{"id":3176,"date":"2018-01-27T01:00:00","date_gmt":"2018-01-26T17:00:00","guid":{"rendered":"http:\/\/swarma.org\/%swarma-paper%\/"},"modified":"2018-12-07T09:37:57","modified_gmt":"2018-12-07T01:37:57","slug":"2017%e5%b9%b4%e4%ba%ba%e5%b7%a5%e6%99%ba%e8%83%bd%e4%b8%8e%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0-%e5%b9%b4%e5%ba%a6%e5%a4%a7%e4%ba%8b%e4%bb%b6","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=3176","title":{"rendered":"2017\u5e74\u4eba\u5de5\u667a\u80fd\u4e0e\u6df1\u5ea6\u5b66\u4e60-\u5e74\u5ea6\u5927\u4e8b\u4ef6"},"content":{"rendered":"
\n
\n
\"\u96c6\u667a\"<\/section>\n<\/section>\n


<\/p>\n

\n
\n
<\/section>\n

2017\u5373\u5c06\u7ed3\u675f\u3002\u6211\u5199\u7684\u4e1c\u897f\u8fdc\u4e0d\u53ca\u6211\u8ba1\u5212\u7684\u90a3\u4e48\u591a\u3002\u4f46\u662f\u6211\u5e0c\u671b\u660e\u5e74\u80fd\u6539\u53d8\u8fd9\u79cd\u72b6\u51b5\uff0c\u5e0c\u671b\u6709\u66f4\u591a\u5173\u4e8e\u5f3a\u5316\u5b66\u4e60\u3001\u8fdb\u5316\u7b97\u6cd5\u548c\u8d1d\u53f6\u65af\u65b9\u6cd5\u7684\u6559\u7a0b\u5c06\u4f1a\u51fa\u73b0\u5728 WildML\uff01\u4ee5\u603b\u7ed32017\u5e74\u6240\u53d1\u751f\u7684\u4e8b\u60c5\u4f5c\u4e3a\u65b0\u7684\u5f00\u59cb\uff0c\u8fd8\u6709\u6bd4\u8fd9\u66f4\u597d\u7684\u65b9\u6cd5\u5417\uff1f\u6240\u4ee5\u6211\u56de\u987e\u4e86\u6211\u7684 Twitter \u5386\u53f2\u548c WildML , \u4e3a\u5927\u5bb6\u603b\u7ed3\u4e3a\u5982\u4e0b\u7684\u51e0\u4e2a\u70ed\u70b9\u4e3b\u9898\u548c\u4e8b\u4ef6\u3002
<\/p>\n<\/section>\n<\/section>\n


<\/p>\n

\n

\u672c\u6587\u7ffb\u8bd1\u81eaDENNY BRITZ\u7684 WildML\u4e2d\u7684\u4e00\u7bc7\u6587\u7ae0AI and Deep Learning in 2017 \u2013 A Year in Review \u3002<\/span><\/p>\n<\/blockquote>\n


<\/p>\n

\n
\n
\n
<\/p>\n

\u5f3a\u5316\u5b66\u4e60\u5728\u6e38\u620f\u4e2d\u6218\u80dc\u4eba\u7c7b<\/span><\/strong><\/span><\/p>\n

<\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/p>\n

2017\u5e74\u4eba\u5de5\u667a\u80fd\u6700\u6210\u529f\u7684\u6545\u4e8b\u662f AlphaGo\u2014\u2014\u4e00\u4e2a\u51fb\u8d25\u4e86\u4eba\u7c7b\u9876\u7ea7\u56f4\u68cb\u68cb\u624b\u7684\u5f3a\u5316\u5b66\u4e60\u667a\u80fd\u4f53\u3002\u7531\u4e8e\u56f4\u68cb\u6e38\u620f\u7684\u641c\u7d22\u7a7a\u95f4\u6781\u5927\uff0c\u4eba\u4eec\u66fe\u7ecf\u8ba4\u4e3a\u673a\u5668\u5b66\u4e60\u6280\u672f\u5728\u591a\u5e74\u5185\u90fd\u65e0\u6cd5\u4e0b\u597d\u56f4\u68cb\u3002\u800c\u73b0\u5728\u7684 AlphaGo \u6240\u5411\u62ab\u9761\uff0c\u771f\u662f\u8ba9\u4eba\u60ca\u559c\uff01<\/p>\n


<\/span><\/p>\n

\u7b2c\u4e00\u7248\u7684 AlphaGo \u5229\u7528\u4eba\u7c7b\u68cb\u624b\u7684\u68cb\u8c31\u6570\u636e\u8bad\u7ec3, \u5e76\u901a\u8fc7\u81ea\u6211\u5bf9\u5f08\u548c\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\u7684\u65b9\u6cd5\u5f97\u5230\u8fdb\u4e00\u6b65\u6539\u8fdb\u3002\u4e0d\u4e45\u4e4b\u540e\uff0c\u540d\u4e3a\u201cAlphaGo Zero\u201d\u7684\u65b0\u7248\u672c\u6709\u4e86\u8fdb\u4e00\u6b65\u7a81\u7834\uff0c \u5b66\u4f1a\u4e86\u4ece\u96f6\u5f00\u59cb\uff0c\u4e0d\u7528\u4efb\u4f55\u4eba\u7c7b\u68cb\u8c31\u6570\u636e\uff0c\u4ec5\u4ec5\u9760\u81ea\u6211\u5bf9\u5f08\u5c31\u5b66\u4f1a\u4e86\u56f4\u68cb\uff0c\u5e76\u4e14\u8f7b\u800c\u6613\u4e3e\u5730\u51fb\u8d25\u4e86 AlphaGo \u7684\u7b2c\u4e00\u4e2a\u7248\u672c\u3002\u52302017\u5e74\u5e74\u5e95\uff0cAlphaGo Zero \u7b97\u6cd5\u62d3\u5c55\u6210\u4e3a Alpha Zero\uff0c\u4e0d\u4ec5\u80fd\u638c\u63e1\u56f4\u68cb\uff0c\u8fd8\u80fd\u4f7f\u7528\u76f8\u540c\u6280\u672f\u638c\u63e1\u56fd\u9645\u8c61\u68cb\u548c\u65e5\u672c\u5c06\u68cb\u3002\u6709\u8da3\u7684\u662f\uff0c\u8fd9\u4e9b\u7a0b\u5e8f\u7684\u68cb\u8def\u751a\u81f3\u80fd\u8ba9\u6700\u6709\u7ecf\u9a8c\u7684\u4eba\u7c7b\u68cb\u624b\u611f\u5230\u60ca\u8bb6\uff0c\u8fd9\u6fc0\u52b1\u7740\u68cb\u624b\u4eec\u4ece AlphaGo \u90a3\u91cc\u5b66\u4e60\u5e76\u8c03\u6574\u81ea\u5df1\u7684\u4e0b\u68cb\u98ce\u683c\u3002DeepMind \u8fd8\u53d1\u5e03\u4e86\u4e00\u6b3e\u5728\u7ebf\u5de5\u5177\uff0c\u7528 AlphaGo \u6280\u672f\u5e2e\u52a9\u4eba\u7c7b\u68cb\u624b\u63a2\u7d22\u56f4\u68cb\u65b0\u73a9\u6cd5\u3002<\/span><\/span><\/p>\n


