{"id":2614,"date":"2025-04-29T11:07:34","date_gmt":"2025-04-29T03:07:34","guid":{"rendered":"https:\/\/www.lanxy.ink\/?p=2614"},"modified":"2025-05-09T20:32:52","modified_gmt":"2025-05-09T12:32:52","slug":"%e4%bb%8e-0-%e5%bc%80%e5%a7%8b%e7%9a%84-transformer-%e6%a2%b3%e7%90%86%ef%bc%882%ef%bc%89%e5%ba%8f%e5%88%97%e6%a8%a1%e5%9e%8b%e4%b8%8e%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9c%ba%e5%88%b6","status":"publish","type":"post","link":"https:\/\/www.lanxy.ink\/?p=2614","title":{"rendered":"\u4ece 0 \u5f00\u59cb\u7684 Transformer \u68b3\u7406\uff082\uff09\u7ecf\u5178\u5e8f\u5217\u6a21\u578b"},"content":{"rendered":"\n

\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08Rerrent Neural Network, RNN\uff09<\/h2>\n\n\n\n

\u4f20\u7edf\u7684\u795e\u7ecf\u7f51\u7edc\u5728\u5904\u7406\u4e0a\u4e00\u6761\u6570\u636e\u548c\u4e0b\u4e00\u6761\u6570\u636e\u65f6\u4e0d\u4f1a\u8fdb\u884c\u4ea4\u6d41\uff0c\u6362\u8a00\u4e4b\uff0c\u6a21\u578b\u5e76\u4e0d\u4f1a\u8bb0\u5f97\u4e0a\u4e00\u6761\u5b83\u5904\u7406\u7684\u6570\u636e\u662f\u4ec0\u4e48\uff0c\u6709\u4ec0\u4e48\u7279\u70b9\u3002\u8fd9\u79cd\u7279\u6027\u5728\u5904\u7406\u56fe\u7247\u65f6\u95ee\u9898\u4e0d\u5927\uff0c\u524d\u4e00\u5f20\u56fe\u7247\u548c\u540e\u4e00\u5f20\u56fe\u7247\u4e4b\u95f4\u5e76\u6ca1\u6709\u4ec0\u4e48\u5e8f\u5217\u7279\u6027<\/strong>\u3002<\/p>\n\n\n\n

\u4f46\u5982\u679c\u662f\u4e00\u53e5\u8bdd\u3001\u4e00\u6bb5\u97f3\u9891\u3001\u6216\u8005\u4e00\u6bb5\u89c6\u9891\u7684\u65f6\u5019\uff0c\u8fd9\u4e9b\u6570\u636e\u91cc\u9762\u5f80\u5f80\u90fd\u542b\u6709\u65f6\u95f4\u4e0a\u7684\u5173\u8054\uff0c\u4e5f\u5373\u65f6\u5e8f\u4fe1\u606f\u3002\u8fd9\u4e2a\u65f6\u5019\u5c31\u9700\u8981\u5faa\u73af\u795e\u7ecf\u7f51\u7edc RNN \u6765\u5904\u7406\u3002”RNN \u5bf9\u5177\u6709\u5e8f\u5217\u7279\u6027\u7684\u6570\u636e\u975e\u5e38\u6709\u6548\uff0c\u5b83\u80fd\u6316\u6398\u6570\u636e\u4e2d\u7684\u65f6\u5e8f\u4fe1\u606f\u4ee5\u53ca\u8bed\u4e49\u4fe1\u606f\uff0c<\/b>\u5229\u7528\u4e86 RNN \u7684\u8fd9\u79cd\u80fd\u529b\uff0c\u4f7f\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u5728\u89e3\u51b3\u8bed\u97f3\u8bc6\u522b\u3001\u8bed\u8a00\u6a21\u578b\u3001\u673a\u5668\u7ffb\u8bd1\u4ee5\u53ca\u65f6\u5e8f\u5206\u6790\u7b49 NLP \u9886\u57df\u7684\u95ee\u9898\u65f6\u6709\u6240\u7a81\u7834\u3002”[1]<\/p>\n\n\n\n

