{"id":63575,"date":"2025-08-18T20:39:25","date_gmt":"2025-08-18T12:39:25","guid":{"rendered":"https:\/\/swarma.org\/?p=63575"},"modified":"2025-08-18T20:39:25","modified_gmt":"2025-08-18T12:39:25","slug":"%e4%bb%8e%e5%8e%86%e5%8f%b2%e3%80%81%e5%8e%9f%e7%90%86%e5%88%b0%e6%9c%aa%e6%9d%a5%e8%b6%8b%e5%8a%bf%ef%bc%8c%e4%b8%80%e7%af%87%e6%96%87%e7%ab%a0%e8%ae%b2%e6%b8%85%e6%a5%9a%e6%b7%b1%e5%ba%a6%e5%ad%a6","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=63575","title":{"rendered":"\u4ece\u5386\u53f2\u3001\u539f\u7406\u5230\u672a\u6765\u8d8b\u52bf\uff0c\u4e00\u7bc7\u6587\u7ae0\u8bb2\u6e05\u695a\u6df1\u5ea6\u5b66\u4e60\u7684\u5e95\u5c42\u903b\u8f91"},"content":{"rendered":"
\n
<\/section>\n
\n

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

\u5bfc\u8bed<\/span><\/span>
<\/span><\/strong><\/span><\/p>\n

\n
\n
\n
\n

\u6df1\u5ea6\u5b66\u4e60\u5e76\u4e0d\u662f\u4e00\u591c\u4e4b\u95f4\u201c\u706b\u8d77\u6765\u201d\u7684\u9b54\u6cd5\uff0c\u800c\u662f\u82e5\u5e72\u8109\u7edc\u4ea4\u7ec7\u3001\u957f\u671f\u79ef\u7d2f\u7684\u7ed3\u679c\uff1a\u7406\u8bba\u4e0a\u7684\u53cd\u5411\u4f20\u64ad\u3001\u6570\u636e\u89c4\u6a21\u7684\u7206\u53d1\u3001\u786c\u4ef6\u52a0\u901f\uff08GPU\uff09\u4e0e\u67b6\u6784\u521b\u65b0\u5171\u540c\u63a8\u52a8\u4e86\u8fd9\u573a\u9769\u547d\u3002\u672c\u7bc7\u6587\u7ae0\u9762\u5411\u60f3\u771f\u6b63\u7406\u89e3\u201c\u539f\u7406\u201d\u7684\u8bfb\u8005\uff1a\u6211\u4eec\u4e0d\u8bf4\u4ee3\u7801\u548c\u6570\u5b66\u516c\u5f0f\uff0c\u4f46\u4f1a\u6df1\u5165\u8bb2\u6e05\u695a\u6bcf\u4e00\u6b65\u4e3a\u4f55\u91cd\u8981\u3001\u5982\u4f55\u76f8\u4e92\u4f5c\u7528\uff0c\u5e76\u6307\u51fa PyTorch \u5728\u7814\u7a76\u4e0e\u5b9e\u8df5\u4e2d\u7684\u5173\u952e\u89d2\u8272\u3002\u672c\u6587\u65e2\u9002\u5408\u5bf9\u6df1\u5ea6\u5b66\u4e60\u5df2\u6709\u521d\u6b65\u4e86\u89e3\u7684\u5de5\u7a0b\u5e08\uff0c\u4e5f\u9002\u5408\u5e0c\u671b\u628a\u6982\u5ff5\u542c\u61c2\u518d\u8fdb\u5165\u5b9e\u64cd\u7684\u79d1\u7814\u4eba\u5458\u4e0e\u5b66\u751f\u3002<\/span><\/span><\/span><\/span><\/span><\/strong><\/span><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/span><\/section>\n
\u5218\u6613\u660e | \u4f5c\u8005<\/span><\/span><\/span><\/section>\n
\u8d75\u601d\u6021 | \u7f16\u8f91<\/span><\/span><\/span><\/section>\n<\/section>\n
\n
\n

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

\u76ee\u5f55<\/span><\/span>
<\/span><\/strong><\/span><\/p>\n

\n
\n
\n
\n
\n

1.\u4eba\u5de5\u667a\u80fd\u4e0e\u6df1\u5ea6\u5b66\u4e60\u53d1\u5c55\u7b80\u53f2<\/span><\/p>\n

\n
\n
\n

  2.\u4e3a\u4ec0\u4e48\u6df1\u5ea6\u5b66\u4e60\u5982\u6b64\u5f3a\u5927\uff08\u67b6\u6784\u3001\u6570\u636e\u3001\u7b97\u529b\uff09<\/span><\/p>\n

  3.\u795e\u7ecf\u7f51\u7edc\u7684\u4e09\u5927\u5173\u952e\u8981\u7d20\uff08\u67b6\u6784 \/ \u8d85\u53c2\u6570 \/ \u8bad\u7ec3\u65b9\u6cd5\uff09<\/span><\/p>\n

  4.PyTorch\uff1a\u4e3a\u4ec0\u4e48\u5728\u7814\u7a76\/\u6559\u5b66\u4e2d\u5982\u6b64\u53d7\u6b22\u8fce\uff08\u52a8\u6001\u8ba1\u7b97\u56fe \/ \u5f20\u91cf \/ \u81ea\u52a8\u5fae\u5206\uff09<\/span><\/p>\n

  5.\u4ece\u7279\u5f81\u5b66\u4e60\u5230\u8fc1\u79fb\u5b66\u4e60\uff08\u591a\u5c3a\u5ea6\u7279\u5f81\u4e0e\u8fc1\u79fb\u7b56\u7565\uff09<\/span><\/p>\n

  6.\u5c55\u671b\u672a\u6765\uff1a\u6df1\u5ea6\u5b66\u4e60\u7684\u673a\u9047\u4e0e\u6311\u6218<\/span><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n


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

 <\/span><\/strong><\/span><\/p>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n

\n
\n
\n
1\u3001\u4eba\u5de5\u667a\u80fd\u4e0e\u6df1\u5ea6\u5b66\u4e60\u53d1\u5c55\u7b80\u53f2<\/span><\/span>
<\/span><\/strong><\/span><\/section>\n<\/section>\n


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


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


<\/span><\/p>\n<\/section>\n<\/h1>\n<\/section>\n<\/section>\n


<\/span><\/h2>\n
<\/section>\n

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

\n
\n
\n
1.1 \u65e9\u671f\u8109\u7edc\uff1a\u4ece\u4eba\u5de5\u795e\u7ecf\u5143\u5230\u611f\u77e5\u673a<\/span>
<\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