<\/p>\n

\n
\"\u96c6\u667a\"<\/section>\n

<\/section>\n<\/section>\n

\u4f46\u56f4\u68cb\u4e0d\u662f\u6211\u4eec\u53d6\u5f97\u91cd\u5927\u8fdb\u6b65\u7684\u552f\u4e00\u6e38\u620f\u3002\u5361\u8010\u57fa\u6885\u9686\u5927\u5b66(CMU)\u7684\u7814\u7a76\u4eba\u5458\u5f00\u53d1\u4e86\u4e00\u4e2a\u540d\u4e3a LIbratus<\/strong> \u7684\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\uff0c\u5728\u4e3a\u671f20\u5929\u7684\u4e00\u5bf9\u4e00\u3001\u65e0\u9650\u6ce8\u5fb7\u5dde\u6251\u514b(Heads-up, No-Limit Texas Hold\u2019em)\u4eba\u673a\u5bf9\u6218\u4e2d\uff0c\u51fb\u8d25\u4e86\u4eba\u7c7b\u9876\u7ea7\u6251\u514b\u73a9\u5bb6\u3002\u66f4\u65e9\u4e4b\u524d\uff0c\u7531\u5e03\u62c9\u683c\u5927\u5b66\u3001\u6377\u514b\u7406\u5de5\u5927\u5b66\u548c\u52a0\u62ff\u5927\u963f\u5c14\u4f2f\u5854\u5927\u5b66\u7684\u7814\u7a76\u4eba\u5458\u5f00\u53d1\u7684 <\/span>DeepStack<\/strong> \u7cfb\u7edf\uff0c\u7b2c\u4e00\u6b21\u51fb\u8d25\u4eba\u7c7b\u804c\u4e1a\u6251\u514b\u73a9\u5bb6\u3002\u8fd9\u4e24\u4e2a\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u90fd\u4f7f\u7528\u4e00\u5bf9\u4e00\u7684\u73a9\u6cd5\uff0c\u662f\u4e24\u4e2a\u73a9\u5bb6\u4e4b\u95f4\u7684\u6e38\u620f\uff0c\u8fd9\u8fdc\u8fdc\u6bd4\u591a\u4eba\u53c2\u4e0e\u7684\u6251\u514b\u6e38\u620f\u5bb9\u6613\u3002\u540e\u8005\u5f88\u53ef\u80fd\u57282018\u5e74\u53d6\u5f97\u66f4\u591a\u8fdb\u5c55\u3002<\/span><\/p>\n


<\/span><\/p>\n

\u5f3a\u5316\u5b66\u4e60\u7684\u4e0b\u4e00\u4e2a\u524d\u6cbf\u5e94\u7528\u53ef\u80fd\u662f\u66f4\u4e3a\u590d\u6742\u7684\u591a\u4eba\u6e38\u620f\uff0c\u5305\u62ec\u591a\u4eba\u6251\u514b\u7b49\u3002DeepMind \u6b63\u5728\u79ef\u6781\u7814\u7a76\u8ba9 AI \u5b66\u4f1a\u73a9\u661f\u9645\u4e89\u97382\uff0c\u4e3a\u6b64\u5b83\u4eec\u53d1\u5e03\u4e86\u4e00\u4e2a\u7814\u7a76\u73af\u5883\uff1bOpenAI \u57281v1\u7684 Dota2\u6e38\u620f\u4e2d\u5df2\u7ecf\u53d6\u5f97\u521d\u6b65\u6210\u679c\uff0c\u5b83\u4eec\u76ee\u6807\u662f\u8ba9 AI \u5728\u4e0d\u8fdc\u7684\u5c06\u6765\u53ef\u4ee5\u73a95v5\u7684 Dota2\u3002<\/span><\/p>\n


<\/p>\n

\n
\"\u96c6\u667a\"<\/section>\n<\/section>\n


<\/p>\n

\n
\n
\n
<\/p>\n

\u8fdb\u5316\u7b97\u6cd5\u5377\u571f\u91cd\u6765<\/span><\/strong><\/span><\/p>\n

<\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/p>\n

\u5728\u76d1\u7763\u5b66\u4e60\u4e2d\uff0c\u57fa\u4e8e\u68af\u5ea6\u7684\u53cd\u5411\u4f20\u64ad\u7b97\u6cd5\u8868\u73b0\u5f88\u597d\uff0c\u8fd9\u4e00\u70b9\u77ed\u671f\u5185\u4e0d\u4f1a\u53d8\u5316\u3002\u7136\u800c\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\uff0c\u8fdb\u5316\u7b56\u7565(Evolution Strategies,ES)\u6b63\u5377\u571f\u91cd\u6765\u3002\u539f\u56e0\u662f\uff0c\u5f3a\u5316\u5b66\u4e60\u7684\u6570\u636e\u901a\u5e38\u4e0d\u662f\u72ec\u7acb\u4e14\u5206\u5e03\u76f8\u540c\u7684\uff1b\u8fd9\u4e9b\u6570\u636e\u7684\u4fe1\u53f7\u8bef\u5dee\u66f4\u5c0f\uff1b\u800c\u4e14\u56e0\u4e3a\u9700\u8981\u63a2\u7d22\u5176\u89c4\u5f8b\uff0c\u4e0d\u4f9d\u8d56\u68af\u5ea6\u7684\u7b97\u6cd5\u4e5f\u53ef\u4ee5\u5f88\u597d\u5de5\u4f5c\u3002\u6b64\u5916\uff0c\u8fdb\u5316\u7b97\u6cd5\u53ef\u4ee5\u7ebf\u6027\u62d3\u5c55\u81f3\u6210\u5343\u4e0a\u4e07\u7684\u673a\u5668\uff0c\u8fdb\u884c\u6781\u5feb\u7684\u5e76\u884c\u8bad\u7ec3\u3002\u53ea\u9700\u8981\u6210\u767e\u4e0a\u5343\u7684\u5ec9\u4ef7 CPU \u5373\u53ef\u8bad\u7ec3\uff0c\u800c\u65e0\u9700\u6602\u8d35\u7684GPU\u3002<\/p>\n


<\/p>\n

2017\u5e74\u65e9\u4e9b\u65f6\u5019\uff0cOpenAI \u7684\u7814\u7a76\u4eba\u5458\u8bc1\u660e\uff0c\u8fdb\u5316\u7b56\u7565\u53ef\u4ee5\u8fbe\u5230\u4e0e\u6df1\u5ea6 Q-learning \u7b49\u6807\u51c6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u76f8\u5ab2\u7f8e\u7684\u6548\u679c\u3002\u52302017\u5e74\u5e74\u5e95\uff0cUber \u7684\u7814\u7a76\u56e2\u961f\u53d1\u5e03\u4e86\u4e00\u7bc7\u535a\u5ba2\u548c\u4e94\u7bc7\u7814\u7a76\u8bba\u6587\uff0c\u5c55\u793a\u4e86\u5bf9\u9057\u4f20\u7b97\u6cd5(Genetic Algorithm,GA)\u7684\u6700\u65b0\u7814\u7a76\u3002\u4ed6\u4eec\u4f7f\u7528\u7684\u9057\u4f20\u7b97\u6cd5\u6781\u5176\u7b80\u5355\uff0c\u65e0\u9700\u4efb\u4f55\u68af\u5ea6\u4fe1\u606f\uff0c\u5c31\u5b66\u4f1a\u4e86\u73a9\u590d\u6742\u7684\u7535\u5b50\u6e38\u620f\u3002\u5728 Frostbite \u8fd9\u6b3e\u6e38\u620f\u4e2d\uff0c\u9057\u4f20\u7b97\u6cd5\u7684\u5f97\u5206\u662f10500\uff0c\u800c DQN\u3001AC3\u548c ES \u7684\u5f97\u5206\u5747\u4e0d\u52301000\u3002<\/p>\n


<\/p>\n

\n
\"\u96c6\u667a\"<\/section>\n<\/section>\n


<\/p>\n

2018\u5e74\uff0c<\/span>\u6211\u4eec\u6781\u6709\u53ef\u80fd\u770b\u5230\u66f4\u591a\u5728\u8fdb\u5316\u7b97\u6cd5\u65b9\u9762\u7684\u5de5\u4f5c\u8fdb\u5c55\u3002<\/p>\n