\u4e00\u4e2a\u8bcd\u8bed\u5728\u4e0d\u540c\u7684\u53e5\u5b50\u4e0a\u4e0b\u6587\u91cc\u4f1a\u6709\u4e0d\u540c\u7684\u610f\u4e49\uff0c\u6bd4\u5982  miss \u8fd9\u4e2a\u8bcd\uff0c\u53ef\u4ee5\u662f\u201c\u9519\u8fc7\u201d\uff0c\u4e5f\u53ef\u4ee5\u662f\u201c\u60f3\u5ff5\u201d\uff0c\u4e0d\u540c\u7684\u8bed\u5883\u91cc\u610f\u4e49\u5b8c\u5168\u4e0d\u540c\uff0c\u6b63\u5982 [1] \u4e2d\u6240\u63cf\u8ff0\u7684\u90a3\u6837\uff0c\u5047\u5982\u6211\u4eec\u6709\u4e00\u5806\u5df2\u7ecf\u7ffb\u8bd1\u597d\u7684\u82f1\u6587\u53e5\u5b50\uff0c\u82f1\u6587\u53e5\u5b50\u4f5c\u4e3a\u6837\u672c\uff0c\u7ffb\u8bd1\u4f5c\u4e3a\u6807\u7b7e\uff0c\u4f9b\u6a21\u578b\u8bad\u7ec3\uff0c\u5982\u679c\u4f7f\u7528\u795e\u7ecf\u7f51\u7edc\u53bb\u505a\u7ffb\u8bd1\u8fd9\u4e2a\u4e8b\uff0c\u5f53\u5b83\u5b66\u4e60 miss \u65f6\u53ef\u80fd\u5c31\u4f1a\u72af\u8ff7\u7cca\uff0c\u56e0\u4e3a\u5582\u7ed9\u6a21\u578b\u7684 miss \u91cc\u6709\u4e9b\u6807\u7b7e\u662f\u201c\u9519\u8fc7\u201d\uff0c\u6709\u4e9b\u6807\u7b7e\u662f\u201c\u60f3\u5ff5\u201d\uff0c\u6a21\u578b\u5bf9 miss \u7684\u7ffb\u8bd1\u5c31\u53d6\u51b3\u4e8e\u54ea\u79cd\u6807\u7b7e\u5360\u591a\u6570\u3002\u8fd9\u79cd\u6a21\u578b\u8bad\u7ec3\u51fa\u6765\u5f97\u5230\u7684\u6027\u80fd\u81ea\u7136\u4e0d\u4f1a\u5f88\u597d\u3002<\/p>\n\n\n\n

