY7F6fZhBH407Q.png" /><figcaption>The business name, Zelina Penus, was correctly predicted by machine learning instead of “Bridgestone”.</figcaption></figure><p>Google keeps updating Google Map and machine learning helps a lot. Check <a href="https://www.justinobeirne.com/google-maps-moat">Google Map’s Moat</a> to see the advantage of Google Map.</p><p><a href="https://research.googleblog.com/2017/05/updating-google-maps-with-deep-learning.html">Google Research Blog</a></p><p><a href="https://www.justinobeirne.com/google-maps-moat">Google Map’s Moat</a></p><h3>Visual to Sound</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FeBrsJpuPol0%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DeBrsJpuPol0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FeBrsJpuPol0%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/bed5e639e9e7084294b1a80cdd1f49c5/href">https://medium.com/media/bed5e639e9e7084294b1a80cdd1f49c5/href</a></iframe><p>Magic! We can image the sounds of a video clip by watching it. Machine can do it too!</p><p><a href="http://bvision11.cs.unc.edu/bigpen/yipin/visual2sound_webpage/visual2sound.html">Visual to Sound: Generating Natural Sound for Videos in the Wild</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e0dd659f617d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/gossip-of-a-i-025-e0dd659f617d">Gossip of A.I. — #025</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gossip of A.I. — #024]]></title>
            <link>https://medium.com/blendvision/gossip-of-a-i-024-333a2e79286d?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/333a2e79286d</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Mon, 18 Dec 2017 08:43:04 GMT</pubDate>
            <atom:updated>2017-12-18T08:48:01.297Z</atom:updated>
            <content:encoded><![CDATA[<h3>Gossip of A.I. — #024</h3><h3>NIPS Top 10 Institutions</h3><h3>Andrew Moore on Twitter</h3><p>Top ten institutions by @NipsConference accepted papers in 2017. Curse you @Google for pipping @CarnegieMellon , but you guys know this whole &quot;internet&quot; thing is just a fad right? We&#39;ll be back at the top soon.</p><ul><li>Google + DeepMind : 13.4%</li><li>Industrial : 19.3%, Academic: 33.1%</li><li>Where is Facebook &amp; OpenAI?</li><li>業界：19.3%，學界：33.1%</li><li>沒有 <a href="https://research.fb.com/category/facebook-ai-research-fair/">Facebook</a> 跟 <a href="https://openai.com/">OpenAI</a> ？</li></ul><h3>Landing.ai</h3><h3>Andrew Ng on Twitter</h3><p>Announcing my new company! Landing.ai will help companies transform for the AI era. We&#39;re starting with the manufacturing industry. https://t.co/QlTLJQwOkL</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F3bs9jRFUIlM%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D3bs9jRFUIlM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F3bs9jRFUIlM%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/765249d2b4f2d0be66087b103aa552d9/href">https://medium.com/media/765249d2b4f2d0be66087b103aa552d9/href</a></iframe><p>Looks like a consulting company to me. The video shows examples that can be improved by A.I.: important, simple but exhausting works.</p><p>看起來像顧問公司。影片中有可用機器學習解決的實例問題：重要、簡單但是費時費力。</p><h3>Microsoft Extends AirSim to Include Autonomous Car Research</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*MT2lbda7OUJ6QWv5RkNApA.jpeg" /></figure><p><a href="https://www.microsoft.com/en-us/research/blog/autonomous-car-research/">Microsoft extends AirSim to include autonomous car research - Microsoft Research</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FCauKo089zm0%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DCauKo089zm0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FCauKo089zm0%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e14f3c241c37d5e76b9835db7e68e52c/href">https://medium.com/media/e14f3c241c37d5e76b9835db7e68e52c/href</a></iframe><p>An environment for autonomous car research. OpenAI had a similar environment but has been removed (<a href="https://deepdrive.io/">deepdrive</a> / <a href="https://github.com/openai/universe/issues/126">OpenAI universe</a>) a year ago.</p><p>微軟 AirSim 提供自駕車研究環境。OpenAI 曾經釋出過類似的環境，但是因為法律問題移除了 (<a href="https://deepdrive.io/">deepdrive</a> / <a href="https://github.com/openai/universe/issues/126">OpenAI universe</a>)。</p><h3>TopoSketch</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FlNjXbKNhtGA%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DlNjXbKNhtGA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FlNjXbKNhtGA%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f0e6b1538c2ee2070f250bf844b1cac0/href">https://medium.com/media/f0e6b1538c2ee2070f250bf844b1cac0/href</a></iframe><p><a href="https://vusd.github.io/toposketch/">TopoSketch Web App</a></p><p>We can interpolate faces if we have their latent vectors. Check the demo to see a new proposed method to create animations.</p><p>如果知道臉部圖片的特徵，就可以內插臉的表情。TopoSketch 用連結特徵向量的方式來建立動畫。</p><h3>Light Cascaded Convolutional Neural Networks for Accurate Player Detection</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fwe0GUTB3wVY%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dwe0GUTB3wVY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fwe0GUTB3wVY%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/296dcead63db73f994c5834839b9c3f2/href">https://medium.com/media/296dcead63db73f994c5834839b9c3f2/href</a></iframe><p><a href="http://www.cs.ubc.ca/~jhchen14/ccnn_player_detection/">Player detection</a></p><p>Use convolutional neural networks to locate players’ on the court. There are some use case here:</p><ul><li>During a live broadcast, reporter may need the positions to show possible strategies.