<\/p>\n

\n
\n
\n
<\/p>\n

Wavenets\uff0cCNNs \u548c\u6ce8\u610f\u529b\u673a\u5236<\/span><\/strong><\/span><\/p>\n

<\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/p>\n

2017\u5e74\uff0cGoogle Tacotron 2\u8bed\u97f3\u5408\u6210\u7cfb\u7edf\u751f\u6210\u7684\u97f3\u9891\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u3002\u5b83\u57fa\u4e8e WaveNet \u8bed\u97f3\u5408\u6210\u6280\u672f\uff0c\u8fd9\u9879\u6280\u672f\u4e5f\u5e94\u7528\u4e8e Google \u52a9\u624b\u3002WaveNet \u7684\u901f\u5ea6\u5728\u8fc7\u53bb\u4e00\u5e74\u83b7\u5f97\u5de8\u5927\u63d0\u5347\u3002\u6b64\u524d\uff0cWaveNet \u4e5f\u66fe\u88ab\u5e94\u7528\u4e8e\u673a\u5668\u7ffb\u8bd1\uff0c\u4ee5\u63d0\u9ad8\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\u7684\u8bad\u7ec3\u901f\u5ea6\u3002<\/span><\/p>\n


<\/p>\n

\"\u96c6\u667a\"<\/p>\n


<\/p>\n

2017\u5e74\uff0c\u5728\u673a\u5668\u5b66\u4e60\u9886\u57df\uff0c\u8fdc\u79bb\u9700\u957f\u65f6\u95f4\u8bad\u7ec3\u7684\u3001\u6602\u8d35\u7684\u5faa\u73af\u67b6\u6784\u4f3c\u4e4e\u662f\u5927\u52bf\u6240\u8d8b\u3002\u201cAttention \u5c31\u662f\u4f60\u9700\u8981\u7684\u5168\u90e8<\/strong>\u201d\uff0c\u7814\u7a76\u4eba\u5458\u53ef\u4ee5\u5b8c\u5168\u6446\u8131\u5faa\u73af\u6216\u5377\u79ef\u7684\u67b6\u6784\uff0c\u4f7f\u7528\u66f4\u590d\u6742\u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4ee5\u6700\u4f4e\u7684\u6210\u672c\u5b9e\u73b0\u6700\u597d\u7684\u7ed3\u679c\u3002<\/span><\/p>\n


<\/p>\n

\n
\n
\n
<\/p>\n

\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u5e74<\/span><\/strong><\/span><\/p>\n

<\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/p>\n

\u5982\u679c\u7b80\u8981\u603b\u7ed32017\u5e74\uff0c\u90a3\u4e48\u53ef\u4ee5\u8bf4\u8fd9\u662f\u201c\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u5e74\u201d\u3002Facebook \u501f\u52a9\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6 PyTorch\uff0c\u6210\u4e3a\u6405\u5c40\u8005\u3002\u7531\u4e8e\u5176\u52a8\u6001\u8ba1\u7b97\u56fe<\/strong>\u7684\u7ed3\u6784\u548c Chainer \u6846\u67b6\u7684\u76f8\u4f3c\uff0cPyTorch \u6df1\u53d7 NLP \u7814\u7a76\u4eba\u5458\u7684\u559c\u7231\uff0c\u56e0\u4e3a\u4ed6\u4eec\u7ecf\u5e38\u8981\u5904\u7406\u52a8\u6001\u548c\u5faa\u73af\u7684\u7ed3\u6784\uff0c\u800c\u8fd9\u4e9b\u7ed3\u6784\u5f88\u96be\u5728 TensorFlow \u4e4b\u7c7b\u7684\u9759\u6001\u56fe\u6846\u67b6\u4e2d\u5c55\u793a\u3002<\/p>\n


<\/p>\n

2017\u5e74\uff0cTensorFlow \u6846\u67b6\u8fd0\u884c\u5f97\u76f8\u5f53\u597d\u30022017\u5e742\u6708\uff0c\u5177\u6709\u7a33\u5b9a\u548c\u540e\u5411\u517c\u5bb9 API \u7684 TensorFlow 1.0\u53d1\u5e03\u3002\u76ee\u524d TensorFlow \u7684\u7248\u672c\u662f1.4.1\u3002\u5728\u4e3b\u5e93\u4e4b\u5916\uff0cGoogle \u8fd8\u53d1\u5e03\u4e86\u51e0\u4e2a TensorFlow \u7684\u4f34\u4fa3\u5e93\uff0c\u5305\u62ec\u53ef\u4ee5\u751f\u6210\u52a8\u6001\u8ba1\u7b97\u56fe\u7684 Tensorflow Fold\u3001\u65b9\u4fbf\u6570\u636e\u9884\u5904\u7406\u7684 TensorFlow Transform\uff0c\u4ee5\u53ca DeepMind \u7684 Higher-Level Sonnet\u3002TensorFlow \u56e2\u961f\u8fd8\u516c\u5e03\u4e86\u4e00\u79cd\u7c7b\u4f3c\u4e8e PyTorch \u52a8\u6001\u8ba1\u7b97\u56fe\u7684 eager execution \u6a21\u5f0f\u3002\u5f53\u7136\uff0c\u8fd9\u4e00\u5e74\u9664\u4e86 Google \u548c Facebook\uff0c\u8bb8\u591a\u5176\u4ed6\u516c\u53f8\u4e5f\u52a0\u5165\u4e86\u673a\u5668\u5b66\u4e60\u6846\u67b6\u7684\u642d\u5efa\uff1a<\/span><\/p>\n

    \n
  • \n

    \u82f9\u679c\u516c\u53f8\u53d1\u5e03 CoreML 2. CoreML \u79fb\u52a8\u673a\u5668\u5b66\u4e60\u5e93<\/span><\/p>\n<\/li>\n

  • \n

    Uber \u7684\u7814\u7a76\u56e2\u961f\u53d1\u5e03\u4e86\u4e00\u79cd\u6df1\u5ea6\u6982\u7387\u7f16\u7a0b\u8bed\u8a00<\/span> Pyro<\/span><\/p>\n<\/li>\n

  • \n

    Amazon \u63a8\u51fa Gluon , \u4e00\u4e2a\u53ef\u4ee5\u5728 MXNet \u4e2d\u4f7f\u7528\u7684\u9ad8\u7ea7 API<\/span><\/p>\n<\/li>\n

  • \n

    Uber \u516c\u5e03\u4e86\u5176\u5185\u90e8\u4f7f\u7528\u7684\u673a\u5668\u5b66\u4e60\u57fa\u7840\u5e73\u53f0 Michelangelo \u7684\u8be6\u7ec6\u4fe1\u606f<\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n


    <\/span><\/p>\n

    \u7531\u4e8e\u73b0\u6709\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u5df2\u7ecf\u5f88\u591a\uff0cFacebook\u3001\u5fae\u8f6f\u8054\u5408\u63a8\u51fa ONNX \u6807\u51c6\uff0c\u7528\u4e8e\u8de8\u5e73\u53f0\u5206\u4eab\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u3002\u6bd4\u5982\u4f60\u53ef\u4ee5\u5728\u4e00\u4e2a\u6846\u67b6\u4e2d\u8bad\u7ec3\u4f60\u7684\u6a21\u578b\uff0c\u7136\u540e\u5728\u8ba9\u8be5\u6a21\u578b\u5728\u53e6\u4e00\u4e2a\u6846\u67b6\u4e2d\u5de5\u4f5c\u3002<\/span><\/p>\n


    <\/p>\n

    \u9664\u4e86\u901a\u7528\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u6211\u4eec\u8fd8\u770b\u5230\u5927\u91cf\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\uff0c\u5176\u4e2d\u5305\u62ec\uff1a<\/p>\n