\u4f46\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u5c31\u4e0d\u4e00\u6837\u4e86\uff0c\u4ece\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u540d\u5b57\u4e0a\u6765\u770b\uff0c\u201c\u5faa\u73af\u201d\u4e24\u4e2a\u5b57\u5c31\u662f\u8fd9\u4e2a\u7f51\u7edc\u7684\u6838\u5fc3\uff0c\u7f51\u7edc\u5185\u90e8\u6709\u5faa\u73af\uff0c\u6a21\u578b\u5904\u7406\u5b8c\u4e0a\u4e00\u4e2a\u6570\u636e\u540e\uff0c\u5185\u90e8\u7684\u72b6\u6001\u4f1a\u4fdd\u7559\u4e0b\u6765\uff0c\u53c2\u4e0e\u5230\u4e0b\u4e00\u4e2a\u6570\u636e\u7684\u5904\u7406\u4e2d\u53bb<\/strong>\u3002\\(x\\) \u662f\u8f93\u5165\uff0c\\(h_t\\) \u662f\u9690\u85cf\u5c42\u8f93\u51fa\uff0c\\(A\\) \u76f8\u5f53\u4e8e\u9690\u85cf\u5c42\u7684\u6743\u91cd\u77e9\u9635\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u4ece\u4e0a\u56fe\u53ef\u4ee5\u770b\u51fa\uff0c\\(h_1\\) \u662f \\(x_1\\) \u548c\u4e0a\u4e00\u4e2a\u65f6\u523b\u7684 \\(A\\) \u505a\u8f93\u5165\uff0c\u7ecf\u9690\u85cf\u5c42\u5904\u7406\u540e\u7684\u8f93\u51fa\u3002\u6b64\u65f6 \\(h_1\\) \u91cc\u65e2\u6709 \\(x_0\\) \u7684\u4fe1\u606f\uff08\u90e8\u5206\u4fdd\u7559\u5728\u4e0a\u4e00\u4e2a\u65f6\u523b\u7684 \\(A\\) \u4e2d\uff09\u53c8\u6709 \\(x_1\\) \u7684\u4fe1\u606f\uff0cRNN \u5c31\u5177\u5907\u4e86\u83b7\u53d6\u4e0a\u6587\u4fe1\u606f\u7684\u80fd\u529b\u3002\u800c\u4e14\u5982\u679c\u4ea4\u6362 \\(x_0\\) \u548c \\(x_1\\) \uff0c\u5f97\u5230\u7684 \\(h_1\\) \u4e5f\u4f1a\u5b8c\u5168\u4e0d\u540c\uff0c\u6a21\u578b\u5bf9\u5e8f\u5217\u4e5f\u53d8\u5f97\u66f4\u654f\u611f\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u6807\u51c6\u7684 RNN \u4e00\u4e2a\u6a21\u5757\u5185\u90e8\u4ec5\u6709\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u5c42\uff08\u9ec4\u8272\u5757\uff09\uff0c\u6fc0\u6d3b\u51fd\u6570\u4e3a tanh\uff0c\u7136\u540e\u7531\u4e00\u4e32\u91cd\u590d\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u5757\u7ec4\u6210\u3002RNN \u7684\u7f3a\u70b9\u5f88\u660e\u663e\uff0c\u6bcf\u6b21\u5b83\u90fd\u53ea\u662f\u7528\u4e0a\u4e00\u4e2a\u65f6\u523b\u7684\u72b6\u6001\uff0c\u66f4\u591a\u65f6\u5019\u6211\u4eec\u9700\u8981\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\uff0c\u7406\u8bba\u4e0a RNN \u53ef\u4ee5\u505a\u5230\u5904\u7406\u957f\u671f\u4f9d\u8d56\u5173\u7cfb\uff0c\u6bd5\u7adf \\(A\\) \u662f\u4e00\u8def\u5f80\u540e\u4f20\u8fc7\u53bb\u7684\uff0c\u4f46\u5728\u5b9e\u8df5\u4e2d RNN \u6ca1\u529e\u6cd5\u505a\u5230\uff0c\u5904\u7406\u5b8c\u65b0\u7684\u6570\u636e\u540e\uff0c\u5f53\u4e0b\u68af\u5ea6\u66f4\u65b0\u65f6\uff0c\u548c\u5f53\u524d\u8ddd\u79bb\u8fc7\u957f\u7684\u8fdc\u53e4\u4fe1\u606f\u5c31\u4f1a\u201c\u8870\u51cf\u201d<\/strong>\uff0c\u5c31\u662f\u8bf4\u524d\u9762\u7684\u68af\u5ea6\u4f1a\u6d88\u5931\uff0c\u68af\u5ea6\u6d88\u5931\u4e86\u5c31\u6ca1\u529e\u6cd5\u7528\u4e0a\u4e86\u3002<\/p>\n\n\n\n

\u6240\u5e78\u4eba\u4eec\u53c8\u63d0\u51fa\u4e86\u957f\u77ed\u671f\u8bb0\u5fc6\u7f51\u7edc<\/strong>\uff0c\u6539\u5584\u4e86 RNN \u7684\u8bf8\u591a\u4e0d\u4fbf\u3002<\/p>\n\n\n\n