<\/span><\/h3>\n

\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u7684\u601d\u60f3\u53ef\u8ffd\u6eaf\u5230 1943 \u5e74 McCulloch \u4e0e Pitts \u63d0\u51fa\u7684\u201c\u4eba\u5de5\u795e\u7ecf\u5143\u201d\u6a21\u578b\uff08\u628a\u795e\u7ecf\u5143\u62bd\u8c61\u4e3a\u7b80\u5355\u7684\u9608\u503c\u5355\u5143\uff09\uff0c\u8fd9\u662f\u628a\u751f\u7269\u795e\u7ecf\u7cfb\u7edf\u7528\u6570\u5b66\u65b9\u5f0f\u6700\u65e9\u7684\u5c1d\u8bd5\u4e4b\u4e00\u3002\u968f\u540e\u51e0\u5341\u5e74\u5185\uff0c\u7814\u7a76\u65e2\u6709\u9ad8\u6f6e\u4e5f\u6709\u4f4e\u6f6e\u2014\u2014\u6bd4\u5982 1969 \u5e74 Minsky \u4e0e Papert \u5bf9\u5355\u5c42\u611f\u77e5\u673a\u7684\u5c40\u9650\u6027\u7684\u8ba8\u8bba\uff08\u5c24\u5176\u662f\u65e0\u6cd5\u89e3\u51b3 XOR \u95ee\u9898\uff09\u66fe\u5bfc\u81f4\u795e\u7ecf\u7f51\u7edc\u7814\u7a76\u4e00\u5ea6\u505c\u6ede\u3002<\/span><\/p>\n


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

\n
\n
\n
1.2 \u53cd\u5411\u4f20\u64ad\u7684\u51fa\u73b0\u4e0e\u590d\u5174<\/span>
<\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


<\/span><\/h3>\n

\u771f\u6b63\u628a\u591a\u5c42\u7f51\u7edc\u53d8\u4e3a\u73b0\u5b9e\u7684\u662f\u53cd\u5411\u4f20\u64ad\uff08backpropagation\uff09<\/span>\u601d\u60f3\u7684\u5e7f\u6cdb\u5e94\u7528\u3002\u53cd\u5411\u4f20\u64ad\u4f9d\u6258\u5fae\u79ef\u5206\uff08\u94fe\u5f0f\u6cd5\u5219\uff09\uff0c\u5141\u8bb8\u6211\u4eec\u628a\u6700\u7ec8\u8f93\u51fa\u8bef\u5dee\u201c\u53cd\u5411\u201d\u4f20\u5230\u6bcf\u4e00\u5c42\uff0c\u4ece\u800c\u7ed9\u6bcf\u4e2a\u53c2\u6570\u5206\u914d\u201c\u8d23\u4efb\u201d\uff08\u68af\u5ea6\uff09\uff0c\u8fd9\u4f7f\u5f97\u591a\u5c42\u7f51\u7edc\u7684\u8bad\u7ec3\u5728\u8ba1\u7b97\u4e0a\u53ef\u884c\u5e76\u9010\u6b65\u53d6\u5f97\u6210\u679c\u300220 \u4e16\u7eaa 80 \u5e74\u4ee3\u4ee5\u6765\uff0c\u968f\u7740\u7b97\u6cd5\u4e0e\u5b9e\u8df5\u7684\u79ef\u7d2f\uff0c\u795e\u7ecf\u7f51\u7edc\u91cd\u65b0\u83b7\u5f97\u4e86\u5b66\u754c\u7684\u5173\u6ce8\u3002<\/span><\/span><\/p>\n


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

\n
\n
\n
1.3 \u201c\u6df1\u5ea6\u201d\u4e0e Hinton \u7684\u5f71\u54cd<\/span>
<\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


<\/span><\/h3>\n

\u201c\u6df1\u5ea6\u201d\u4e00\u8bcd\u5728 2006 \u5e74\u540e\u6b63\u5f0f\u8fdb\u5165\u5b66\u672f\u6d41\u884c\u8bed\uff0cGeoffrey Hinton \u7b49\u4eba\u7684\u5de5\u4f5c\u8868\u660e\uff1a\u5c06\u7f51\u7edc\u505a\u6df1\uff08\u66f4\u591a\u5c42\uff09\u53ef\u4ee5\u5728\u67d0\u4e9b\u95ee\u9898\u4e0a\u663e\u8457\u63d0\u5347\u8868\u73b0\uff0c\u4f46\u8fd9\u9700\u8981\u914d\u5957\u7684\u8bad\u7ec3\u6280\u5de7\u4e0e\u5927\u91cf\u6570\u636e\u30022006 \u5e74\u5f00\u59cb\u7684\u8fd9\u4e00\u6ce2\u7406\u8bba\u4e0e\u5b9e\u9a8c\u7ed3\u5408\uff0c\u4e3a\u540e\u6765\u7684\u7a81\u7834\u5960\u5b9a\u4e86\u57fa\u7840\u3002<\/span><\/p>\n


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

\n
\n
\n
1.4 ImageNet\uff1a\u6570\u636e\u9a71\u52a8\u7684\u5173\u952e\u8f6c\u6298<\/span>
<\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


<\/span><\/h3>\n

\u6df1\u5ea6\u5b66\u4e60\u771f\u6b63\u8fdb\u5165\u5927\u4f17\u89c6\u7ebf\uff0c\u662f\u5728\u6709\u4e86\u8db3\u591f\u5927\u89c4\u6a21\u3001\u6807\u6ce8\u826f\u597d\u7684\u6570\u636e\u96c6\u4e4b\u540e\u3002\u674e\u98de\u98de\u7b49\u4eba\u53d1\u8d77\u5e76\u6784\u5efa\u7684 ImageNet\uff0c\u63d0\u4f9b\u4e86\u6d77\u91cf\u7684\u6709\u6807\u6ce8\u56fe\u7247\uff0c\u4f7f\u5f97\u201c\u5b66\u5230\u590d\u6742\u89c6\u89c9\u7279\u5f81\u201d\u6210\u4e3a\u53ef\u80fd\u3002\u968f\u540e AlexNet \u5728 ImageNet \u6bd4\u8d5b\u4e2d\u7684\u5de8\u5927\u8dc3\u5347\uff08\u76f8\u6bd4\u5f53\u65f6\u7ecf\u9a8c\u65b9\u6cd5\u63d0\u9ad8\u4e86\u663e\u8457\u7684\u7cbe\u5ea6\uff09\u76f4\u63a5\u89e6\u53d1\u4e86\u5de5\u4e1a\u754c\u4e0e\u5b66\u754c\u5bf9\u6df1\u5ea6\u7f51\u7edc\u7684\u6295\u5165\u6f6e\u6d41\u3002Li Fei-Fei \u4f7f\u7528\u4f17\u5305\uff08Amazon Mechanical Turk\uff09\u6765\u5b8c\u6210\u6807\u6ce8\u5de5\u4f5c\uff0c\u8fd9\u4e2a\u5b9e\u8df5\u8bf4\u660e\u4e86\u5927\u5de5\u7a0b\u5728\u7814\u7a76\u8fdb\u6b65\u4e2d\u7684\u91cd\u8981\u6027\u3002<\/span><\/p>\n