</li><li>Sometimes people may want to build animations of a game and it’s difficult to label all positions (over entire game).</li></ul><p>使用卷積網路定位球場上的選手。實際需求：</p><ul><li>賽事轉播中球評可能需要球員的位置來講解戰術。</li><li>把比賽的錄影以動畫方式呈現（視角可能不同，標記所有選手在整場比賽的位置費時費力）</li></ul><h3>High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs</h3><p><a href="https://tcwang0509.github.io/pix2pixHD/">High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F3AIpPlzM_qs%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D3AIpPlzM_qs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F3AIpPlzM_qs%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/2015a125d137efafd6b4a6cabacdab2f/href">https://medium.com/media/2015a125d137efafd6b4a6cabacdab2f/href</a></iframe><p><a href="https://github.com/phillipi/pix2pix">Pix2Pix</a> is HD now! Check the video to see the future of photo creator. (I was thinking about the “creator”. Should I use creator or editor? It’s possible to edit pictures base on the pix2pix technic. It might work by translating a picture to its labeled version, editing the labeled image and translating it back to picture domain. Cycle consistence is required.)</p><p><a href="https://github.com/phillipi/pix2pix">Pix2Pix</a> 的高解析度版本，影片中有修圖的應用。直接編輯相片是可行的（目前的技術效果不會好），首先將相片轉換成色塊標籤版本，修改標籤再轉換回相片的版本。</p><p><a href="https://github.com/imironhead/ml_gans/tree/master/pix2pix">ironhead replicated pix2pix</a></p><p><a href="https://github.com/imironhead/ml_gans/tree/master/cyclegan">ironhead replicated cyclegan</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=333a2e79286d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/gossip-of-a-i-024-333a2e79286d">Gossip of A.I. — #024</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gossip of A.I. — #023]]></title>
            <link>https://medium.com/blendvision/gossip-of-a-i-023-f7c429a8d75d?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/f7c429a8d75d</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Thu, 14 Dec 2017 10:04:46 GMT</pubDate>
            <atom:updated>2017-12-14T10:04:46.960Z</atom:updated>
            <content:encoded><![CDATA[<h3>Gossip of A.I. — #023</h3><h3>Counting Crowds and Lines</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SLJ050AqQOx9SAHebGh05g.jpeg" /></figure><p><a href="https://medium.com/@dimroc/counting-crowds-and-lines-fe08c98612b3">Medium</a></p><p>出發買漢堡前想知道需不需要排隊嗎？訓練一個卷積網路吧！</p><p>Want to know how much time you have to stay in a line before getting your burger? Train a convolutional neural network!</p><h3>Artificial intelligence, video games and the mysteries of the mind</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fmqma6GpM7vM%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dmqma6GpM7vM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fmqma6GpM7vM%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/28c59e7e609f2d3ae1a05c1ce6444b95/href">https://medium.com/media/28c59e7e609f2d3ae1a05c1ce6444b95/href</a></iframe><p>DeepMind 的 Raia Hadsell 介紹強化學習與挑戰：</p><ul><li>缺乏生活經驗（無法直接取用）。</li><li>訓練好的模型變成飽合狀況，沒辦法學新習新的任務。</li><li>缺乏記憶。</li></ul><p>Raia Hadsell from DeepMind introduced reinforcement learning and its challenge:</p><ul><li>lack of experience</li><li>saturate on learnt tasks</li><li>memory issue</li></ul><h3>A Robotic Spy among the Fish</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FeoElxhjXfpQ%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DeoElxhjXfpQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FeoElxhjXfpQ%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/156afceb3f8e4079dd568a305f8290a2/href">https://medium.com/media/156afceb3f8e4079dd568a305f8290a2/href</a></iframe><p><a href="https://www.digitaltrends.com/cool-tech/robot-spy-fish/">DIGITAL TRENDS</a></p><p>用機器魚研究魚群的溝通方式與影響魚群。</p><p>Use robot fish to study how fish communicate and influence fish.</p><h3>2017 AI Index Report</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9ZXzKlHaPOWsQ4aqMSUnDg.png" /></figure><p><a href="http://www.aiindex.org/">AI Index Report</a></p><p>一份 2017 年關於 A.I. 進展的報告。</p><p>A report about A.I. status so far. I explore this report mainly because Jack Clark is one of its team members.</p><h3>Interactive Example Based Terrain Authoring with Conditional Adversarial Networks</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F5w685udM838%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D5w685udM838&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F5w685udM838%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/dddc99456318b1533c1abc857172ed4f/href">https://medium.com/media/dddc99456318b1533c1abc857172ed4f/href</a></iframe><p><a href="http://liris.cnrs.fr/eric.guerin/interactive-example-based-terrain-authoring-with-conditional-generative-adversarial-networks/">paper</a></p><p>用生成式對抗網路與使用者互動，快速生成地貌模型。</p><p>Use GANs to help artists to generate terrains.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f7c429a8d75d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/gossip-of-a-i-023-f7c429a8d75d">Gossip of A.I. — #023</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[KKTV Data Game 17.11 Benchmarks]]></title>