\u957f\u77ed\u671f\u8bb0\u5fc6\u7f51\u7edc\uff08Long Short Term Memory networks, LSTM\uff09<\/h2>\n\n\n
\n
\"\"<\/div><\/figure>\n<\/div>\n\n
\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

LSTM \u4e5f\u662f\u8fd9\u79cd\u94fe\u5f0f\u7ed3\u6784\uff0c\u4f46\u662f\u6bcf\u4e2a\u6a21\u5757\u91cc\u9762\u6709\u56db\u4e2a\u795e\u7ecf\u7f51\u7edc\u5c42\u3002\u5206\u522b\u62c5\u5f53\u4e09\u4e2a\u95e8\u7684\u5de5\u4f5c\u3002\u5206\u522b\u662f\u9057\u5fd8\u95e8\uff08Forget Gate\uff09\u3001\u8f93\u5165\u95e8\uff08Input Gate\uff09\u548c\u8f93\u51fa\u95e8\uff08Output Gate\uff09\u3002\u6a21\u5757\u5185\u6700\u4e0a\u65b9\u90a3\u6761\u76f4\u7ebf\u76f8\u5f53\u4e8e\u8bb0\u5fc6\u6d41\u6c34\u7ebf\uff08\u8bb0\u5fc6\u7ec6\u80de\uff09\uff0c\u5728\u4e09\u4e2a\u95e8\u7684\u5f71\u54cd\u4e0b\u66f4\u65b0\u5176\u4e2d\u7684\u8bb0\u5fc6\u3002<\/p>\n\n\n\n

\u9057\u5fd8\u95e8\uff08Input Gate\uff09<\/h3>\n\n\n\n

\u7528\u6765\u51b3\u5b9a\u4ece\u8bb0\u5fc6\u4e2d\u4e22\u6389\u54ea\u4e9b\u4fe1\u606f\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u8bb0\u5fc6\u5c31\u76f8\u5f53\u4e8e\u4e00\u4e2a\u4ed3\u5e93\uff08\u77e9\u9635\uff09\u91cc\u7684\u5f88\u591a\u8d27\u7269\uff08\u5143\u7d20\uff09\uff0c\u9057\u5fd8\u95e8\u5c31\u76f8\u5f53\u4e8e\u6e05\u7406\u5de5\uff0c\u5bf9\u6bcf\u4e00\u4e2a\u8d27\u7269\u8fdb\u884c\u68c0\u67e5\uff0c\u51b3\u5b9a\u54ea\u4e9b\u8be5\u7559\u4e0b\u3001\u54ea\u4e9b\u8be5\u6254\u6389\u3002\u4e0a\u4e00\u65f6\u523b\u7684\u8bb0\u5fc6 \\( C_{t-1} \\) \u8f93\u5165\u8fdb\u6765\uff0c\u4e0e  \\( f_t \\) \u505a\u9010\u5143\u7d20\u76f8\u4e58\uff0c\u8fbe\u6210\u9057\u5fd8\u7684\u76ee\u7684\u3002<\/p>\n\n\n\n

\\( f_t \\) \u53c8\u7531\u4e0a\u4e00\u65f6\u523b\u7684\u9690\u85cf\u5c42\u8f93\u51fa \\( h_{t-1} \\) \u548c\u8fd9\u4e00\u65f6\u523b\u7684\u8f93\u5165 \\( x_{t} \\) \u5171\u540c\u51b3\u5b9a\uff0c\u62fc\u63a5\u540e\u7ecf\u8fc7\u4e00\u4e2a Sigmoid \u6fc0\u6d3b\u7684\u795e\u7ecf\u7f51\u7edc\u5c42\uff0c\u56e0\u4e3a Sigmoid \u4f1a\u5c06\u503c\u538b\u7f29\u5728 0 \u5230 1 \u4e4b\u95f4\uff0c\u7c7b\u4f3c\u95e8\u63a7\u4fe1\u53f7\uff0c\u8fd9\u6837\u5c31\u505a\u5230\u4e86\u201c\u6e05\u7406\u201d\uff0c1 \u5c31\u5b8c\u5168\u4fdd\u7559\uff0c0 \u5c31\u5b8c\u5168\u4e22\u5f03\u3002<\/p>\n\n\n\n