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

\n
\n
\n
1.5 \u4ece GPT-1 \u5230 GPT-5\uff1a\u6df1\u5ea6\u5b66\u4e60\u7684\u201c\u5927\u6a21\u578b\u65f6\u4ee3\u201d<\/span>
<\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


<\/span><\/strong><\/h3>\n

\u5982\u679c\u8bf4AlexNet \u5728\u89c6\u89c9\u9886\u57df\u5f15\u53d1\u4e86\u7b2c\u4e00\u6b21\u6df1\u5ea6\u5b66\u4e60\u6d6a\u6f6e\uff0c\u90a3\u4e48 GPT \u7cfb\u5217\u7684\u8bde\u751f\u5219\u5728\u8bed\u8a00\u9886\u57df\u6380\u8d77\u4e86\u7b2c\u4e8c\u6b21\u9769\u547d\u3002  2017 \u5e74 Transformer \u67b6\u6784\u95ee\u4e16\uff0c\u4e3a\u957f\u7a0b\u4f9d\u8d56\u5efa\u6a21\u548c\u5927\u89c4\u6a21\u5e76\u884c\u8bad\u7ec3\u5960\u5b9a\u57fa\u7840\u3002\u57fa\u4e8e\u6b64\uff0cOpenAI \u4e8e 2018 \u5e74\u63a8\u51fa GPT-1\uff0c\u9a8c\u8bc1\u4e86<\/span>\u5927\u89c4\u6a21\u9884\u8bad\u7ec3 + \u4e0b\u6e38\u9002\u914d<\/span><\/span><\/strong>\u7684\u53ef\u884c\u6027\uff1b\u968f\u540e GPT-2 \u4e0e GPT-3 \u4e0d\u65ad\u6269\u5c55\u53c2\u6570\u89c4\u6a21\u4e0e\u751f\u6210\u80fd\u529b\uff0c\u5c55\u73b0\u51fa\u60ca\u4eba\u7684\u96f6\u6837\u672c\u4e0e\u5c11\u6837\u672c\u5b66\u4e60\u80fd\u529b\u3002GPT-4 \u5219\u8fc8\u5411\u591a\u6a21\u6001\uff0c\u80fd\u540c\u65f6\u5904\u7406\u6587\u672c\u4e0e\u56fe\u50cf\uff0c\u8fdb\u4e00\u6b65\u62d3\u5c55\u4e86\u5e94\u7528\u8fb9\u754c\u3002\u521a\u521a\u53d1\u5e03\u7684 <\/span>GPT-5<\/span><\/span><\/strong> \u6574\u5408\u4e86\u8bed\u8a00\u3001\u63a8\u7406\u3001\u591a\u6a21\u6001\u5904\u7406\u4e8e\u4e00\u4f53\uff0c\u5177\u5907\u66f4\u5f3a\u7684\u903b\u8f91\u63a8\u7406\u4e0e\u6267\u884c\u80fd\u529b\uff0c\u652f\u6301\u8d85\u957f\u4e0a\u4e0b\u6587\uff08\u53ef\u8fbe 256K token\uff09\uff0c\u88ab OpenAI \u79f0\u4e3a “\u535a\u58eb\u7ea7\u4e13\u5bb6\u968f\u65f6\u5f85\u547d”\uff0c\u5b9a\u4f4d\u4e3a\u96c6\u6210\u8bed\u8a00\u3001\u63a8\u7406\u3001\u591a\u6a21\u6001\u7684\u7edf\u4e00\u667a\u80fd\u7cfb\u7edf\u3002<\/span><\/span><\/p>\n

\u53ef\u4ee5\u8bf4\uff0c<\/span>GPT \u4e4b\u540e\u7684\u6df1\u5ea6\u5b66\u4e60\u8fdb\u5165\u4e86\u201c\u5927\u6a21\u578b\u65f6\u4ee3\u201d<\/span><\/span><\/strong>\uff0c\u7814\u7a76\u548c\u5e94\u7528\u7684\u6838\u5fc3\u95ee\u9898\u4ece\u201c\u5982\u4f55\u8bbe\u8ba1\u7279\u5b9a\u4efb\u52a1\u6a21\u578b\u201d\u8f6c\u5411\u201c\u5982\u4f55\u9ad8\u6548\u8bad\u7ec3\u901a\u7528\u6a21\u578b\u201d\uff0c\u8fd9\u5bf9\u672a\u6765\u6df1\u5ea6\u5b66\u4e60\u7684\u7b97\u6cd5\u3001\u6570\u636e\u3001\u7b97\u529b\u548c\u67b6\u6784\u90fd\u63d0\u51fa\u4e86\u5168\u65b0\u7684\u6311\u6218\u3002<\/span><\/span><\/p>\n


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

\n
\n
\n
2\u3001\u4e3a\u4ec0\u4e48\u6df1\u5ea6\u5b66\u4e60\u5982\u6b64\u5f3a\u5927<\/span><\/span>
<\/span><\/strong><\/span><\/section>\n<\/section>\n


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


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


<\/span><\/p>\n<\/section>\n<\/h1>\n<\/section>\n<\/section>\n


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

<\/section>\n

\u6df1\u5ea6\u5b66\u4e60\u7684\u6d41\u884c\u5e76\u975e\u5355\u70b9\u539f\u56e0\uff0c\u800c\u662f<\/span>\u67b6\u6784\u521b\u65b0 + \u5927\u6570\u636e + \u8ba1\u7b97\u529b<\/span><\/span><\/strong>\u4e09\u8005\u534f\u540c\u7684\u7ed3\u679c\u3002\u4e0b\u9762\u9010\u6761\u62c6\u89e3\u201c\u4e3a\u4ec0\u4e48\u201d\u3002<\/span><\/span><\/p>\n


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

\n
\n
\n
2.1 \u67b6\u6784\u521b\u65b0\uff1a\u4e0d\u540c\u95ee\u9898\u7528\u4e0d\u540c\u201c\u5927\u8111\u7ed3\u6784\u201d<\/span>
<\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


<\/span><\/h3>\n

\u6df1\u5ea6\u5b66\u4e60\u5e76\u4e0d\u7b49\u4e8e\u5355\u4e00\u201c\u6df1\u5c42\u7f51\u7edc\u201d\u2014\u2014\u4e0d\u540c\u7684\u4efb\u52a1\u50ac\u751f\u4e86\u4e0d\u540c\u7684\u67b6\u6784\uff0c\u8fd9\u4e9b\u67b6\u6784\u5404\u81ea\u5229\u7528\u4e86\u6570\u636e\u7684\u7ed3\u6784\u5316\u4fe1\u606f\uff1a<\/span><\/p>\n