      \n
    • \n

      OpenAI\uff0c\u4e00\u4e2a\u7528\u4e8e\u673a\u5668\u4eba\u6a21\u62df\u7684\u5f00\u6e90\u8f6f\u4ef6<\/p>\n<\/li>\n

    • \n

      OpenAI Baselines<\/span>\uff0c\u4e00\u7ec4\u9ad8\u8d28\u91cf\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u5b9e\u73b0<\/p>\n<\/li>\n

    • \n

      Tensorflow Agents\uff0c\u4e00\u4e2a\u6846\u67b6\uff0c\u53ef\u4ee5\u4f7f\u7528 Tensorflow \u5bf9 RL agent \u8fdb\u884c\u4f18\u5316<\/span><\/span><\/p>\n<\/li>\n

    • \n

      Unity ML Agents\uff0c<\/span>\u5141\u8bb8\u7814\u7a76\u8005\u548c\u5f00\u53d1\u8005\u4f7f\u7528 Unity \u7f16\u8f91\u5668\u521b\u5efa\u6e38\u620f\u548c\u6a21\u62df\u4f7f\u7528, \u5e76\u7528\u5f3a\u5316\u5b66\u4e60\u6765\u8bad\u7ec3<\/span><\/span><\/p>\n<\/li>\n

    • \n

      Nervana Coach\uff0c\u5141\u8bb8\u4f7f\u7528\u6700\u5148\u8fdb\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u8fdb\u884c\u5b9e\u9a8c<\/span><\/p>\n<\/li>\n

    • \n

      Facebook\u2019s ELF Facebook\uff0c\u4e00\u4e2a\u6e38\u620f AI \u7814\u7a76\u5e73\u53f0<\/p>\n<\/li>\n

    • \n

       <\/span>DeepMind Pycolab\uff0c\u4e00\u4e2a\u53ef\u5b9a\u5236\u7684 gridworld \u6e38\u620f\u5f15\u64ce<\/span><\/p>\n<\/li>\n

    • \n

      Geek.ai MAgent\uff0c\u4e00\u4e2a\u6e38\u620f AI \u7814\u7a76\u5e73\u53f0\uff0c\u7528\u4e8e\u591a\u4e3b\u4f53(many-agent)\u6e38\u620f\u7684\u5f3a\u5316\u5b66\u4e60<\/p>\n<\/li>\n<\/ul>\n


      <\/p>\n

      \u6211\u4eec\u8fd8\u770b\u5230\u4e00\u4e9b\u4e3a\u4e86\u8ba9\u6df1\u5ea6\u5b66\u4e60\u66f4\u6613\u7528\u7684 web \u6846\u67b6, \u6bd4\u5982 Google \u7684 deeplearn.js \u548c MIL WebDNN \u6267\u884c\u6846\u67b6\u3002\u4f46\u81f3\u5c11\u6709\u4e00\u4e2a\u66fe\u7ecf\u975e\u5e38\u53d7\u6b22\u8fce\u7684\u6846\u67b6\u5df2\u7ecf\u505c\u6b62\u7ef4\u62a4\uff0c\u90a3\u5c31\u662f Theano\u3002\u5f00\u53d1\u8005\u5728 Theano \u90ae\u4ef6\u5217\u8868\u7684\u516c\u544a\u4e2d\u5ba3\u5e03\uff0cTheano <\/span>1.0\u5c06\u662f\u6700\u540e\u4e00\u4e2a\u7248\u672c\u3002<\/p>\n


      <\/p>\n

      \n
      \n
      \n
      <\/p>\n

      \u5e94\u7528: \u4eba\u5de5\u667a\u80fd\u548c\u533b\u7597<\/span><\/strong><\/span><\/p>\n

      <\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      2017\u5e74\uff0c\u770b\u5230\u8bb8\u591a\u7c7b\u4f3c\u201c\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u89e3\u51b3\u533b\u5b66\u95ee\u9898\uff0c\u51fb\u8d25\u4eba\u7c7b\u4e13\u5bb6\u201d\u7684\u7092\u4f5c\u3002\u5bf9\u4e8e\u6ca1\u6709\u533b\u7597\u80cc\u666f\u7684\u4eba\uff0c\u8981\u4e86\u89e3 AI \u5728\u533b\u7597\u9886\u57df\u771f\u6b63\u7684\u6280\u672f\u7a81\u7834\u7edd\u975e\u6613\u4e8b\u3002\u60f3\u5bf9\u6b64\u6709\u5168\u9762\u4e86\u89e3\uff0c\u63a8\u8350\u9605\u8bfb Luke Oakden-Ranyner \u7684\u300a\u4eba\u7c7b\u533b\u751f\u7684\u7ec8\u7ed3\u300b\u7cfb\u5217\u6587\u7ae0(https:\/\/lukeoakdenrayner.wordpress.com\/2017\/04\/20\/the-end-of-human-doctors-introduction\/<\/span>)\u3002\u8fd9\u91cc\u6211\u4ec5\u7b80\u8981\u4ecb\u7ecd\u4e00\u4e9b\u8fdb\u5c55\uff1a<\/span><\/p>\n


      <\/p>\n

      \u4eca\u5e74\u533b\u7597\u9886\u57df\u6700\u70ed\u95e8\u7684\u65b0\u95fb\u662f\u65af\u5766\u798f\u5927\u5b66\u7684\u4e00\u4e2a\u7814\u7a76\u56e2\u961f\u53d1\u5e03\u7684\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5\uff0c\u8fd9\u4e2a\u7b97\u6cd5\u53ef\u4ee5\u5e2e\u76ae\u80a4\u79d1\u533b\u751f\u8bc6\u522b\u76ae\u80a4\u764c\u3002\u65af\u5766\u798f\u5927\u5b66\u7684\u53e6\u4e00\u4e2a\u7814\u7a76\u5c0f\u7ec4\u5f00\u53d1\u4e86\u4e00\u79cd\u53ef\u4ee5\u8bca\u65ad\u5fc3\u5f8b\u4e0d\u9f50\u7684\u6a21\u578b\uff0c\u53ef\u4ee5\u6bd4\u5fc3\u810f\u75c5\u4e13\u5bb6\u66f4\u597d\u5730\u4ece\u5355\u5bfc\u8054\u5fc3\u7535\u56fe\u4fe1\u53f7\u4e2d\u8bca\u65ad\u51fa\u5fc3\u5f8b\u5931\u5e38\u3002<\/p>\n


      <\/p>\n

      \n
      \"\u96c6\u667a\"<\/section>\n<\/section>\n


      <\/p>\n

      \u4f46\u662f\u4eca\u5e74\u5e76\u975e\u6ca1\u6709\u9057\u61be\u3002 Deepmind \u4e0e\u82f1\u56fd NHS \u7684\u6570\u636e\u5171\u4eab\u534f\u8bae\u5145\u6ee1\u4e86\u201c\u4e0d\u53ef\u539f\u8c05\u7684\u201d\u9519\u8bef\u3002\u7f8e\u56fd\u56fd\u5bb6\u536b\u751f\u7814\u7a76\u9662\u5411\u79d1\u5b66\u754c\u516c\u5e03\u4e86\u4e00\u4e2a\u80f8\u90e8 x \u5c04\u7ebf\u6570\u636e\u96c6, \u4f46\u66f4\u7ec6\u81f4\u7684\u7814\u7a76\u53d1\u73b0\uff0c\u5b83\u5e76\u4e0d\u771f\u6b63\u9002\u5408\u8bad\u7ec3\u7528\u4e8e\u75be\u75c5\u8bca\u65ad\u7684 AI \u6a21\u578b\u3002<\/p>\n