\u8f93\u5165\u95e8\uff08Input Gate\uff09<\/h3>\n\n\n\n

\u7528\u6765\u51b3\u5b9a\u5f80\u8bb0\u5fc6\u4e2d\u5b58\u50a8\u54ea\u4e9b\u65b0\u4fe1\u606f\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u8bb0\u5fc6\u4ed3\u5e93\u5f97\u5230\u4e86\u6e05\u7406\uff0c\u73b0\u5728\u5f00\u59cb\u5f80\u91cc\u9762\u5b58\u65b0\u7684\u8d27\u7269\uff08\u5143\u7d20\u76f8\u52a0\uff09\u3002\u4f46\u662f\u5b58\u65b0\u8d27\u7269\u4e4b\u524d\uff0c\u9700\u8981\u201c\u5ba1\u6838\u5458\u201d\u8fd9\u4e2a\u8f93\u5165\u95e8\u6765\u5bf9\u8d27\u7269\u7684\u8d28\u91cf\u8fdb\u884c\u201c\u628a\u5173\u201d\uff08PS\uff1a\u8fd9\u4e5f\u662f RNN \u4e0d\u5177\u5907\u7684\u7279\u70b9\uff0cRNN \u662f\u5bf9\u6240\u6709\u8f93\u5165\u5168\u76d8\u63a5\u6536\uff09<\/em>\uff0c\u5c06 \\( \\tilde{C}_t \\) \u548c \\( i_{t} \\) \u76f8\u4e58\uff0c\u5f97\u5230\u201c\u5ba1\u6838\u201d\u540e\u7684\u8d27\u7269\uff0c\u518d\u548c\u6e05\u7406\u540e\u7684\u8bb0\u5fc6\u76f8\u52a0\uff0c\u8fbe\u6210\u8f93\u5165\u7684\u76ee\u7684\u3002<\/p>\n\n\n\n

\\( i_{t} \\) \u540c\u6837\u662f\u7531\u4e0a\u4e00\u65f6\u523b\u7684\u9690\u85cf\u5c42\u8f93\u51fa \\( h_{t-1} \\) \u548c\u8fd9\u4e00\u65f6\u523b\u7684\u8f93\u5165 \\( x_{t} \\) \u5171\u540c\u51b3\u5b9a\uff0c\u540c\u6837\u7ecf\u8fc7\u4e00\u4e2a Sigmoid \u6fc0\u6d3b\u7684\u795e\u7ecf\u7f51\u7edc\u5c42\uff0c\u4e5f\u76f8\u5f53\u4e8e\u95e8\u63a7\uff0c\u53ea\u4e0d\u8fc7\u662f\u5bf9\u8f93\u5165\u7684\u95e8\u63a7\u3002<\/p>\n\n\n\n

\\( \\tilde{C}_t \\) \u7684\u8f93\u5165\u4e5f\u662f\u4e00\u6837\u7684\uff0c\u552f\u4e00\u6709\u533a\u522b\u7684\u5730\u65b9\u5c31\u662f\u7ecf\u8fc7\u7684\u795e\u7ecf\u7f51\u7edc\u5c42\uff0c\u5176\u6fc0\u6d3b\u51fd\u6570\u662f tanh\u3002<\/p>\n\n\n