    \n
  • \n

    \u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09<\/span><\/span><\/strong>\uff1a\u8bbe\u8ba1\u4e4b\u521d\u5c31\u662f\u4e3a\u4e86\u89e3\u51b3\u56fe\u50cf\u4e2d\u5c40\u90e8\u6027\u4e0e\u5e73\u79fb\u4e0d\u53d8\u6027\u95ee\u9898\u3002\u5377\u79ef\u6838\uff08filter\uff09\u5728\u7a7a\u95f4\u4e0a\u6ed1\u52a8\uff0c\u53c2\u6570\u5171\u4eab\u4f7f\u5f97\u76f8\u540c\u7684\u7279\u5f81\u68c0\u6d4b\u5668\u80fd\u5728\u56fe\u50cf\u4e0d\u540c\u4f4d\u7f6e\u91cd\u590d\u4f7f\u7528\uff0c\u4ece\u800c\u663e\u8457\u51cf\u5c11\u53c2\u6570\u6570\u91cf\u5e76\u63d0\u9ad8\u6837\u672c\u6548\u7387\u3002\u5377\u79ef\u7684\u201c\u611f\u53d7\u91ce\u201d\u4e0e\u9010\u5c42\u5806\u53e0\u80fd\u4ece\u5c40\u90e8\u8fb9\u7f18\u9010\u7ea7\u62bd\u8c61\u81f3\u9ad8\u9636\u8bed\u4e49\uff08\u8fb9->\u7eb9\u7406->\u90e8\u5206->\u7269\u4f53\uff09\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48 CNN \u5728\u89c6\u89c9\u4efb\u52a1\u4e0a\u957f\u671f\u5360\u4f18\u3002<\/span><\/span><\/p>\n<\/li>\n

  • \n

    \u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09\u53ca\u5176\u53d8\u79cd<\/span><\/span><\/strong>\uff1a\u5904\u7406\u5e8f\u5217\u6570\u636e\uff08\u6587\u672c\u3001\u8bed\u97f3\u3001\u65f6\u95f4\u5e8f\u5217\uff09\u65f6\uff0c\u6570\u636e\u7684\u987a\u5e8f\u53ca\u4e0a\u4e0b\u6587\u5173\u7cfb\u6781\u4e3a\u91cd\u8981\u3002RNN \u7528\u9690\u85cf\u72b6\u6001\uff08hidden state\uff09\u643a\u5e26\u524d\u5e8f\u4fe1\u606f\uff0c\u7406\u8bba\u4e0a\u80fd\u591f\u5efa\u6a21\u4efb\u610f\u957f\u5ea6\u7684\u4f9d\u8d56\u3002\u4f46\u5728\u5b9e\u9645\u8bad\u7ec3\u4e2d\uff0c\u4f20\u7edf RNN \u4f1a\u906d\u9047\u68af\u5ea6\u8870\u51cf\/\u7206\u70b8\u95ee\u9898\uff0cLSTM \u548c GRU \u7b49\u95e8\u63a7\u7ed3\u6784\u88ab\u63d0\u51fa\u4ee5\u7f13\u89e3\u957f\u5e8f\u5217\u4f9d\u8d56\u7684\u5b66\u4e60\u95ee\u9898\u3002<\/span><\/span><\/p>\n<\/li>\n

  • \n

    \u6ce8\u610f\u529b\u673a\u5236\uff08Transformer\uff09<\/span><\/span><\/strong>\uff1a2017 \u5e74\u63d0\u51fa\u7684 Transformer \u7528\u201c\u81ea\u6ce8\u610f\u529b\uff08self-attention\uff09\u201d\u66ff\u4ee3\u5e8f\u5217\u4e2d\u7684\u5faa\u73af\u64cd\u4f5c\uff0c\u4ee4\u6bcf\u4e2a\u4f4d\u7f6e\u80fd\u591f\u76f4\u63a5\u201c\u5173\u6ce8\u201d\u5176\u4ed6\u4f4d\u7f6e\uff0c\u4ece\u800c\u9ad8\u6548\u6355\u6349\u957f\u7a0b\u4f9d\u8d56\uff1b\u5173\u952e\u4f18\u52bf\u5728\u4e8e\u80fd\u591f\u5e76\u884c\u8ba1\u7b97\u3001\u6269\u5c55\u6027\u5f3a\u3001\u5bf9\u5927\u89c4\u6a21\u8bed\u6599\u8bad\u7ec3\u5c24\u4e3a\u53cb\u597d\u3002Transformer \u6210\u4e3a NLP \u7684\u57fa\u7840\u7ec4\u4ef6\uff0c\u8fdb\u4e00\u6b65\u884d\u751f\u51fa BERT\u3001GPT \u7b49\u9884\u8bad\u7ec3\u67b6\u6784\u3002<\/span><\/span><\/p>\n<\/li>\n<\/ul>\n


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

    \n
    \n
    \n
    2.2 \u5927\u6570\u636e\u5582\u517b\u4e0b\u7684\u201c\u6050\u9f99\u6548\u5e94\u201d <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


    <\/span><\/h3>\n

    \u6df1\u5ea6\u7f51\u7edc\u50cf\u201c\u5de8\u578b\u6a21\u578b\u201d\uff0c\u9700\u8981\u5927\u91cf\u201c\u98df\u7269\u201d\uff08\u6570\u636e\uff09\u6765\u8bad\u7ec3\u51fa\u6cdb\u5316\u80fd\u529b\u3002\u6ca1\u6709\u8db3\u591f\u591a\u6837\u4e14\u4ee3\u8868\u6027\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u518d\u597d\u7684\u7f51\u7edc\u4e5f\u5bb9\u6613\u8fc7\u62df\u5408\u6216\u5b66\u4e0d\u5230\u7a33\u5065\u7684\u8bed\u4e49\u8868\u793a\u3002ImageNet \u7684\u5efa\u7acb\u548c\u968f\u540e\u7684\u5927\u89c4\u6a21\u8bed\u6599\uff08\u4f8b\u5982\u5927\u89c4\u6a21\u6587\u672c\u8bed\u6599\uff09\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u5b66\u4e60\u5230\u901a\u7528\u3001\u53ef\u8fc1\u79fb\u7684\u7279\u5f81\u8868\u5f81\uff0c\u8fd9\u6b63\u662f\u6df1\u5ea6\u5b66\u4e60\u771f\u6b63\u8d70\u5411\u5de5\u4e1a\u5e94\u7528\u7684\u5173\u952e\u4e00\u73af\u3002<\/span><\/p>\n