            <link>https://medium.com/blendvision/kktv-data-game-17-11-benchmarks-97fffc46fa23?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/97fffc46fa23</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Fri, 08 Dec 2017 10:33:45 GMT</pubDate>
            <atom:updated>2017-12-08T16:22:52.592Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mD52q7BBDSmmRokU5QNEBQ.png" /><figcaption>Data Game banner designed by Joshua Chao</figcaption></figure><h3>KKTV Data Game 17.11</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1_jyNbTp7g_KdQPlsS1olw.png" /><figcaption>5 benchmarks and their public scores</figcaption></figure><p>KKTV Data Game 17.11 剛結束，這是今年第四次辦資料遊戲，同時也是第二次公開競賽。題目是提供用戶過往的使用記錄，推測用戶未來一週的觀影時機（資料都經過去識別化的處理）。比賽前經過多次討論，主題定作我隨口說出的 <strong>I’m gonna make him a notification he can’t refuse</strong>，推播並不是遊戲的主軸，然而這個題目的應用之一是挑選推播的時機與對象。我依然是主辦人之一，除了設定平台設定與資料檢查之外，最有趣的部份當屬提供遊戲的參考基準。這次共有五個不同的基準，呼應教父的經典台詞，五個基準也對應到五個跟教父有關的角色上。</p><h3>Benchmark: 生倉新八</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*PQKnD4qfJ85VCR6fx9MTqw.png" /><figcaption>AUC is 0.5 if the predictions are randomly generated</figcaption></figure><p>生倉新八是漫畫《內衣教父》裡的配角。這個基準由隨機猜測產生。這次比賽的評分方式是 <a href="https://medium.com/kkstream/auc-explained-85537509e3b3">AUC</a>，AUC 的特性是隨機猜測的結果會接近 0.5。</p><h3>Benchmark: 近藤靜也</h3><p>近藤靜也是漫畫《內衣教父》的主角。我們假設所有的用戶作息都很規律，有些生活充滿朝氣，用日劇通勤開始新的一天；有些生活多采多姿，以韓劇在睡前劃下美好句點。如果所有用戶的生活都如此規律，這週看劇的模式該跟上週相去不遠。所以我們決定把上週的記錄當作答案直接遞交，得到 AUC 0.678。還好分數不漂亮，不然 KKStream 的資料科學家就要失業了。不過這也是我們在設計比賽時做的第一項檢查：資料裡是否暗藏某種規律？0.678 雖然不是高分，但也顯示用戶習慣與時間的關聯性。</p><p>一個有趣的問題是：既然隨機猜測的 AUC 是 0.5，什麼人有辦法交出這個比賽的最低分數？為什麼？</p><h3>Benchmark: Santino Corleone</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*2AQiUhmNrQ7qWL_1GijLlg.png" /></figure><p><a href="https://en.wikipedia.org/wiki/Sonny_Corleone">Santino Corleone</a> 是<a href="https://en.wikipedia.org/wiki/Vito_Corleone">教父</a>的長子。驗證資料時，直覺是用戶在 T + 1 週的觀影時間點會與第 T 週的的記錄相關。但是如果第 T 週時有些用戶正在準備資料競賽呢？如果有些用戶去員工旅遊了呢？那就用上所有的資料吧！把每週的資料拿來作線性迴歸，可以達到 AUC 0.82。</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/473ae2d0a6fa152816507d867e94ffd5/href">https://medium.com/media/473ae2d0a6fa152816507d867e94ffd5/href</a></iframe><p>一個很笨的方法，而且這個做法會在最後一步再次出現。實際上可以直接用函式庫完成。</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/ef4e15fdb3ba7c547bc29941883f9755/href">https://medium.com/media/ef4e15fdb3ba7c547bc29941883f9755/href</a></iframe><h3>Benchmark: Michael Corleone</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*qXMEthAyMclahh3xZOgR4w.png" /></figure><p><a href="https://en.wikipedia.org/wiki/Michael_Corleone">Michael Corleone</a> 是<a href="https://en.wikipedia.org/wiki/Vito_Corleone">教父</a>的最年輕的兒子。現在流行把能夠變成圖片的資料交給卷積網路，所以做了一些嚐試。不過還是想試 RNN，所以沒花太多時間在這裡。</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/ea393126c7582c6175252038c69f73f0/href">https://medium.com/media/ea393126c7582c6175252038c69f73f0/href</a></iframe><p>Loss function 還是選 cross entropy。另外用了 <a href="https://arxiv.org/abs/1607.08022">instance norm</a> 跟 leaky ReLu。</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b4c097d12369cfeb8e48cf3716e85cd8/href">https://medium.com/media/b4c097d12369cfeb8e48cf3716e85cd8/href</a></iframe><h3>Benchmark: Vito Corleone</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*3SYN4ohTCd3D648dhcYYNQ.png" /></figure><p><a href="https://en.wikipedia.org/wiki/Vito_Corleone">Vito Corleone</a> 是<a href="https://en.wikipedia.org/wiki/Vito_Corleone">教父</a>。</p><p>因為這個題目跟時間有關，而且平常少有機會接觸 RNN，所以這次決定用 RNN 嚐試。</p><ul><li>建立基本的模型並確定效能跟卷積網路相近，然後隨意調整參數到超越卷積網路的效能。但是調校模型應該有脈胳可尋，不該隨機嚐試，這是我需要特別練習的地方。</li><li>尋找改善 LSTM 效果的方法，第一次看到把 initial state 當作變數訓練會有比較好的效果。</li><li>把整個程式改成隨機產生模型（比如 LSTM 一至五層隨機決定）、訓練、驗證並存下每次的結果。這時候發現兩個有趣的現象：隨機模型的效能有個頂點，偶爾會產生接近頂點的結果，但是機會不多。另外上傳的幾次結果裡，驗證集跟測試集間的 AUC 差距變化很小（AUC(validation) — AUC(test) ~ D）。目前為止所有的模型都是用前四萬個用戶的資料當訓練集，餘下的部份當驗證集，每次訓練過程都保留驗證集 AUC 最高時的結果。</li><li>開始想試 Bagging，修改程式，一次生成數組的訓練集（最後有 90 組），每次都用隨機的模型在隨機的資料分佈上訓練（但是基礎模型應該有缺陷，所以最後單一模型的 AUC 不高）。到這裡其實純粹只想讓分數變高而已，整個方法越多組越不實用。</li><li>最後手上有大量的結果（約 90 種資料分佈，619 組結果）。這時候發現另一個有趣的現象：有些分佈會出現很好的驗證 AUC，但是跟測試 AUC 相差很遠，我想這是這組資料該特別花時間研究的部份，為什麼不同的訓練資料分佈結果會差這麼多（驗證 AUC 0.8850 ~ 0.8945）。</li><li>這時有 90 種資料分佈，該如何 ensemble 呢？發揮想像力的時間到了，用了類似線性組合的方式完成。最後的結果 LSTM 最佳 private score 是 0.88957，ensemble 後到 0.89112，增加 0.00155。</li><li>變種的Ensemble ：目的是整合所有夠好的驗證結果，難處是每一個驗證集的的資料分佈不同。做法是列出所有的驗證集結果，假設所有的驗證資料能覆蓋所有的訓練集與驗證集，透過合理的權重加總後可以得到更好的結果。因為每個驗證集不完全吻合，也就是某些用戶的分數只來自部份的模型，這樣的用戶的分數在加權後同時除以這些模型的權重和。然後持續微調權重讓每個測試集以外的用戶加權分數有更好的 AUC，接著把這組權重用在測試集用戶上得到最後的結果。</li></ul><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/acdb45247d600f6d3f245cc277d24e5c/href">https://medium.com/media/acdb45247d600f6d3f245cc277d24e5c/href</a></iframe><h3>Next Step</h3><p>看到排行榜上的分數就想超越，這是我覺得最有趣的部份。下次我會嚐試不同的方法，比如 gradient boost，再用不同的 model 做 ensemble。</p><p>Happy data game!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97fffc46fa23" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/kktv-data-game-17-11-benchmarks-97fffc46fa23">KKTV Data Game 17.11 Benchmarks</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gossip of A.I. — #022]]></title>
            <link>https://medium.com/blendvision/gossip-of-a-i-022-b348bcecb09e?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/b348bcecb09e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Wed, 29 Nov 2017 08:25:31 GMT</pubDate>