\u4e3a\u4ec0\u4e48\u9009\u7528 Tanh\uff1f \u4e2a\u4eba\u8ba4\u4e3a\u6700\u4e3b\u8981\u7684\u539f\u56e0\u662f\uff1aTanh \u53ef\u4ee5\u628a\u795e\u7ecf\u7f51\u7edc\u5c42\u7684\u8f93\u51fa\u6574\u5408\u5230 [-1, 1] \u4e4b\u95f4\uff0c\u8fd9\u6837\u9001\u5230\u8bb0\u5fc6\u4ed3\u5e93\u91cc\u9762\u7684\u8d27\u7269\u6709\u6b63\u4e5f\u6709\u8d1f\uff0c\u5141\u8bb8\u8bb0\u5fc6\u5f80\u201c\u6b63\u5411\u589e\u5f3a\u201d\u548c\u201c\u53cd\u5411\u589e\u5f3a\u201d\u4e24\u4e2a\u65b9\u5411\u8fdb\u884c\u8c03\u6574\u3002Sigmoid \u6fc0\u6d3b\u51fd\u6570\u7684\u8f93\u51fa\u8303\u56f4\u53ea\u6709 [0, 1] \u4e4b\u95f4\uff0c\u53ea\u80fd\u201c\u6b63\u5411\u589e\u5f3a\u201d\u6216\u8005\u201c\u5f31\u5316\u201d\u8bb0\u5fc6\uff0c\u8bb0\u5fc6\u91cc\u9762\u751a\u81f3\u53ef\u80fd\u53ea\u6709\u6b63\u503c\u3002<\/div><\/div>\n\n\n
\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u7ecf\u8fc7\u4e86\u9057\u5fd8\u95e8\u548c\u8f93\u5165\u95e8\uff0c\u5f53\u524d\u5c31\u4e0d\u4f1a\u518d\u52a8\u8bb0\u5fc6\u4ed3\u5e93\u91cc\u7684\u4e1c\u897f\uff08\u5982\u4e0a\u56fe\u6240\u793a\uff09\uff0c\u800c\u662f\u8003\u8651\u5982\u4f55\u4ece\u8bb0\u5fc6\u4ed3\u5e93\u91cc\u53d6\u4e1c\u897f\u3002<\/p>\n\n\n\n

\u8f93\u51fa\u95e8\uff08Output Gate\uff09<\/h3>\n\n\n\n

\u7528\u6765\u51b3\u5b9a\u4ece\u8bb0\u5fc6\u4e2d\u53d6\u51fa\u54ea\u4e9b\u4fe1\u606f\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u8bb0\u5fc6\u4ed3\u5e93\u5df2\u7ecf\u66f4\u65b0\u5b8c\u6bd5\uff0c\u73b0\u5728\u8981\u4ece\u91cc\u9762\u53d6\u4fe1\u606f\u3002\u8fd9\u4e2a\u65f6\u5019\u8f93\u51fa\u95e8\u5c31\u76f8\u5f53\u4e8e\u4e13\u7ba1\u51fa\u8377\u7684\u201c\u5ba1\u6838\u5458\u201d\uff0c\u5bf9\u4ece\u4ed3\u5e93\u4e2d\u53d6\u51fa\u6765\u7684\u8d27\u7269\u8fdb\u884c\u201c\u628a\u5173\u201d\uff0c\u5c06\u7ecf\u8fc7 Tanh \u6574\u5408\u540e\u7684 \\( {C}_t \\) \u548c \\( o_{t} \\) \u76f8\u4e58\uff0c\u5f97\u5230\u201c\u5ba1\u6838\u201d\u540e\u7684\u51fa\u8377\u8d27\u7269\uff0c\u8fbe\u6210\u8f93\u51fa\u7684\u76ee\u7684\u3002<\/p>\n\n\n\n

\u5c06 \\( {C}_t \\) \u7528 Tanh \u51fd\u6570\u6574\u5408\u539f\u56e0\uff0c\u524d\u9762\u5df2\u7ecf\u63d0\u5230 Tanh \u7684\u4f18\u52bf\uff1b\\( {o}_t \\) \u548c\u524d\u9762\u7684 \\( {i}_t \\) \u3001 \\( {f}_t \\) \u4e00\u6837\uff0c\u7531 Sigmoid \u751f\u6210\u7684\u95e8\u63a7\u4fe1\u53f7\uff0c\u8fd9\u91cc\u5747\u4e0d\u518d\u8d58\u8ff0\u3002<\/p>\n\n\n\n