      <\/p>\n

      \n
      \n
      \n
      <\/p>\n

      \u5e94\u7528\uff1aArt & GANs<\/span><\/strong><\/span><\/p>\n

      <\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      \u4eca\u5e74\u5f00\u59cb\u8d8a\u6765\u8d8a\u53d7\u5173\u6ce8\u7684\u5e94\u7528\u8fd8\u6709\u56fe\u7247\u3001\u97f3\u4e50\u3001\u7ed8\u753b\u548c\u89c6\u9891\u7684\u751f\u6210\u6a21\u578b\u3002 2017\u5e74\u7684 NIPS \u4f1a\u8bae\u9996\u6b21\u4e3e\u529e\u4e86\u673a\u5668\u5b66\u4e60\u521b\u9020\u529b\u4e0e\u8bbe\u8ba1\u7814\u8ba8\u4f1a<\/strong>(Machine Learning for Creativity and Design workshop)\u3002<\/p>\n


      <\/p>\n

      \u8fd9\u7c7b\u5e94\u7528\u7a0b\u5e8f\u4e2d\u6700\u6d41\u884c\u7684\u662f Google \u7684 QuickDraw\uff0c\u5b83\u4f7f\u7528\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u6765\u8bc6\u522b\u4f60\u7684\u6d82\u9e26\u662f\u4ec0\u4e48\u3002\u4f7f\u7528\u516c\u5f00\u6570\u636e\u96c6\uff0c\u4f60\u751a\u81f3\u53ef\u4ee5\u8ba9\u673a\u5668\u4e3a\u4f60\u753b\u753b\u3002<\/p>\n


      <\/p>\n

      \u751f\u6210\u5f0f\u5bf9\u6297\u7f51\u7edc(GANs)\u4eca\u5e74\u53d6\u5f97\u4e86\u91cd\u5927\u8fdb\u5c55\u3002\u4f8b\u5982\uff0cCycleGAN\u3001DiscoGAN \u548c StarGAN \u7b49\u5728\u4eba\u8138\u6a21\u578b\u65b9\u9762\u53d6\u5f97\u4e86\u4ee4\u4eba\u77a9\u76ee\u7684\u6210\u679c\u3002\u4f20\u7edf\u4e0a\uff0cGANs \u5f88\u96be\u751f\u6210\u771f\u5b9e\u7684\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\uff0c\u4f46 pix2pixHD \u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u7ed3\u679c\u8868\u660e\uff0c\u6211\u4eec\u6b63\u5728\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\u3002\u6240\u4ee5 GANs \u4f1a\u6210\u4e3a\u65b0\u7684\u753b\u7b14\u5417\uff1f<\/p>\n


      <\/p>\n

      \n
      \"\u96c6\u667a\"<\/section>\n<\/section>\n


      <\/p>\n


      <\/p>\n

      \n
      \n
      \n
      <\/p>\n

      \u5e94\u7528: \u81ea\u52a8\u9a7e\u9a76\u6c7d\u8f66<\/span><\/strong><\/span><\/p>\n

      <\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      \u81ea\u52a8\u9a7e\u9a76\u6c7d\u8f66\u7684\u5927\u73a9\u5bb6\u662f\u5171\u4eab\u6c7d\u8f66\u516c\u53f8 Uber \u548c Lyft\uff0c\u4ee5\u53ca Alphabet \u7684 Waymo \u548c Tesla\u3002Uber \u5728\u4eca\u5e74\u5e74\u521d\u906d\u9047\u4e00\u4e9b\u632b\u6298\uff0c\u4ed6\u4eec\u7684\u81ea\u52a8\u9a7e\u9a76\u6c7d\u8f66\u5728\u65e7\u91d1\u5c71\u95ef\u4e86\u51e0\u6b21\u7ea2\u706f\uff0c\u539f\u56e0\u662f\u8f6f\u4ef6\u9519\u8bef<\/span>\uff0c\u800c\u975e\u4e4b\u524d\u62a5\u9053\u6240\u8bf4\u7684\u4eba\u4e3a\u5931\u8bef\u3002\u4e8b\u540e Uber \u5206\u4eab\u4e86\u5176\u5185\u90e8\u4f7f\u7528\u7684\u6c7d\u8f66\u53ef\u89c6\u5316\u5e73\u53f0\u7684\u8be6\u7ec6\u4fe1\u606f\u3002\u523012\u6708\uff0cUber \u7684\u81ea\u52a8\u9a7e\u9a76\u6c7d\u8f66\u9879\u76ee\u5df2\u7ecf\u8fbe\u5230\u4e86200\u4e07\u82f1\u91cc\u7684\u91cc\u7a0b\u3002<\/p>\n


      <\/p>\n

      \u4e0e\u6b64\u540c\u65f6\uff0cWaymo \u7684\u81ea\u52a8\u9a7e\u9a76\u6c7d\u8f66\u5728\u56db\u6708\u5f00\u59cb\u4e86\u4eba\u7c7b\u9a7e\u9a76\u5458\u7684\u7b2c\u4e00\u6b21\u8bd5\u884c\uff0c\u540e\u6765\u5728\u4e9a\u5229\u6851\u90a3\u5dde\u7684\u51e4\u51f0\u57ce\u8bd5\u884c\u65f6\uff0c\u5219\u5b8c\u5168\u6ca1\u6709\u4e86\u4eba\u7c7b\u9a7e\u9a76\u5458\u3002Waymo <\/span>\u8fd8\u516c\u5e03\u4e86\u4ed6\u4eec\u7684\u6d4b\u8bd5\u548c\u6a21\u62df\u6280\u672f\u7684\u7ec6\u8282\u3002<\/p>\n


      <\/p>\n

      \n
      \"\u96c6\u667a\"<\/section>\n<\/section>\n


      <\/p>\n

      Lyft \u5ba3\u5e03\u6b63\u5728\u5f00\u53d1\u81ea\u52a8\u9a7e\u9a76\u7684\u786c\u4ef6\u548c\u8f6f\u4ef6\uff0c\u5b83\u5728\u6ce2\u58eb\u987f\u7684\u7b2c\u4e00\u4e2a\u8bd5\u9a8c\u70b9\u5df2\u7ecf\u542f\u7528\u3002\u7279\u65af\u62c9\u7684 Autpilot \u6ca1\u6709\u770b\u5230\u592a\u591a\u7684\u66f4\u65b0\uff0c\u4f46\u662f\u8fd9\u4e2a\u9886\u57df\u53c8\u52a0\u5165\u4e86\u4e00\u4e2a\u65b0\u4eba\uff1aApple\u3002\u8482\u59c6\u2022\u5e93\u514b(Tim Cook)\u8bc1\u5b9e\uff0cApple \u516c\u53f8\u6b63\u81f4\u529b\u4e8e\u5f00\u53d1\u81ea\u52a8\u9a7e\u9a76\u6c7d\u8f66\u7684\u8f6f\u4ef6\uff0cApple \u7684\u7814\u7a76\u4eba\u5458\u5df2\u7ecf\u5728 arXiv \u53d1\u8868\u4e86\u4e00\u7bc7\u5173\u4e8e\u4e0e\u5730\u56fe\u76f8\u5173\u7684\u8bba\u6587\u3002<\/p>\n


      <\/p>\n

      \n
      \n
      \n
      <\/p>\n

      \u6df1\u5ea6\u5b66\u4e60\u3001\u53ef\u91cd\u590d\u6027\u548c\u70bc\u91d1\u672f<\/span><\/strong><\/span><\/p>\n