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

    \n
    \n
    \n
    2.3 GPU \u5e26\u6765\u7684\u8ba1\u7b97\u9769\u547d<\/span>
    <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


    <\/span><\/h3>\n

    \u6df1\u5ea6\u7f51\u7edc\u7684\u8bad\u7ec3\u4e3b\u8981\u662f\u5bc6\u96c6\u7684\u77e9\u9635\/\u5f20\u91cf\u8fd0\u7b97\uff08\u7ebf\u6027\u4ee3\u6570\uff09\uff0cGPU \u6700\u521d\u7528\u4e8e\u56fe\u5f62\u6e32\u67d3\u7684\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\u6070\u597d\u4e0e\u6b64\u9ad8\u5ea6\u5951\u5408\u3002GPU \u63d0\u4f9b\u4e86\u5927\u89c4\u6a21\u5e76\u884c\u7684\u77e2\u91cf\/\u77e9\u9635\u64cd\u4f5c\u80fd\u529b\uff0c\u4f7f\u5f97\u8bad\u7ec3\u65f6\u95f4\u4ece\u201c\u51e0\u5468\u201d\u7f29\u77ed\u5230\u201c\u51e0\u5929\u201d\u751a\u81f3\u201c\u51e0\u5c0f\u65f6\u201d\uff0c\u8fd9\u76f4\u63a5\u52a0\u901f\u4e86\u5b9e\u9a8c\u7684\u8fed\u4ee3\u901f\u5ea6\u4e0e\u6a21\u578b\u89c4\u6a21\u7684\u6269\u5c55\u3002\u7b80\u8a00\u4e4b\uff1a\u6709\u4e86 GPU\uff0c\u7814\u7a76\u8005\u53ef\u4ee5\u5c1d\u8bd5\u66f4\u6df1\u3001\u66f4\u5bbd\u7684\u7f51\u7edc\u5e76\u5728\u73b0\u5b9e\u65f6\u95f4\u5185\u89c2\u5bdf\u7ed3\u679c\u3002<\/span><\/p>\n


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

    \n
    \n
    \n
    3\u3001\u795e\u7ecf\u7f51\u7edc\u7684\u4e09\u5927\u5173\u952e\u8981\u7d20<\/span><\/span>
    <\/span><\/strong><\/span><\/section>\n<\/section>\n


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


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


    <\/span><\/p>\n<\/section>\n<\/h1>\n<\/section>\n<\/section>\n


    <\/span><\/h2>\n
    <\/section>\n

    \u65e0\u8bba\u4f60\u9762\u5bf9\u600e\u6837\u7684\u65b0\u4efb\u52a1\uff0c\u628a\u95ee\u9898\u62c6\u6210\u8fd9\u4e09\u5757\u5f80\u5f80\u80fd\u8fc5\u901f\u7406\u6e05\u601d\u8def\uff1a\u7f51\u7edc\u7684\u201c\u5f62\u201d\uff0c\u8bad\u7ec3\u7684\u201c\u672f\u201d\uff0c\u4e0e\u53c2\u6570\u7684\u201c\u5ea6\u201d\u3002<\/span><\/p>\n


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

    \n
    \n
    \n
    3.1 \u7f51\u7edc\u67b6\u6784 \u2014 \u9009\u62e9\u5408\u9002\u7684\u201c\u5f62\u201d<\/span>
    <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


    <\/span><\/h3>\n

    \u67b6\u6784\u5b9a\u4e49\u4e86\u201c\u4fe1\u606f\u5982\u4f55\u6d41\u52a8\u3001\u5982\u4f55\u7ec4\u5408\u201d\u3002\u9009\u62e9\u5408\u9002\u7684\u67b6\u6784\u9700\u8981\u628a\u95ee\u9898\u4e0e\u6570\u636e\u7684\u7ed3\u6784\u5bf9\u5e94\u8d77\u6765\uff1a<\/span><\/p>\n

      \n
    • \n

      \u56fe\u50cf -> \u5377\u79ef\uff08\u5229\u7528\u7a7a\u95f4\u5c40\u90e8\u6027\u4e0e\u5e73\u79fb\u4e0d\u53d8\u6027\uff09\u3002<\/span><\/p>\n<\/li>\n

    • \n

      \u5e8f\u5217\uff08\u6587\u672c\/\u8bed\u97f3\uff09-> \u5e8f\u5217\u5efa\u6a21\uff08RNN\u3001Transformer\uff09\u3002<\/span><\/p>\n<\/li>\n

    • \n

      \u56fe\u7ed3\u6784\u6570\u636e\uff08\u793e\u4ea4\u7f51\u7edc \/ \u5206\u5b50 \/ \u5173\u7cfb\u56fe\uff09-> \u56fe\u795e\u7ecf\u7f51\u7edc\uff08GNN\uff09\u3002<\/span><\/p>\n<\/li>\n<\/ul>\n

      \u8bbe\u8ba1\u67b6\u6784\u65f6\uff0c\u4e5f\u8981\u8003\u8651\u9c81\u68d2\u6027\u3001\u53c2\u6570\u6570\u76ee\u3001\u8ba1\u7b97\u91cf\uff08FLOPs\uff09\u3001\u4ee5\u53ca\u662f\u5426\u4fbf\u4e8e\u5e76\u884c\u5316\u8bad\u7ec3\u7b49\u5de5\u7a0b\u7ea6\u675f\u3002<\/span><\/p>\n


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

      \n
      \n
      \n
      3.2 \u8d85\u53c2\u6570 \u2014 \u6a21\u578b\u4e4b\u5916\u7684\u91cd\u8981\u201c\u8bbe\u7f6e\u201d<\/span>
      <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


      <\/span><\/h3>\n

      \u8d85\u53c2\u6570\uff08Hyperparameters\uff09<\/span>\u4e0d\u662f\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u88ab\u5b66\u4e60\u7684\u6743\u91cd\uff08\u90a3\u4e9b\u662f\u53c2\u6570\uff09\uff0c\u800c\u662f\u5728\u6a21\u578b\u8bbe\u8ba1\u6216\u8bad\u7ec3\u524d\u8bbe\u5b9a\u7684\u503c\uff0c\u4f8b\u5982\u5c42\u6570\u3001\u6bcf\u5c42\u795e\u7ecf\u5143\u6570\u3001\u5b66\u4e60\u7387\u3001\u6279\u91cf\u5927\u5c0f\u3001\u6b63\u5219\u5316\u5f3a\u5ea6\u3001dropout \u6bd4\u4f8b\u7b49\u3002\u8fd9\u4e9b\u8bbe\u5b9a\u76f4\u63a5\u51b3\u5b9a\u4e86\u6a21\u578b\u7684\u8868\u73b0\u4e0e\u8bad\u7ec3\u884c\u4e3a\u3002\u8d85\u53c2\u6570\u7684\u641c\u7d22\uff08\u7f51\u683c\u641c\u7d22\u3001\u8d1d\u53f6\u65af\u4f18\u5316\u3001\u8d85\u53c2\u8c03\u5ea6\uff09\u5f80\u5f80\u662f\u5de5\u7a0b\u5b9e\u8df5\u4e2d\u975e\u5e38\u8017\u65f6\u4f46\u5fc5\u8981\u7684\u6b65\u9aa4\u3002<\/span><\/span><\/p>\n