\u7531\u6b64\u6700\u57fa\u672c\u7684 LSTM \u6a21\u578b\u5c31\u5b66\u4e60\u5b8c\u6bd5\uff0c\u4e09\u9053\u95e8\u8ba9\u6a21\u578b\u80fd\u201c\u6709\u9009\u62e9\u201d\u5730\u5fd8\u8bb0\u3001\u5438\u6536\u548c\u8f93\u51fa\u4fe1\u606f\uff0c\u65e2\u80fd\u6355\u6349\u77ed\u671f\u7684\u201c\u65b0\u4fe1\u606f\u201d\uff0c\u4e5f\u80fd\u8bb0\u5f97\u201c\u5f88\u4e45\u4ee5\u524d\u7684\u91cd\u8981\u4e8b\u4ef6\u201d\u3002<\/p>\n\n\n\n

\u4f46\u5e76\u975e\u6240\u6709 LSTM \u90fd\u4e0e\u4e0a\u8ff0\u7684\u57fa\u672c\u6a21\u578b\u76f8\u540c\uff0c\u6709\u4e9b\u6539\u8fdb\u8fc7\u7684\u6a21\u578b\u66f4\u53d7\u6b22\u8fce\uff0c\u6bd4\u5982\u95e8\u63a7\u5faa\u73af\u5355\u5143\uff08GRU\uff09\u3002<\/strong><\/p>\n\n\n\n

\u95e8\u63a7\u5faa\u73af\u5355\u5143\uff08Gated Recurrent Unit\uff09<\/h2>\n\n\n\n

GRU \u662f LSTM \u7684\u4e00\u79cd\u53d8\u4f53\uff0c\u628a\u9057\u5fd8\u95e8\u548c\u8f93\u5165\u95e8\u5408\u5e76\u6210\u4e00\u4e2a\u201c\u66f4\u65b0\u95e8\uff08Update Gate\uff09\u201d\uff0c\u8fd8\u628a\u8bb0\u5fc6\u548c\u9690\u85cf\u72b6\u6001\u5408\u5e76\u4e86\uff0c\u6ca1\u6709\u4e86\u8f93\u51fa\u95e8\uff0c\u800c\u662f\u589e\u52a0\u4e86\u201c\u91cd\u7f6e\u95e8\uff08Reset Gate\uff09\u201d\uff0c\u548c\u66f4\u65b0\u95e8\u4e00\u8d77\uff0c\u4e00\u4e2a\u8d1f\u8d23\u5199\u5165\uff0c\u4e00\u4e2a\u8d1f\u8d23\u8bfb\u53d6\uff0c\u7528\u8d77\u6765\u66f4\u7b80\u5355\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u91cc\u9762\u53ea\u6709\u4e09\u4e2a\u795e\u7ecf\u7f51\u7edc\u5c42\uff0c\u4e24\u4e2a Sigmoid \u8f93\u51fa\u95e8\u63a7\u4fe1\u53f7\u548c\u4e00\u4e2a Tanh \u5bf9\u4fe1\u606f\u8fdb\u884c\u6574\u5408\u3002<\/p>\n\n\n\n

\u5de6\u8fb9\u7684 \\( {r}_t \\) \u5c31\u662f\u91cd\u7f6e\u95e8\u63a7\u4fe1\u53f7\u4e86\uff0c\u76f8\u5f53\u4e8e\u4ece\u8bb0\u5fc6\u91cc\u9009\u62e9\u6027\u5730\u8bfb\u53d6\u4e00\u90e8\u5206\u4fe1\u606f\uff0c\u4e5f\u76f8\u5f53\u4e8e\u4ece\u4ed3\u5e93\u91cc\u627e\u5408\u683c\u7684\u8d27\u7269\u3002<\/p>\n\n\n\n