            <atom:updated>2017-11-29T08:28:02.271Z</atom:updated>
            <content:encoded><![CDATA[<h3>Gossip of A.I. — #022</h3><h3>Winning Data Science Competitions</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VK4KLEbZpbeAD8Jn8v5zDg.png" /></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FClAZQI_B4t8%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DClAZQI_B4t8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FClAZQI_B4t8%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/4d95c8df72d8108d8e5f977728fbdcbd/href">https://medium.com/media/4d95c8df72d8108d8e5f977728fbdcbd/href</a></iframe><p><a href="http://bit.ly/kktv_dg_1711">KKTV Data Game 17.11</a> 即將結束，我們很興奮地發現排行榜上出現一位<a href="https://www.kaggle.com/jeongyoonlee">大師</a>。歡迎有興趣的朋友在結束前進來試試身手！</p><p><a href="http://bit.ly/kktv_dg_1711">KKTV Data Game 17.11</a> is nearing its end. We are surprised to have a <a href="https://www.kaggle.com/jeongyoonlee">Competitions Master</a> in our competition. Happy data game!</p><h3>Software 2.0</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CEOekss9H6mD0dZVwN9ZFw.png" /></figure><p>我是從 Andrej Karpathy 的 <a href="http://cs231n.stanford.edu/">CS231n</a> 開始入門深度學習的，之後對軟體的未來或多或少有些想像，但都不如他最近的一篇文章具體。<a href="https://medium.com/@karpathy/software-2-0-a64152b37c35">這篇文章</a>主要講述想像中深度學習將會如何改變軟體現況，值得一讀。</p><p>I started to study deep learning by following Andrej Karpathy’s <a href="http://cs231n.stanford.edu/">CS231n</a> and it was the best course I have found so far. Since then, I have some thoughts about how software will be changing in the near future, but none of them are as concrete as <a href="https://medium.com/@karpathy/software-2-0-a64152b37c35">Andrej Karpathy’s new blog post abot software 2.0</a>.</p><h3>Waymo Public Testing</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/482/1*bOBij_3sxpbKiuEMplK3DA.jpeg" /><figcaption><a href="http://costumefail.com/2012/11/03/driverless-car-costume/">http://costumefail.com/2012/11/03/driverless-car-costume/</a></figcaption></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FaaOB-ErYq6Y&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DaaOB-ErYq6Y&amp;image=http%3A%2F%2Fi.ytimg.com%2Fvi%2FaaOB-ErYq6Y%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/73de1660c32e7245f260413d235e381c/href">https://medium.com/media/73de1660c32e7245f260413d235e381c/href</a></iframe><p><a href="https://techcrunch.com/2017/11/07/waymo-now-testing-its-self-driving-cars-on-public-roads-with-no-one-at-the-wheel/">TechCrunch News</a></p><p>Alphabet Inc. Waymo 的自駕車開始新的測試，以後無人車的梗不能用了，因為這次方向盤前沒有駕駛員。</p><p>Waymo, subsidiary of Alphabet Inc., started their driverless car test without a safety driver at the driver position. The driverless costume will soon to be useless.</p><h3>Are You Talking to Me?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HuaT_Jwu9jWU1wGvg2F6ng.png" /></figure><p><a href="https://arxiv.org/abs/1711.07613">arXiv: Ary you talking to me?</a></p><p>看到一篇有趣的論文，用生成式對抗網路做圖片問答。像上圖一樣，訓練好的模型根據圖文（Image+Caption 欄）對問題（Question 欄）給出回應（Ours 欄）。</p><p>Found an interesting paper which use GANs. The image from the paper shows that a learnt model answers (Ours column) questions (Questing column) according to the pictures (Image+Caption column).</p><h3>Fused Video Stabilization</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fx5rHog6RnNQ%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dx5rHog6RnNQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fx5rHog6RnNQ%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fd019fcba37ee7843fe8641a0265bedc/href">https://medium.com/media/fd019fcba37ee7843fe8641a0265bedc/href</a></iframe><p><a href="https://research.googleblog.com/2017/11/fused-video-stabilization-on-pixel-2.html">Google Research Blog</a></p><p>Google 介紹手機相機的震動問題。然而，這篇最有趣的地方在收集了各種問題的影片。</p><p>Google introduced the shaking issue of mobile camera. However, the most valuable part is the collection of video clips about different issues.</p><p><a href="http://blog.ironhead.ninja/2017/11/27/gossip-of-ai-022.html">Ironhead’s Blog</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b348bcecb09e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/gossip-of-a-i-022-b348bcecb09e">Gossip of A.I. — #022</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AUC Explained]]></title>
            <link>https://medium.com/blendvision/auc-explained-85537509e3b3?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/85537509e3b3</guid>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Wed, 22 Nov 2017 14:51:03 GMT</pubDate>