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

      \n
      \n
      \n
      3.3 \u8bad\u7ec3\u65b9\u6cd5 \u2014 \u8ba9\u6a21\u578b\u5b66\u4f1a\u201c\u5982\u4f55\u5b66\u4e60\u201d<\/span>
      <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


      <\/span><\/h3>\n

      \u8bad\u7ec3\u65b9\u6cd5\u5305\u542b\u4f18\u5316\u7b97\u6cd5\uff08SGD\u3001Momentum\u3001Adam \u7b49\uff09\u3001\u635f\u5931\u51fd\u6570\u9009\u62e9\uff08\u56de\u5f52\u3001\u5206\u7c7b\u6216\u5bf9\u6bd4\u635f\u5931\uff09\u3001\u6b63\u5219\u5316\uff08L2\u3001dropout\u3001\u6570\u636e\u589e\u5f3a\uff09\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u3001\u65e9\u505c\uff08early stopping\uff09\u7b49\u3002\u4e24\u70b9\u503c\u5f97\u5f3a\u8c03\uff1a<\/span><\/p>\n

        \n
      1. \n

        \u4f18\u5316\u5668\u4e0e\u5b66\u4e60\u7387<\/span><\/span><\/strong>\uff1a\u5b66\u4e60\u7387\u662f\u6700\u654f\u611f\u7684\u8d85\u53c2\u6570\u3002\u5408\u9002\u7684\u5b66\u4e60\u7387\u4e0e\u8c03\u5ea6\u7b56\u7565\u80fd\u8ba9\u6a21\u578b\u5feb\u901f\u6536\u655b\u5e76\u53d6\u5f97\u66f4\u597d\u7684\u6cdb\u5316\u3002<\/span><\/span><\/p>\n<\/li>\n

      2. \n

        \u8bad\u7ec3\u6d41\u7a0b\u4e0e\u4efb\u52a1\u8bbe\u8ba1<\/span><\/span><\/strong>\uff1a\u6709\u4e9b\u4efb\u52a1\u9760\u7eaf\u76d1\u7763\u5b66\u4e60\u5c31\u591f\uff0c\u800c\u590d\u6742\u4efb\u52a1\uff08\u5982 AlphaGo\uff09\u7ed3\u5408\u76d1\u7763\u5b66\u4e60\u4e0e\u5f3a\u5316\u5b66\u4e60\uff0c\u5148\u4ece\u4eba\u7c7b\u6570\u636e\u5b66\u4e60\uff08\u5148\u9a8c\uff09\uff0c\u518d\u901a\u8fc7\u81ea\u5bf9\u5f08\u6216\u65e0\u76d1\u7763\u65b9\u5f0f\u5f3a\u5316\uff08\u63a2\u7d22\uff09\uff0c\u8fd9\u79cd\u6df7\u5408\u8bad\u7ec3\u6d41\u7a0b\u80fd\u663e\u8457\u63d0\u5347\u6027\u80fd\u3002<\/span><\/span><\/p>\n<\/li>\n<\/ol>\n


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

        \n
        \n
        \n
        4\u3001PyTorch\uff1a\u4e3a\u4ec0\u4e48\u5728\u7814\u7a76\/\u6559\u5b66\u4e2d\u5982\u6b64\u53d7\u6b22\u8fce<\/span><\/span>
        <\/span><\/strong><\/span><\/section>\n<\/section>\n


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


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


        <\/span><\/p>\n<\/section>\n<\/h1>\n<\/section>\n<\/section>\n


        <\/span><\/h2>\n
        <\/section>\n

        \u5728\u4f17\u591a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u91cc\uff0cPyTorch \u88ab\u8bb8\u591a\u7814\u7a76\u8005\u4e0e\u6559\u5e08\u504f\u7231\uff0c\u539f\u56e0\u4e0d\u4ec5\u662f\u8bed\u6cd5\u4e0a\u7684\u201c\u6f02\u4eae\u201d\uff0c\u66f4\u5728\u4e8e\u5b83\u628a<\/span>\u8ba1\u7b97\u62bd\u8c61<\/span><\/span><\/strong>\u505a\u5f97\u65e2\u7075\u6d3b\u53c8\u900f\u660e\uff0c\u8fd9\u5bf9\u7406\u89e3\u6df1\u5ea6\u5b66\u4e60\u539f\u7406\u975e\u5e38\u6709\u5e2e\u52a9\u3002<\/span><\/span><\/p>\n


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

        \n
        \n
        \n
        4.1 \u52a8\u6001\u8ba1\u7b97\u56fe<\/span>
        <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


        <\/span><\/h3>\n

        PyTorch \u7684\u6838\u5fc3\u7279\u8272\u4e4b\u4e00\u662f<\/span>\u52a8\u6001\u8ba1\u7b97\u56fe\uff08Dynamic Computation Graph\uff09<\/span><\/span><\/strong>\uff1a\u6bcf\u6b21\u524d\u5411\u8ba1\u7b97\u65f6\uff0c\u6846\u67b6\u90fd\u4f1a\u5373\u65f6\u6784\u5efa\u4e00\u4e2a\u8ba1\u7b97\u56fe\uff0c\u8bb0\u5f55\u8fd0\u7b97\u8282\u70b9\u4e0e\u4f9d\u8d56\u5173\u7cfb\uff1b\u5728\u53cd\u5411\u4f20\u64ad\u65f6\uff0c\u6846\u67b6\u6cbf\u7740\u8fd9\u4e2a\u56fe\u81ea\u52a8\u8ba1\u7b97\u68af\u5ea6\u3002\u8fd9\u79cd\u6309\u9700\u6784\u5efa\u4e14\u53ef\u4ee5\u968f\u8fd0\u884c\u52a8\u6001\u6539\u53d8\u7684\u673a\u5236\uff0c\u4f7f\u5f97\u5b9e\u73b0\u9012\u5f52\u3001\u6761\u4ef6\u5206\u652f\u6216\u590d\u6742\u63a7\u5236\u6d41\u7684\u6a21\u578b\u975e\u5e38\u81ea\u7136\u3002\u5bf9\u6559\u5b66\u4e0e\u7814\u7a76\u6765\u8bf4\uff0c\u8fd9\u79cd\u53ef\u89c2\u5bdf\u3001\u53ef\u8c03\u8bd5\u7684\u7279\u6027\u6781\u5176\u6709\u4ef7\u503c\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u65e9\u671f\u7684\u9759\u6001\u56fe\u6846\u67b6\uff08\u9700\u8981\u5148\u5b9a\u4e49\u5b8c\u6574\u56fe\u518d\u8fd0\u884c\uff09\u5728\u8c03\u8bd5\u4e0e\u7075\u6d3b\u6027\u4e0a\u4e0d\u5982\u52a8\u6001\u56fe\u76f4\u89c2\u3002<\/span><\/span><\/p>\n