      <\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      \u5728\u8fc7\u53bb\u4e00\u6574\u5e74\uff0c\u4e00\u4e9b\u7814\u7a76\u4eba\u5458\u5bf9\u5b66\u672f\u8bba\u6587\u7ed3\u679c\u7684\u53ef\u91cd\u590d\u6027\u63d0\u51fa\u4e86\u62c5\u5fe7\u3002\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u901a\u5e38\u4f9d\u8d56\u5927\u91cf\u7684\u8d85\u53c2\u6570\uff0c\u5fc5\u987b\u4f18\u5316\u8fd9\u4e9b\u53c2\u6570\uff0c\u624d\u80fd\u83b7\u5f97\u8db3\u591f\u597d\u7684\u7ed3\u679c\u5e76\u53d1\u8868\u3002\u8fd9\u79cd\u53c2\u6570\u4f18\u5316\u7684\u6210\u672c\u53ef\u80fd\u8d8a\u6765\u8d8a\u9ad8\uff0c\u4ee5\u81f3\u4e8e\u53ea\u6709 Google \u548c Facebook \u8fd9\u6837\u7684\u516c\u53f8\u624d\u80fd\u8d1f\u62c5\u5f97\u8d77\u3002\u7814\u7a76\u4eba\u5458\u5e76\u4e0d\u603b\u4f1a\u516c\u5f00\u4ed6\u4eec\u7684\u4ee3\u7801\uff0c\u6709\u65f6\u4ed6\u4eec\u4f1a\u5fd8\u8bb0\u628a\u91cd\u8981\u7684\u7ec6\u8282\u5199\u8fdb\u8bba\u6587\uff0c\u6216\u8005\u4f7f\u7528\u7565\u5fae\u4e0d\u540c\u7684\u8bc4\u4f30\u6b65\u9aa4\uff0c\u6216\u8005\u901a\u8fc7\u5728\u540c\u4e00\u4e2a\u5206\u88c2\u70b9\u4e0a\u53cd\u590d\u4f18\u5316\u8d85\u53c2\u6570\u6765\u9002\u5e94\u6570\u636e\u96c6\uff0c\u4ece\u800c\u51fa\u73b0\u8fc7\u62df\u5408\uff0c\u8fd9\u4e9b\u4f7f\u5f97\u53ef\u91cd\u590d\u6027\u6210\u4e3a\u4e00\u4e2a\u5927\u95ee\u9898\u3002\u5728\u5f3a\u5316\u5b66\u4e60\u7814\u7a76\u4e2d\uff0c\u7814\u7a76\u4eba\u5458\u53d1\u73b0\u540c\u6837\u7684\u7b97\u6cd5\u4ece\u4e0d\u540c\u7684\u4ee3\u7801\u5e93\u4e2d\u83b7\u5f97\u7684\u7ed3\u679c\u5927\u4e0d\u76f8\u540c\uff0c\u800c\u4e14\u5dee\u5f02\u5f88\u5927:<\/p>\n


      <\/p>\n

      \n
      \"\u96c6\u667a\"<\/section>\n<\/section>\n


      <\/p>\n

      \u5728\u300aGANs \u662f\u5426\u9020\u6210\u4e86\u4e0d\u516c\uff1f\u4e00\u9879\u5927\u578b\u7814\u7a76\u300b(Are GANs Created Equal? A Large-Scale Study)\u4e00\u6587\u4e2d\uff0c\u7814\u7a76\u4eba\u5458\u53d1\u73b0\u8c03\u6574\u826f\u597d\u7684 GAN \u5982\u679c\u4f7f\u7528\u6602\u8d35\u7684\u8d85\u53c2\u6570\u641c\u7d22\uff0c\u5c31\u53ef\u4ee5\u51fb\u8d25\u90a3\u4e9b\u81ea\u79f0\u6700\u4f18\u7684\u66f4\u590d\u6742\u7684\u65b9\u6cd5\u3002\u540c\u6837, \u5728\u300a\u795e\u7ecf\u8bed\u8a00\u6a21\u578b\u8bc4\u4ef7\u7684\u73b0\u72b6\u300b(On the State of the Art of Evaluation in Neural Language Models<\/span>)\u4e00\u6587\u4e2d\uff0c\u7814\u7a76\u4eba\u5458\u8868\u793a\uff0c\u7b80\u5355\u7684 LSTM \u67b6\u6784\uff0c\u5982\u679c\u7ecf\u8fc7\u9002\u5f53\u7684\u89c4\u8303\u548c\u8c03\u6574, \u4e5f\u53ef\u4ee5\u80dc\u8fc7\u6700\u65b0\u7684\u6a21\u578b\u3002<\/span><\/p>\n

      <\/span><\/p>\n


      <\/span><\/p>\n

      \u5728\u4e00\u6b21 NIPS \u6f14\u8bb2\u4e2d\uff0cAli Rahimi \u628a\u6700\u8fd1\u6df1\u5ea6\u5b66\u4e60\u65b9\u6cd5\u548c\u4e2d\u4e16\u7eaa\u7684\u70bc\u91d1\u672f\u505a\u4e86\u6bd4\u8f83\uff0c\u547c\u5401\u66f4\u4e25\u683c\u7684\u5b9e\u9a8c\u8bbe\u8ba1\uff0c\u5f15\u8d77\u7814\u7a76\u4eba\u5458\u7684\u5e7f\u6cdb\u5171\u9e23\u3002Yann LeCun \u7b2c\u4e8c\u5929\u7acb\u5373\u4f5c\u51fa\u56de\u5e94\uff0c\u8ba4\u4e3a\u8fd9\u662f\u5bf9\u7814\u7a76\u8005\u7684\u5192\u72af\u3002<\/span>
      <\/p>\n


      <\/p>\n

      \n
      \n
      \n
      <\/p>\n

      \u52a0\u62ff\u5927\u548c\u4e2d\u56fd\u7684\u4eba\u5de5\u667a\u80fd<\/span><\/strong><\/span><\/p>\n

      <\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      \u968f\u7740\u7f8e\u56fd\u79fb\u6c11\u653f\u7b56\u7684\u6536\u7d27\uff0c\u5bf9\u56fd\u9645\u4eba\u624d\u9700\u6c42\u5f88\u5927\u7684\u79d1\u6280\u516c\u53f8\u6b63\u8d8a\u6765\u8d8a\u591a\u5730\u5728\u6d77\u5916\u5f00\u8bbe\u529e\u516c\u5ba4\uff0c\u52a0\u62ff\u5927\u662f\u9996\u9009\u76ee\u7684\u5730\u3002Google \u5728\u591a\u4f26\u591a\u5f00\u4e86\u4e00\u95f4\u65b0\u7684\u529e\u516c\u5ba4\uff0cDeepMind  \u5728\u52a0\u62ff\u5927\u7684\u57c3\u5fb7\u8499\u987f\u4e5f\u5f00\u4e86\u4e00\u95f4\u65b0\u529e\u516c\u5ba4\uff0cFacebook AI \u7814\u7a76\u9662\u4e5f\u5728\u5411\u52a0\u62ff\u5927\u8499\u7279\u5229\u5c14\u6269\u5f20\u3002<\/p>\n


      <\/p>\n

      \u4e2d\u56fd\u662f\u53e6\u4e00\u4e2a\u53d7\u5230\u5e7f\u6cdb\u5173\u6ce8\u7684\u76ee\u7684\u5730\u3002\u7531\u4e8e\u4e2d\u56fd\u62e5\u6709\u5927\u91cf\u8d44\u91d1\u3001\u5e9e\u5927\u7684\u4eba\u624d\u5e93\u4ee5\u53ca\u968f\u65f6\u53ef\u4ee5\u83b7\u5f97\u7684\u653f\u5e9c\u6570\u636e\uff0c\u5728\u4eba\u5de5\u667a\u80fd\u5f00\u53d1\u548c\u751f\u4ea7\u90e8\u7f72\u65b9\u9762\uff0c\u5b83\u5df2\u7ecf\u4e0e\u7f8e\u56fd\u5c55\u5f00\u4e86\u6b63\u9762\u7ade\u4e89\u3002Google \u8fd8\u5ba3\u5e03\u4e0d\u4e45\u5c06\u5728\u5317\u4eac\u5f00\u8bbe\u4e00\u4e2a\u65b0\u7684\u5b9e\u9a8c\u5ba4\u3002<\/p>\n