            <atom:updated>2017-11-29T17:27:34.365Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*ouPUL22RxrNblhXBzToKmQ.gif" /><figcaption>Construct ROC curve and calculate AUC</figcaption></figure><p>KKTV + KKStream 在 Kaggle 上辦了一個<a href="http://bit.ly/kktv_dg_1711">比賽</a>，目標是猜測用戶在未來一週裡的上線時段（已去識別化）。前一陣子我利用週會的時間跟同事解釋過 AUC 的算法與用途，以下是投影片的逐字稿。</p><p>I am hosting <a href="http://bit.ly/kktv_dg_1711">KKTV Data Game 17.11</a>, a competition about predicting user behaviors, on Kaggle. We chose AUC as our metric to evaluate the performance. This is the slide I used to explain AUC to colleagues.</p><h3>What is AUC</h3><p>AUC 是 Area Under ROC Curve 的縮寫，用來衡量二元分類模型的表現。</p><p>AUC stands for Area Under ROC Curve. It is a metric to evaluate the performance of binary classification problems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MeMmqhQOtDImI-ZP0LzF_w.gif" /><figcaption>I am a little zombie, a happy jumpy little zombie.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*NOi63lmzKGvA7x4S-DIXmg.png" /></figure><p><a href="https://en.wikipedia.org/wiki/Receiver_operating_characteristic">ROC curve</a> 是「接收者操作特徵曲線」。現在我們要建立這種曲線來評估「僵屍病毒檢測器」的效能，首先準備十個可供檢驗的案例，並且先隱藏<strong>僵屍標籤</strong>。</p><p><a href="https://en.wikipedia.org/wiki/Receiver_operating_characteristic">ROC stands for Receiver Operating Characteristic</a>. We will build the ROC curve of a “<em>zombie disease detection device</em>” and use this curve to evaluate this device. There are 10 subjects here and we know whether they would become zombies in the future or not. But let’s hide the <strong>zombie labels</strong> for now.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*IkOnFXu5VynqGVnZLj2MCA.png" /></figure><p>用「僵屍病毒檢測器」檢驗所有的案例，列出檢測器的信心指數。信心指數表示檢測器對判斷案例已被感染有多高的信心。</p><p>Use the device to predict a value (confidence) for each subject. It’s a value of how strong the device believes a subject is infected.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*MShgs0Vsq3esw8iu1wnUDQ.png" /></figure><p>按照每個案例的信心指數將所有的案例重新排列。</p><p>Then we use the confidence as keys to sort all subjects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*j5dYl3SKxieuGw8Bl9Jb6g.png" /></figure><p>比較各個案例被感染的信心指數跟標籤，開始畫出曲線。</p><p>The ROC curve is drew by comparing the predictions and true labels.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*lIdWSQyyzuVs1g-cUKhRRg.png" /></figure><p>我們想根據信心指數把所有的案例分成兩組：是否被感染。所以挑一個閾值，只要信心指數高過閾值，就判定該案例已被感染；反之則判定為未受感染。首先將這個閾值設定為 1.0。因為最高的信心指數是 1.0，也就是所有案例的信心指數都不大於閾值，TPR 跟 FPR 都是 0.0，於是在圖表上 （0.0, 0.0） 的位置畫上一個小圈圈（左下角）。</p><ul><li><strong>TPR</strong> : 在所有<strong>已被感染</strong>的案例裡，被檢測器跟閾值歸類到已被感染的比例。</li><li><strong>FPR</strong> : 在所有<strong>未被感染</strong>的案例裡，被檢測器跟閾值歸類到已被感染的比例。</li></ul><p>We want to split all subjects into 2 groups: infected or not. A threshold is chosen for the judgement. If the confidence value is greater than the threshold, we consider it’s infected. All the others are considered as non-infected. In the beginning, we choose 1.0 as threshold. Because all confidence values are less than or equal to the threshold, both TPR and FPR are 0.0. So we draw a little brown dot on (0.0, 0.0) (the left-bottom corner).</p><ul><li><strong>TPR</strong> : True positive rate. Among all <strong>infected subjects</strong>, what’s the ratio of subjects are classified as infected.</li><li><strong>FPR</strong> : False positive rate. Among all <strong>non-infected subjects</strong>, what’s the ratio of subjects are classified as infected.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*MiqmqXFJ5T6A9rxCjv5ZbA.png" /></figure><p>將閾值降為 <strong>0.7</strong> 。TPR 變成 0.6 （有五個案例已被感染，但是只有三個藍色案例的信心指數比 0.7 高，所以 TPR = 3/5）且 FPR 維持不變。在這個位子也畫上一個小圈圈。</p><p>Let’s choose another threshold: <strong>0.7</strong>. TPR become 0.6 (5 subjects are infected but only 3 have confidence values greater than 0.7, so TPR = 3/5) while FPR is still 0.0. Put another little brown dot on the chart.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*kZ0YZX4UMM4yO1yAEeeEkg.png" /></figure><p>將閾值降為 <strong>0.6</strong> 。FPR 變成 0.2 （有五個案例未被感染，但是其中有一個白色案例因為信心指數高過閾值被判定成已感染，所以 FPR = 1/5）而 TPR 維持不變。在這個位子也畫上一個小圈圈。</p><p>Let’s choose another threshold: <strong>0.6</strong>. FPR become 0.2 (5 subjects are non-infected, but one of them are miss-classified as infected because its confidence value is greater than threshold, so FTR = 1/5) while TPR is still 0.6. Put another little brown dot on the chart.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*03MFVWGe6r13mD-lfb7kiA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*jxL09mTfv7UupMRJg-BKFA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*EwG0lwi3vrUdQIFc1-fahg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*sTP7GlU2MJAKicZOWoNgZw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*L5QubNEyrPKKI1mZHLXpaw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*kT-vcEMdK-mNIFI82jKm1A.png" /></figure><p>重復這個方法直到閾值變為 <strong>0.0</strong>，TPR 與 FPR 也變成 0.0 了，所以表格的右上角有一個小圈圈。</p><p>Repeat the process until the threshold become <strong>0.0</strong>. Both TPR and FTR are 0.0 so there is a brown dot on the right-top corner of the chart.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*-NOpaa0LT3DfnUCoickwVg.png" /></figure><p>連接所有的小圈圈就得到一條曲線，這條曲線就是「接收者操作特徵曲線」。</p><p>Connecting all little brown dots gives us a curve. This is the ROC curve.</p><p>將曲線下方的區塊塗上藍色，藍色區塊的面積就是 AUC 了。這個曲線的 AUC 是 <strong>0.84</strong> ，因為在表格上所有的 25 格裡，有 21 格被塗成藍色，AUC = 21/25 。</p><p>Area under the ROC curve, blue blocks on the chart, is AUC. In this chart, AUC is <strong>0.84</strong> (21 blue blocks, area of each block is 1/25 = 0.04).</p><h3>How can We Use AUC?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*nqkNG2LC3RHhuQSfRRYLqg.png" /></figure><ul><li>假設有100個案例。</li><li>根據統計結果，案例中的一成已被感染，也就是十個案例。</li><li>我們還沒辦法確定哪些案例已經被感染了。</li><li>我們有「僵屍病毒檢測器」了。</li><li>運氣不錯，已經能做出解毒劑了；運氣不好，只做得出35份。</li><li>我們該怎麼運用這些解毒劑？</li><li>Assume there are 100 subjects.</li><li>According to statistic, 10% of subjects are infected (10 subjects).</li><li>We do not know who is infected.