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

        \n
        \n
        \n
        4.2 \u4ec0\u4e48\u662f\u5f20\u91cf<\/span>
        <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


        <\/span><\/h3>\n

        \u5f20\u91cf\uff08Tensor\uff09<\/span>\u662f\u5bf9\u201c\u5411\u91cf\/\u77e9\u9635\u201d\u5728\u66f4\u9ad8\u7ef4\u5ea6\u4e0a\u7684\u81ea\u7136\u63a8\u5e7f\uff08\u591a\u7ef4\u6570\u7ec4\uff09\u3002\u5728\u6df1\u5ea6\u5b66\u4e60\u4e2d\uff0c\u56fe\u50cf\u3001\u6279\u91cf\u6570\u636e\u3001\u6743\u91cd\u3001\u68af\u5ea6\u7b49\u90fd\u53ef\u4ee5\u7edf\u4e00\u770b\u4f5c\u5f20\u91cf\u3002\u4e0e Numpy \u6570\u7ec4\u76f8\u6bd4\uff0c\u5f20\u91cf\u80fd\u591f\u900f\u660e\u5730\u5728 GPU \u4e0a\u8fd0\u7b97\uff0c\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u6211\u4eec\u5728\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u4e2d\u5927\u91cf\u4f7f\u7528\u5f20\u91cf\u800c\u975e\u5355\u7eaf\u7684\u6570\u7ec4\u3002\u7406\u89e3\u5f20\u91cf\u7684\u5f62\u72b6\uff08shape\uff09\u3001\u7ef4\u5ea6\uff08rank\uff09\u3001\u5e7f\u64ad\u673a\u5236\uff08broadcasting\uff09\u4e0e\u8f6c\u7f6e\u3001\u5c55\u5f00\u7b49\u57fa\u672c\u64cd\u4f5c\uff0c\u662f\u638c\u63e1\u6df1\u5ea6\u5b66\u4e60\u5b9e\u73b0\u7ec6\u8282\u7684\u57fa\u7840\u3002<\/span><\/span><\/p>\n


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

        \n
        \n
        \n
        4.3 \u81ea\u52a8\u5fae\u5206\u4e0e\u53cd\u5411\u4f20\u64ad\u7684\u81ea\u52a8\u5316<\/span>
        <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


        <\/span><\/h3>\n

        \u81ea\u52a8\u5fae\u5206\uff08Autograd\uff09<\/span>\u662f\u4e00\u79cd\u673a\u5236\uff0c\u5b83\u8bb0\u5f55\u524d\u5411\u8fd0\u7b97\u4e2d\u6bcf\u4e2a\u64cd\u4f5c\uff0c\u5e76\u5728\u53cd\u5411\u9636\u6bb5\u4f9d\u636e\u94fe\u5f0f\u6cd5\u5219\u81ea\u52a8\u7ec4\u5408\u5c40\u90e8\u5bfc\u6570\u6765\u5f97\u5230\u76ee\u6807\u5bf9\u6bcf\u4e2a\u53c2\u6570\u7684\u68af\u5ea6\u3002PyTorch \u5b9e\u73b0\u7684\u81ea\u52a8\u5fae\u5206 \u8ba9\u4f7f\u7528\u8005\u65e0\u987b\u624b\u5de5\u63a8\u5bfc\u590d\u6742\u6a21\u578b\u7684\u504f\u5bfc\u6570\uff0c\u4ece\u800c\u628a\u7cbe\u529b\u653e\u5728\u6a21\u578b\u8bbe\u8ba1\u4e0e\u95ee\u9898\u5efa\u6a21\u4e0a\u3002\u6559\u5b66\u4e0a\u8bb2\uff0c\u5148\u7406\u89e3\u201c\u8ba1\u7b97\u56fe + \u94fe\u5f0f\u6cd5\u5219\u201d\u5982\u4f55\u628a\u590d\u6742\u51fd\u6570\u7684\u5bfc\u6570\u62c6\u89e3\u6210\u7b80\u5355\u5c40\u90e8\u5bfc\u6570\u7684\u4e58\u79ef\uff0c\u662f\u7406\u89e3\u6df1\u5ea6\u5b66\u4e60\u5b66\u4e60\u672c\u8d28\u7684\u5173\u952e\u3002<\/span><\/span><\/p>\n


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

        \n
        \n
        \n
        4.4 PyTorch \u5728\u8bfe\u7a0b\/\u7814\u7a76\u4e2d\u7684\u4f18\u52bf<\/span>
        <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


        <\/span><\/h3>\n
          \n
        • \n

          \u4ee3\u7801\u98ce\u683c\u66f4\u50cf\u5e38\u89c4 Python\uff0c\u964d\u4f4e\u5b66\u4e60\u95e8\u69db\uff0c\u4f7f\u7406\u8bba\u4e0e\u5b9e\u73b0\u7684\u5bf9\u5e94\u5173\u7cfb\u66f4\u76f4\u89c2\u3002<\/span><\/p>\n<\/li>\n

        • \n

          \u52a8\u6001\u56fe\u4fbf\u4e8e\u8c03\u8bd5\u590d\u6742\u7ed3\u6784\u3001\u5b9e\u9a8c\u6027\u6539\u52a8\u4e0e\u7b97\u6cd5\u539f\u578b\u9a8c\u8bc1\u3002<\/span><\/p>\n<\/li>\n

        • \n

          \u4e30\u5bcc\u7684\u5de5\u5177\u5e93\u4e0e\u6a21\u578b\u4ed3\u5e93\u6709\u5229\u4e8e\u8fc1\u79fb\u5b66\u4e60\u4e0e\u5feb\u901f\u539f\u578b\u3002<\/span><\/p>\n<\/li>\n<\/ul>\n


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

          \n
          \n
          \n
          5\u3001\u4ece\u7279\u5f81\u5b66\u4e60\u5230\u8fc1\u79fb\u5b66\u4e60<\/span><\/span>
          <\/span><\/strong><\/span><\/section>\n<\/section>\n


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


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


          <\/span><\/p>\n<\/section>\n<\/h1>\n<\/section>\n<\/section>\n


          <\/span><\/h2>\n
          <\/section>\n

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

          \n
          \n
          \n
          5.1 \u4ec0\u4e48\u662f\u201c\u7279\u5f81\u5b66\u4e60\u201d\uff1f<\/span>
          <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