      <\/p>\n

      \"\u96c6\u667a\"<\/p>\n


      <\/p>\n

      \n
      \n
      \n
      <\/p>\n

      \u786c\u4ef6\u6218\u4e89\uff1aNvidia\uff0cIntel\uff0cGoogle\uff0cTesla<\/span><\/strong><\/span><\/p>\n

      <\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      \u4f17\u6240\u5468\u77e5\uff0c\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u9700\u8981\u6602\u8d35\u7684 GPU \u6765\u8bad\u7ec3\u6700\u5148\u8fdb\u7684\u6a21\u578b\u3002\u622a\u81f3\u76ee\u524d\uff0cNVIDIA \u662f\u6700\u5927\u8d62\u5bb6\u3002\u4eca\u5e74\uff0c\u5b83\u516c\u5e03\u4e86\u6cf0\u57665\u53f7\u65d7\u8230 GPU\u3002\u987a\u4fbf\u8bf4\u4e00\u53e5, \u5b83\u662f\u91d1\u8272\u7684\u3002<\/p>\n


      <\/p>\n

      \"\u96c6\u667a\"<\/p>\n


      <\/p>\n

      \u4f46\u7ade\u4e89\u6b63\u5728\u52a0\u5267\u3002Google \u7684 TPUs \u73b0\u5728\u53ef\u4ee5\u5728\u5176\u4e91\u5e73\u53f0\u4e0a\u4f7f\u7528\uff0c\u82f1\u7279\u5c14\u4e0b\u5c5e Nervana \u516c\u53f8\u63a8\u51fa\u4e86\u4e00\u5957\u65b0\u7684\u82af\u7247, \u751a\u81f3 Tesla \u4e5f\u627f\u8ba4\u5b83\u6b63\u5728\u5f00\u53d1\u81ea\u5df1\u7684\u4eba\u5de5\u667a\u80fd\u786c\u4ef6\u3002\u7ade\u4e89\u4e5f\u53ef\u80fd\u6765\u81ea\u4e2d\u56fd\uff0c\u90a3\u91cc\u4e13\u95e8\u4ece\u4e8b\u6bd4\u7279\u5e01\u91c7\u77ff\u7684\u786c\u4ef6\u5236\u9020\u5546\uff0c\u4e5f\u5e0c\u671b\u8fdb\u5165 GPU \u9886\u57df\u3002<\/p>\n

      \n
      \n
      \n
      <\/p>\n

      \u7092\u4f5c\u548c\u5931\u8d25<\/span><\/strong><\/span><\/p>\n

      <\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      \u8fc7\u5ea6\u7092\u4f5c\u8981\u62c5\u8d23\u3002\u4e3b\u6d41\u5a92\u4f53\u7684\u62a5\u9053\u51e0\u4e4e\u4ece\u4e0d\u7b26\u5408\u5b9e\u9a8c\u5ba4\u6216\u5de5\u4e1a\u754c\u7684\u5b9e\u9645\u60c5\u51b5\u3002 IBM Watson \u5c31\u662f\u4e00\u4e2a\u8fc7\u5ea6\u8425\u9500\u7684\u5178\u578b\u4ee3\u8868\u30022017\u5e74\uff0c\u6bcf\u4e2a\u4eba\u90fd\u8ba8\u538c IBM Watson\uff0c\u8fd9\u5e76\u4e0d\u5947\u602a\uff0c\u56e0\u4e3a\u5b83\u88ab\u5ba3\u4f20\u5f97\u5f88\u5389\u5bb3\uff0c\u4f46\u5728\u533b\u7597\u65b9\u9762\u5374\u5c61\u6b21\u5931\u8d25\u3002<\/p>\n


      <\/p>\n

      \"\u96c6\u667a\"<\/p>\n


      <\/p>\n

      \u6700\u5f15\u4eba\u6ce8\u76ee\u7684\u53ef\u80fd\u662f Facebook \u7684\u201c\u7814\u7a76\u4eba\u5458\u5173\u95ed\u4e86\u53d1\u660e\u81ea\u5df1\u8bed\u8a00\u7684\u4eba\u5de5\u667a\u80fd\u201d\u8fd9\u4ef6\u4e8b\uff0c\u6211\u5e76\u4e0d\u4e86\u89e3\u5a92\u4f53\u8fd9\u6837\u62a5\u9053\u7684\u771f\u5b9e\u610f\u56fe\u662f\u4ec0\u4e48\u3002\u4f46\u5b83\u5df2\u7ecf\u9020\u6210\u5f88\u5927\u7684\u8d1f\u9762\u5f71\u54cd\uff0c\u4f60\u751a\u81f3\u53ef\u4ee5\u5728 Google \u4e0a\u641c\u7d22\u5230\u76f8\u5173\u62a5\u9053\u3002\u5f53\u7136\u4e86\uff0c\u8fd9\u4e2a\u6807\u9898\u4e0e\u4e8b\u5b9e\u76f8\u53bb\u751a\u8fdc\u3002\u5b9e\u9645\u60c5\u51b5\u53ea\u662f\u7814\u7a76\u8005\u505c\u6b62\u4e86\u4e00\u9879\u7ed3\u679c\u4e0d\u7406\u60f3\u7684\u5b9e\u9a8c\u3002<\/p>\n


      <\/p>\n

      \u4f46\u662f\uff0c\u5bf9\u7092\u4f5c\u8d1f\u6709\u8d23\u4efb\u7684\u4e0d\u4ec5\u4ec5\u662f\u5a92\u4f53\u3002\u786e\u5b9e\u4e5f\u6709\u4e00\u90e8\u5206\u7814\u7a76\u8005\u7a81\u7834\u4e86\u5e95\u7ebf\uff0c\u5728\u8bba\u6587\u4e2d\u4f7f\u7528\u4e0d\u80fd\u53cd\u6620\u5b9e\u9645\u5b9e\u9a8c\u7ed3\u679c\u7684\u6807\u9898\u548c\u6458\u8981\uff0c\u6bd4\u5982\u8fd9\u7bc7\u81ea\u7136\u8bed\u8a00\u7814\u7a76\u8bba\u6587(https:\/\/medium.com\/@yoav.goldberg\/an-adversarial-review-of-adversarial-generation-of-natural-language-409ac3378bd7<\/span>)\uff0c\u548c\u8fd9\u7bc7\u673a\u5668\u5b66\u4e60\u7814\u7a76\u8bba\u6587(http:\/\/zacharydavid.com\/2017\/08\/06\/fitting-to-noise-or-nothing-at-all-machine-learning-in-markets\/<\/span>)\u3002<\/p>\n


      <\/p>\n

      \n
      \n
      \n
      <\/p>\n

      \u77e5\u540d\u7684\u8058\u4efb\u548c\u79bb\u804c<\/span><\/strong><\/span><\/p>\n