</li><li>We have zombie disease detection devices.</li><li>We have only 35 cures.</li><li>How can we use those precious cures?</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*ZAaloy5PB5ecg9UFUD1HwA.png" /></figure><ul><li>我們隨機挑35個案例打針。</li><li>但是只有一成的案例被感染。</li><li>31 份解毒濟被浪費掉了！</li><li>We choose 35 subjects to cure, randomly.</li><li>But there are only 10% of subjects are infected.</li><li>31 cures are wasted!</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*VKkkaHg0p7cQfRH49js2GQ.png" /></figure><ul><li>「僵屍病毒檢測器」登場。</li><li>我們知道 ROC 了，挑了一個閾值使得 TPR 是 0.8 且 FPR 是 0.3。</li><li>Here comes the zombie disease detection devices.</li><li>We know the AUC and we choose a threshold so that th TPR is 0.8 and FPR is 0.3.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*Kr1irh2ln3fb_7kDCDXZ7A.png" /></figure><ul><li>根據這個閾值，對所有被判斷已感染的案例打針。</li><li>因為 TPR 是 0.8，所以已感染的案例中有八成的機會被檢測出，也就是八個已感染的案例被治癒。</li><li>因為 FPR 是 0.3，所以未感染的案例中有三成的機會被檢測錯，也就是 27 個末感染的案例被打針。</li><li>只有 27 份解毒劑被浪費，而且多治癒四個案例。</li><li>According to the threshold, we give cures to those who are classified as infected subjects.</li><li>Because TPR is 0.8, 80% of infected subjects are detected. So 8 subjects are cured.</li><li>Because FPR is 0.3, 30% of non-infected subjects are miss-classified. So 27 subjects are bitten…pity.</li><li>Only 27 cures are wasted, and 4 more infected subjects are cured.</li></ul><p>簡單說，有了 AUC ，就可以根據 AUC 跟需求調整閾值，對資源做更有效的運用。</p><p>So we can use our resource more efficiently if we know the AUC of prediction model and pick a better threshold.</p><p><a href="http://blog.ironhead.ninja/2017/11/22/auc-explained.html">ironhead’s blog</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=85537509e3b3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/auc-explained-85537509e3b3">AUC Explained</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gossip of A.I. — #021]]></title>
            <link>https://medium.com/blendvision/gossip-of-a-i-021-80a3cbd77ad5?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/80a3cbd77ad5</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Sun, 19 Nov 2017 05:33:57 GMT</pubDate>
            <atom:updated>2017-11-19T05:33:57.611Z</atom:updated>
            <content:encoded><![CDATA[<h3>Ironhead Wanted to be a Game Programmer</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*zJvoc2_pXE_jXMCJ11MI-w.jpeg" /><figcaption>My very first programming project.</figcaption></figure><p>2003 年剛退伍時，我想去遊戲公司做遊戲。當時完全不會寫程式，所以找了一家補習班上課，三個月後的期末專題就是上面這張圖。當時很喜歡暴雪的遊戲，雖然幾乎沒玩星海爭霸，不過最後決定用星海爭霸的圖片做一個 2.5D 的小遊戲。把這些圖片倒出來之後，我才開始感覺到暴雪花在開發遊戲上的心力。現在已經不想寫遊戲了，反倒對用機器學習的方法玩遊戲感到興趣，或許過一陣子會再試著玩強化學習吧！</p><p>卦點：星海爭霸只有 256 色。</p><p>Back to 2003, I wanted to be a game programmer. This picture is a screenshot of my term project in a game programming class. I like Blizzard so I found tools to extract sprites of StarCraft and made a small 2.5 D shooting game. Them I admired Blizzard more after exploring those sprites. I no longer want to make games anymore, but I am interested in playing video games with A.I.. Maybe I’ll try reinforcement learning again.</p><p>Fun fact: depth of those sprites is 8 bits.</p><h3>Learning Macromanagement in StarCraft using Deep Learning</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FgCa5OejosYA%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DgCa5OejosYA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FgCa5OejosYA%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/c206f15f89998d0f88e481c2a3d9ec71/href">https://medium.com/media/c206f15f89998d0f88e481c2a3d9ec71/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FPs_5P8Xx3_s%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DPs_5P8Xx3_s&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FPs_5P8Xx3_s%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/dbf9ae21aa07a0906af05fbf4135f831/href">https://medium.com/media/dbf9ae21aa07a0906af05fbf4135f831/href</a></iframe><p><a href="https://arxiv.org/abs/1707.03743">arXiv: Learning Macromanagement in StarCraft from Replays using Deep Learning</a></p><p>這篇論文嚐試模仿人類玩星海爭霸的生產策略。</p><p>This paper is about imitate how human players build their units.</p><h3>Procedural Terrain Generation with GANs</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FmkYQnErDIOE%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DmkYQnErDIOE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FmkYQnErDIOE%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ec9e2665e9ea46fd93fb3d55fa3b69f9/href">https://medium.com/media/ec9e2665e9ea46fd93fb3d55fa3b69f9/href</a></iframe><p><a href="https://arxiv.org/abs/1707.03383">arXiv: A step towards procedural terrain generation with GANs</a></p><p><a href="https://visibleearth.nasa.gov/view.php?id=73934">NASA Visible Earth</a></p><p>又是生成式對抗網路，我應該更早看這段影片的，這樣就可以在人工智慧年會上提了。這篇論文做的事主要是根據兩個生成式對抗網路的模型（DCGAN + pix2pix），用來產生遊戲中的地表。訓練資料是由美國太空總署提供的高解析度地圖（高度加顏色，一個像素代表一平方公里），由第一個生成網路產生高度表，再由第二個生成網路產生對應的地貌顏色。</p><p>Use high resolution hight map and texture map (21600 x 10800, 1 square kilo-meter per pixel) provided by NASA, they trained two GANs to produce realistic terrain. The first GAN is DCGAN which receive a noise vector and generate a hight-map to fool its discriminator. The second GAN is pix2pix which try to translate the height map into texture map. The technic may be used to generate large scale realistic terrain for some game genre (maybe dog fight?).