          <\/span><\/h3>\n

          \u4f20\u7edf\u65b9\u6cd5\u5f80\u5f80\u628a\u201c\u7279\u5f81\u5de5\u7a0b\u201d\u4f5c\u4e3a\u4eba\u5de5\u5236\u5b9a\u7684\u6b65\u9aa4\uff08\u4f8b\u5982\u624b\u5de5\u63d0\u53d6 SIFT\u3001HOG \u7b49\u7279\u5f81\uff09\uff1b\u6df1\u5ea6\u5b66\u4e60\u6700\u5927\u7684\u4e0d\u540c\u5728\u4e8e\u7aef\u5230\u7aef\u5b66\u4e60\uff1a\u7f51\u7edc\u4ece\u539f\u59cb\u6570\u636e\u4e2d\u81ea\u52a8\u5b66\u4e60\u5230\u5c42\u6b21\u5316\u7684\u7279\u5f81\u8868\u793a\u3002\u65e9\u671f\u5c42\u901a\u5e38\u6355\u6349\u8fb9\u7f18\u3001\u89d2\u70b9\u7b49\u5c40\u90e8\u6a21\u5f0f\uff1b\u4e2d\u95f4\u5c42\u7ec4\u5408\u6210\u7eb9\u7406\u6216\u5c40\u90e8\u7ed3\u6784\uff1b\u9ad8\u5c42\u5219\u62bd\u8c61\u51fa\u6982\u5ff5\u7ea7\u522b\u7684\u8bed\u4e49\uff08\u4f8b\u5982\u201c\u4eba\u8138\u201d\u6216\u201c\u732b\u201d\uff09\u3002\u8fd9\u79cd\u591a\u5c3a\u5ea6\u3001\u591a\u5c42\u6b21\u7684\u7279\u5f81\u5b66\u4e60\u662f\u6df1\u5ea6\u7f51\u7edc\u5e7f\u6cdb\u6210\u529f\u7684\u91cd\u8981\u539f\u56e0\u3002<\/span><\/p>\n


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

          \n
          \n
          \n
          5.2 \u8fc1\u79fb\u5b66\u4e60\u4e3a\u4ec0\u4e48\u6709\u6548<\/span>
          <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


          <\/span><\/h3>\n

          \u8fc1\u79fb\u5b66\u4e60\uff08Transfer Learning\uff09<\/span>\u7684\u57fa\u672c\u601d\u8def\u662f\uff1a\u5728\u4e00\u4e2a\u5927\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u51fa\u4e00\u7ec4\u901a\u7528\u7684\u8868\u5f81\uff0c\u7136\u540e\u628a\u8fd9\u4e9b\u8868\u5f81\u8fc1\u79fb\u5230\u5c0f\u6570\u636e\u96c6\u7684\u76ee\u6807\u4efb\u52a1\u4e0a\uff0c\u53ea\u5fae\u8c03\u5c11\u91cf\u53c2\u6570\u6216\u4f7f\u7528\u5176\u4f5c\u4e3a\u7279\u5f81\u63d0\u53d6\u5668\u3002\u8fd9\u5728\u5b9e\u8df5\u4e2d\u6781\u5176\u6709\u6548\uff0c\u5c24\u5176\u662f\u5f53\u76ee\u6807\u4efb\u52a1\u6570\u636e\u7a00\u7f3a\u65f6\u3002\u8fc1\u79fb\u5b66\u4e60\u7684\u6210\u529f\u4f9d\u8d56\u4e8e\u6e90\u4efb\u52a1\u4e0e\u76ee\u6807\u4efb\u52a1\u7684\u76f8\u4f3c\u6027\u4ee5\u53ca\u7f51\u7edc\u5b66\u4e60\u5230\u7684\u7279\u5f81\u7684\u666e\u9002\u6027\u3002<\/span><\/span><\/p>\n


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

          \n
          \n
          \n
          5.3 \u5b9e\u9645\u5e94\u7528\u573a\u666f<\/span> <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


          <\/span><\/h3>\n
            \n
          • \n

            \u89c6\u89c9\u5206\u7c7b\uff1a\u7528\u5728 ImageNet \u4e0a\u9884\u8bad\u7ec3\u7684\u6a21\u578b\u4f5c\u4e3a backbone\uff0c\u5728\u533b\u5b66\u5f71\u50cf\u3001\u5de5\u4e1a\u68c0\u6d4b\u4e0a\u5fae\u8c03\uff0c\u5f80\u5f80\u80fd\u5728\u5c11\u91cf\u6807\u6ce8\u4e0b\u53d6\u5f97\u9ad8\u6027\u80fd\u3002<\/span><\/p>\n<\/li>\n

          • \n

            \u81ea\u7136\u8bed\u8a00\uff1a\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff08\u4f8b\u5982 BERT\uff09\u5728\u4e0b\u6e38\u4efb\u52a1\uff08\u60c5\u611f\u5206\u6790\u3001\u95ee\u7b54\u3001\u6587\u672c\u5206\u7c7b\uff09\u5fae\u8c03\u53ef\u663e\u8457\u63d0\u5347\u6548\u679c\u3002<\/span><\/p>\n<\/li>\n

          • \n

            \u7279\u6b8a\u9886\u57df\uff1a\u9065\u611f\u3001\u8bed\u97f3\u3001\u5316\u5b66\u5206\u5b50\u9884\u6d4b\u7b49\u9886\u57df\u90fd\u53ef\u501f\u52a9\u8fc1\u79fb\u6216\u5c11\u91cf\u5fae\u8c03\u5b9e\u73b0\u5b9e\u9645\u5e94\u7528\u3002<\/span><\/p>\n<\/li>\n<\/ul>\n


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

            \n
            \n
            \n
            6 \u5c55\u671b\u672a\u6765\uff1a\u6df1\u5ea6\u5b66\u4e60\u7684\u673a\u9047\u4e0e\u6311\u6218<\/span><\/span>
            <\/span><\/strong><\/span><\/section>\n<\/section>\n


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


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


            <\/span><\/p>\n<\/section>\n<\/h1>\n<\/section>\n<\/section>\n


            <\/span><\/h2>\n
            \n

            \n
            \n
            \n
            6.1 \u65b0\u5174\u67b6\u6784\u4e0e\u7814\u7a76\u65b9\u5411<\/span>
            <\/span><\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/h1>\n<\/section>\n


            <\/span><\/h3>\n

            \u5f53\u524d\u6d3b\u8dc3\u7684\u65b9\u5411\u5305\u62ec\u4f46\u4e0d\u9650\u4e8e\uff1a<\/span><\/p>\n