      <\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      Coursera \u7684\u8054\u5408\u521b\u59cb\u4eba Andrew Ng\uff08\u5434\u6069\u8fbe\uff09\u4eca\u5e74\u591a\u6b21\u51fa\u73b0\u5728\u65b0\u95fb\u62a5\u9053\u4e2d\uff0c\u4ed6\u56e0\u4e3a\u673a\u5668\u5b66\u4e60 MOOC \u8bfe\u7a0b\u89c6\u9891\u800c\u51fa\u540d\u3002\u4eca\u5e743\u6708\uff0cAndrew Ng \u79bb\u5f00\u4e86\u7531\u81ea\u5df1\u62c5\u4efb AI \u56e2\u961f\u9886\u5bfc\u8005\u7684\u767e\u5ea6\u516c\u53f8\uff0c\u7b79\u96c61.5\u4ebf\u7f8e\u5143\u8d44\u91d1\uff0c\u6210\u7acb\u4e86\u540d\u4e3a landing.AI \u7684\u521b\u4e1a\u516c\u53f8\uff0c\u4e13\u6ce8\u4e8e\u5236\u9020\u4e1a\u3002\u5176\u4ed6\u8fd9\u7c7b\u65b0\u95fb\u8fd8\u6709\uff1aGary Marcus \u8f9e\u53bb\u4e86 Uber \u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4\u7684\u4e3b\u7ba1\u804c\u52a1\uff0cFacebook \u96c7\u4f63\u4e86 Siri \u7684\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u4e3b\u7ba1\uff0c\u51e0\u4f4d\u6770\u51fa\u7684\u7814\u7a76\u4eba\u5458\u79bb\u5f00\u4e86 OpenAI\uff0c\u5f00\u529e\u4e86\u4e00\u5bb6\u65b0\u7684\u673a\u5668\u4eba\u516c\u53f8\u3002<\/p>\n


      <\/p>\n

      \"\u96c6\u667a\"<\/p>\n


      <\/p>\n

      \u5b66\u672f\u754c\u6d41\u5931\u79d1\u5b66\u5bb6\u7684\u8d8b\u52bf\u4e5f\u5728\u7ee7\u7eed\u3002\u5927\u5b66\u5b9e\u9a8c\u5ba4\u7eb7\u7eb7\u62b1\u6028\uff1a\u4ed6\u4eec\u80fd\u4e3a AI \u4eba\u624d\u63d0\u4f9b\u7684\u85aa\u6c34\uff0c\u65e0\u6cd5\u4e0e\u5de5\u4e1a\u5de8\u5934\u7ade\u4e89\u3002<\/p>\n


      <\/strong><\/p>\n

      \u4ee5\u4e0b\u4e3a\u82f1\u6587\u539f\u6587\uff0c\u70b9\u51fb\u9605\u8bfb\u539f\u6587\u4e5f\u53ef\u67e5\u770b\u3002<\/strong><\/span><\/p>\n

      <\/span><\/p>\n

      \n
      \n

      AI and Deep Learning in 2017 \u2013 A Year in Review<\/h1>\n<\/header>\n

      The year is coming to an end. I did not write nearly as much as I had planned to. But I\u2019m hoping to change that next year, with more tutorials around Reinforcement Learning, Evolution, and Bayesian Methods coming to WildML! And what better way to start than with a summary of all the amazing things that happened in 2017? Looking back through my Twitter history and the WildML newsletter, the following topics repeatedly came up. I\u2019ll inevitably miss some important milestones, so please let me know about it in the comments!<\/p>\n

      Reinforcement Learning beats humans at their own games<\/h3>\n

      The biggest success story of the year was probably AlphaGo (Nature paper), a Reinforcement Learning agent that beat the world\u2019s best Go players. Due to its extremely large search space, Go was thought to be out of reach of Machine Learning techniques for a couple more years. What a nice surprise!<\/p>\n

      The first version of AlphaGo was bootstrapped using training data from human experts and further improved through self-play and an adaptation of Monte-Carlo Tree Search. Soon after, AlphaGo Zero (Nature Paper) took it a step further and learned to play Go from scratch, without human training data whatsoever, using a technique simultaneously published in the Thinking Fast and Slow with Deep Learning and Tree Search paper. It also handily beat the first version of AlphaGo. Towards the end of the year, we saw yet another generalization of the AlphaGo Zero algorithm, called AlphaZero, which not only mastered Go, but also Chess and Shogi, using the exact same techniques. Interestingly, these programs made moves that surprised even the most experienced Go players, motivating players to learn from AlphaGo and adjusting their own play style accordingly. To make this easier, DeepMind also released an AlphaGo Teach tool.<\/p>\n

      \"\u96c6\u667a\"But Go wasn\u2019t the only game where we made significant progress. Libratus (Science paper), a system developed by researchers from CMU, managed to beat top Poker players in a 20-day Heads-up, No-Limit Texas Hold\u2019em tournament. A little earlier, DeepStack, a system developed by researchers from Charles University, The Czech Technical University, and the University of Alberta, became the first to beat professional poker players. Note that both of these systems played Heads-up poker, which is played between two players and a significantly easier problem than playing at a table of multiple players. The latter will most likely see additional progress in 2018.<\/p>\n

      The next frontiers for Reinforcement Learning seem to be more complex multi-player games, including multi-player Poker. DeepMind is actively working on Starcraft 2, releasing a research environment, and OpenAI demonstrated initial success in 1v1 Dota 2, with the goal of competing in the the full 5v5 game in the near future.<\/p>\n

      \"\u96c6\u667a\"<\/p>\n

      Evolution Algorithms make a Comeback<\/span><\/p>\n

      For supervised learning, gradient-based approaches using the back-propagation algorithm have been working extremely well. And that isn\u2019t likely to change anytime soon. However, in Reinforcement Learning, Evolution Strategies (ES) seem to be making a comeback. Because the data typically is not iid (independent and identically distributed), error signals are sparser, and because there is a need for exploration, algorithms that do not rely on gradients can work quite well. In addition, evolutionary algorithms can scale linearly to thousands of machines enabling extremely fast parallel training. They do not require expensive GPUs, but can be trained on a large number (typically hundreds to thousands) of cheap CPUs.<\/p>\n

      Earlier in the year, researchers from OpenAI demonstrated that Evolution Strategies can achieve performance comparable to standard Reinforcement Learning algorithms such as Deep Q-Learning. Towards the end of the year, a team from Uber released a blog post and a set of five research papers, further demonstrating the potential of Genetic Algorithms and novelty search. Using an extremely simple Genetic Algorithm, and no gradient information whatsoever, their algorithm learns to play difficult Atari Games. Here\u2019s a video of the GA policy scoreing 10,500 on Frostbite. DQN, AC3, and ES score less than 1,000 on this game.\"\u96c6\u667a\"<\/p>\n

      Most likely, we\u2019ll see more work in this direction in 2018.<\/p>\n

      WaveNets, CNNs, and Attention Mechanisms<\/h3>\n

      Google\u2019s Tacotron 2 text-to-speech system produces extremely impressive audio samples and is based on WaveNet, an autoregressive model which is also deployed in the Google Assistant and has seen massive speed improvements in the past year. WaveNet had previously been applied to Machine Translation as well, resulting in faster training times that recurrent architectures.<\/p>\n

      The move away from expensive recurrent architectures that take long to train seems to be larger trend in Machine Learning subfields. In Attention is All you Need, researchers get rid of recurrence and convolutions entirely, and use a more sophisticated attention mechanism to achieve state of the art results at a fraction of the training costs.<\/p>\n

      The Year of Deep Learning frameworks<\/h3>\n

      If I had to summarize 2017 in one sentence, it would be the year of frameworks. Facebook made a big splash with PyTorch. Due to its dynamic graph construction similar to what Chainer offers, PyTorch received much love from researchers in Natural Language Processing, who regularly have to deal with dynamic and recurrent structures that hard to declare in a static graph frameworks such as Tensorflow.<\/p>\n

      Tensorflow had quite a run in 2017. Tensorflow 1.0 with a stable and backwards-compatible API was released in February. Currently, Tensorflow is at version 1.4.1. In addition to the main framework, several Tensorflow companion libraries were released, including Tensorflow Fold for dynamic computation graphs, Tensorflow Transform for data input pipelines, and DeepMind\u2019s higher-level Sonnet library. The Tensorflow team also announced a new eager execution mode which works similar to PyTorch\u2019s dynamic computation graphs.<\/p>\n

      In addition to Google and Facebook, many other companies jumped on the Machine Learning framework bandwagon:<\/p>\n