</p><h3>NIPS 2017: Learning to Run</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*YIyXUukt2h9JtL-2KOjmUA.gif" /></figure><p><a href="https://github.com/stanfordnmbl/osim-rl">GitHub: Reinforcement learning environments with musculoskeletal models</a></p><p><a href="https://www.crowdai.org/challenges/nips-2017-learning-to-run">crowdAI</a></p><p>NIPS 2017 辦的比賽之一，目標是控制骷髏人的下半身探索世界。這是一個強化學習的問題：</p><ul><li>環境：取得一個 41D 的向量。</li><li>動作：控制肌肉，輸出一個 18D 的向量。</li><li>獎勵：有移動就有分數，走得越遠分數越高，但是如果做出超人的動作會扣分（韌帶不能亂拉）。</li></ul><p>One of NIPS 2017 competition track. The objective is to control the skeleton to explore the environment. This is a reinforcement learning problem:</p><ul><li>environment: In each frame, you receive a 41D vector.</li><li>actions: According to the environment, output a 18D vector to control muscles.</li><li>rewards: Moving in the world and the final distance get scores. Using ligament get penalty (split? :D).</li></ul><h3>MariFlow — Self-Driving Mario Kart w/Recurrent Neural Network</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FIpi40cb_RsI%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DIpi40cb_RsI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FIpi40cb_RsI%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ab2eaf80e4b0b0ce8d4f9ceeaf5c8abc/href">https://medium.com/media/ab2eaf80e4b0b0ce8d4f9ceeaf5c8abc/href</a></iframe><p>上次 Seth Bling 用 AI 玩馬利兄弟，這次他建了另一種 AI 來模仿人類玩馬利歐賽車。他用的方法是 LSTM / TensorFlow，先錄了 50 個小時的遊戲過程來當訓練資料。網路的輸入是遊戲畫面的灰階版本，輸出則是按八個不同按鍵的機率。我覺得最有趣的地方在他怎麼解決卡關的問題。有時候 AI 會不停的往牆上移動，這時候人類導師會出場修正問題，讓遊戲能繼續進行，過程也同時被記錄下來當作新的訓練資料，最後能訓練出冠軍車手。</p><p>You probably remember somebody played Mario with neural networks, it’s Seth Bling. This time he tried to build another kind of neural network to play Mario Kart. He built a recurrent neural network using TensorFlow to mimic how human played the game. The training data is 50 hours human game play records. For the LSTM, the input is gray scale video frames and the output is probabilities of pushing buttons. The most interesting part is how he handled the failed case. Sometimes the game was stuck (keep moving into the wall) and would not move forward, Seth Bling would take over the game and <strong>recorded how human fix the problems</strong>. Eventually, the A.I. can win the race.</p><h3>NeuralKart</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FEo07BAsyQ24%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DEo07BAsyQ24&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FEo07BAsyQ24%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d785737110f0102a282b73c0ea6d3fa2/href">https://medium.com/media/d785737110f0102a282b73c0ea6d3fa2/href</a></iframe><p><a href="https://github.com/rameshvarun/NeuralKart">GitHub: NeuralKart</a></p><p>這是 <a href="http://cs231n.stanford.edu/">CS231n</a> 學生的期末專題！目標是用卷積網路玩賽車，但是因為他們的 AI 學習諭練資料的結果不穩定，所以在遇到問題的時候會交給另一個 AI 控制，讓時間倒退，尋找更種可能的改進方式，然後記下好的結果當作新的訓練資料。</p><p>Awesome, this is students’ term project in <a href="http://cs231n.stanford.edu/">CS231n</a>. They want to play Mario Kart with convolutional neural network. But they found when AI learnt from human game play records, Mario was stuck sometimes. So they designed second A.I., every time when A.I. was stuck, the second A.I. would take over and search all possibilities to improve the playing. Then the searching results were recoded as new training data.</p><h3>Mario AI</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FL4KBBAwF_bE%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DL4KBBAwF_bE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FL4KBBAwF_bE%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/52fded76739a0314e473440553e26a32/href">https://medium.com/media/52fded76739a0314e473440553e26a32/href</a></iframe><p><a href="https://github.com/aleju/mario-ai">GitHub: Mario AI</a></p><p>用強化學習玩馬利兄弟！DQN 是我第一個認識的強化學習方法，這個專案有趣的地方在輸入資料是遊戲畫面（像素），結果有目的性（高分）。</p><p>Play Mario with deep q-learning. <a href="http://blog.ironhead.ninja/2016/09/08/openai-cartpole.html">Deep Q-Learning is also the first reinforcement learning algorithm I have learnt</a>. The environment is the video frame pixels and the actions are how to control Mario. And the reward is to finish the level as soon as possible.</p><p><a href="http://blog.ironhead.ninja/2017/11/20/gossip-of-ai-021.html">Ironhead’s Blog: Gossip of A.I. — #021</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=80a3cbd77ad5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/gossip-of-a-i-021-80a3cbd77ad5">Gossip of A.I. — #021</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gossip of A.I. — #020]]></title>
            <link>https://medium.com/blendvision/gossip-of-a-i-020-9d8c06b9dcd7?source=rss-be351d550baa------2</link>
            <guid isPermaLink="false">https://medium.com/p/9d8c06b9dcd7</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[莊鐵頭]]></dc:creator>
            <pubDate>Mon, 13 Nov 2017 10:31:29 GMT</pubDate>
            <atom:updated>2017-11-19T05:49:36.181Z</atom:updated>
            <content:encoded><![CDATA[<h3>Gossip of A.I. — #020</h3><h3>What is Gossip of A.I.</h3><p>四月開始，我固定在 KKStream 內部週會時演講，長度五至十五分鐘不等。內容全是關於機器學習與人工智慧的舊聞新聞，從名人八卦到有趣的論文都有。目的有二，一是協助建立公司的資料文化，二是幫助大家成長。由於與會的同事有四成左右為非技術背景，所以內容不深，旨在宣揚目前人工智慧的進度。最近跟同事討論，覺得既然都整理這些內容了，何不寫出來讓更多人看到？所以從這篇開始寫線上版，內容會跟週會的版本相近，每一則可能會有、也可能不會有我個人的註解。前面十九份就看我有沒有心力了。另外為了順便練習英文，每段都會有中英文版，但是不保證是中英對照。</p><p>I started to share A.I. and machine learning news regularly in April 2017. Each talk is 5 ~ 15 minutes long. The contents range from gossip to interesting papers. Beside building data culture in KKStream, I also want to help colleagues gain basic knowledge about A.I.. Around 40% of the attendees are non-developers, so there is almost no difficult topics among those contents. I had a discussion with colleagues recently, and we decided to share the contents on internet. They may or may not contain my comments. Like the title says, consider this a gossip column.</p><h3>AI/DS Conference 2017</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1lGhpdXAK0zVR1F2nBHUCg.jpeg" /><figcaption>4 speakers from KKBOX.</figcaption></figure><p>上週跟同事們一起去人工智慧年會跟資料科學年會。公司這次有四個講者，我也去介紹生成式對抗網路。</p><p>We attended AI/DS Conference 2017 last week. There were 4 speakers from KKBOX. I was there to talk about GANs.</p><h3>KKTV Data Game 17.11</h3><p>KKTV KKStrea 在 Kaggle 上共同主持一個關於資料科學的小比賽。比賽的目標是預測使用者的行為。具體來說，如果我們提個一個使用者前六個月觀影的時間記錄，你能告訴我們這個使用者未來一週可能會看影片的時間點嗎？</p><p><a href="https://www.kaggle.com/t/1594f541360d406d83cf7ec3cc57ff03">加入我們！！！</a></p><p>We are hosting an open data competition on Kaggle. The goal of this game is to better predict users’ future usages. In brief, if we tell you when a user watched video in the last 6 months, can you tell us when will the user watch videos in the following week?</p><p><a href="https://www.kaggle.com/t/1594f541360d406d83cf7ec3cc57ff03">Join Us!!!</a></p><h3>SC2 A.I. Workshop</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8V5qZ8CkDJ8IqSoDLH7k5g.png" /></figure><p>期待在圍棋之後，人工智慧也能超越電競選手！</p><p>Oriol Vinyals is from DeepMind. We all know that DeepMind is trying to build an A.I. that can defeat StarCraft professional players. Stay tight!</p><p><a href="https://twitter.com/OriolVinyalsML/status/926466086857457665">Oriol Vinyals on Twitter</a></p><h3>Introducing: Unity Machine Learning Agents</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Ffq0JBaiCYNA%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dfq0JBaiCYNA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Ffq0JBaiCYNA%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/af9368ff7a9e8f7aa67d1fb38e4243c6/href">https://medium.com/media/af9368ff7a9e8f7aa67d1fb38e4243c6/href</a></iframe><p>Unity 是一家很酷的公司，遊戲做得不好，結果把遊戲引擎推向全世界。現在他們也把觸角伸進機器學習了，遊戲確實是模擬強化學習的好方法，也更能引起人們的興趣。影片裡同時訓練 12 個 agent ，試圖控制平板把球維持在平板上。</p><p>Unity is a cool company and they rejected my application years ago though I match the requirement 100% (come on! who still knows about Windows COM?). Anyway, game engine is quite good on simulating the environment of reinforcement learning. In the video, Unity demonstrate training 12 agents simultaneously to balance a ball on a panel.</p><p><a href="https://blogs.unity3d.com/2017/09/19/introducing-unity-machine-learning-agents/">Unity Blog</a></p><h3>The State of Data Science &amp; Machine Learning</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YOiS6QFWnfjOOjKqIGhEBQ.png" /><figcaption>What data science methods are used at work?</figcaption></figure><p>Kaggle 做了一個關於資料科學跟機器學習的調察，並將結果公佈在上面的連結中。除了各種數據之外，視覺化的方式也很有趣。上圖是其中一項統計：在工作中最常用到的技術。我花最多時間玩的生成式對抗網路最不實用，難怪兩次演講被問到應用面時都只能說：「你問倒我了！」</p><p>Kaggle recently had a survey on data science and machine learning. Among so many interesting topics, I found this one interests me most. I have spent a lot of my spare time on playing GANs, and it is the least practical one. In my talks of GANs, I also have said “You Got Me” twice. Fortunately, I do it for fun.</p><p><a href="https://www.kaggle.com/surveys/2017">Kaggle survey</a></p><h3>The Treachery of ImageNet</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aD7EpQngFQDPy0_5uE2OxA.png" /></figure><p>雖然影相辨識有長足的進步，但是還是會在奇怪的地方失敗啊！</p><p>Image recognition failed on strange images.</p><p><a href="https://twitter.com/dribnet/status/925871135849193472">Tom White on Twitter</a></p><h3>MoCoGAN</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fTwwyAXf6BTPNgZAzBJe7g.gif" /></figure><p>看起來是篇有趣的論文。產生臉的時候有兩個控制項：內容與動作。固定內容項並改變動作項時，產生同一個人的不同表情（縱軸）。固定動作項並改變內容項時，產生不同人的同一個表情（橫軸）。</p><blockquote>When fixing the content code and changing the motion code, it generated the same person performs different expressions. When fixing the motion code and changing the content code, it generated different people performs the same expression.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/704/1*1beDhtmvzBrwEJdLFnswnw.gif" /></figure><p>生成式對抗網路產生的影片。</p><p>GAN generated videos.</p><p><a href="https://github.com/sergeytulyakov/mocogan">GitHub</a></p><p><a href="https://arxiv.org/abs/1707.04993">Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, Jan Kautz, “MoCoGAN: Decomposing Motion and Content for Video Generation”</a></p><p><a href="http://blog.ironhead.ninja/2017/11/13/gossip-of-ai-020.html">鐵頭的 blog</a></p><p><a href="http://blog.ironhead.ninja/2017/11/13/gossip-of-ai-020.html">ironhead’s blog</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9d8c06b9dcd7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blendvision/gossip-of-a-i-020-9d8c06b9dcd7">Gossip of A.I. — #020</a> was originally published in <a href="https://medium.com/blendvision">BlendVision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>