\\( {r}_t \\) \u7531\u4e0a\u4e00\u65f6\u523b\u7684\u9690\u85cf\u5c42\u8f93\u51fa \\( h_{t-1} \\) \u548c\u8fd9\u4e00\u65f6\u523b\u7684\u8f93\u5165 \\( x_{t} \\) \u5171\u540c\u51b3\u5b9a\uff0c\u7136\u540e\u53c8\u4f5c\u7528\u5230  \\( h_{t-1} \\) \u4e0a\uff0c\u5f97\u5230\u7ecf\u8fc7\u5ba1\u6838\u7684\u8d27\u7269\uff0c\u52a0\u5230\u65b0\u6765\u7684\u8d27\u7269 \\( x_{t} \\) \u4e0a\uff0c\u4e0e\u65b0\u8d27\u4e00\u8d77\u53c2\u8003\u3002\u518d\u7ecf\u8fc7 Tanh \u8fdb\u884c\u6574\u5408\u5f97\u5230 \\( \\tilde{h}_t \\) \u3002<\/p>\n\n\n\n

\\( {z}_t \\) \u662f\u66f4\u65b0\u95e8\u63a7\u4fe1\u53f7\uff0c\u540c\u65f6\u62c5\u5f53\u7740\u201c\u6e05\u7406\u5de5\u201d\u548c\u201c\u5ba1\u6838\u5458\u201d\u7684\u5de5\u4f5c\u3002\\( {z}_t \\) \u9001\u5165 (1-) \u5206\u652f\u548c \\( {h}_{t-1} \\) \u76f8\u4e58\u76f8\u5f53\u4e8e\u505a\u4e86\u539f\u6765\u7684\u201c\u6e05\u7406\u5de5\u201d\u9057\u5fd8\u95e8\u7684\u5de5\u4f5c\uff0c\u6700\u53f3\u8fb9\u548c \\( \\tilde{h}_t \\) \u76f8\u4e58\uff0c\u7136\u540e\u5f80\u4e0a\u8d70\u548c \\( {h}_{t-1} \\) \u76f8\u52a0\u76f8\u5f53\u4e8e\u505a\u4e86\u8f93\u5165\u95e8\u201c\u5ba1\u6838\u5458\u201d\u7684\u5de5\u4f5c\u3002<\/p>\n\n\n\n

GRU \u5c11\u4e86\u4e00\u7ec4\u95e8\u63a7\u4fe1\u53f7\uff0c\u5c11\u4e86\u4e00\u5c42\u795e\u7ecf\u7f51\u7edc\uff0c\u53c2\u6570\u66f4\u5c11\uff0c\u8ba1\u7b97\u91cf\u4e5f\u66f4\u5c0f\uff0c\u5728\u67d0\u4e9b\u6570\u636e\u96c6\u548c\u4efb\u52a1\u4e0a\uff0cGRU \u7684\u6027\u80fd\u548c LSTM \u76f8\u5f53\uff0c\u4f46\u8bad\u7ec3\u901f\u5ea6\u66f4\u5feb\uff0c\u53ef\u80fd\u5728\u5bf9\u6781\u5176\u957f\u671f\u7684\u590d\u6742\u4f9d\u8d56\u5efa\u6a21\u4e0a\u5f31\u4e8e LSTM\u3002<\/p>\n\n\n\n

RNN \/ LSTM \/ GRU \u6bd4\u8f83<\/h2>\n\n\n\n

\u7ed3\u6784\u590d\u6742\u7a0b\u5ea6\uff08\u8ba1\u7b97\u590d\u6742\u5ea6 \/ \u53c2\u6570\u91cf \/ \u8d44\u6e90\u6d88\u8017\uff09\uff1a LSTM > GRU > RNN<\/p>\n\n\n\n

\u957f\u671f\u4f9d\u8d56\u5efa\u6a21\u6027\u80fd\u6bd4\u8f83\uff1aLSTM \u2248 GRU > RNN<\/p>\n\n\n\n

\u81f3\u4e8e\u4e09\u8005\u7684\u8054\u7cfb\uff0cGRU \u662f LSTM \u7684\u53d8\u4f53\uff0c\u66f4\u6613\u7528\uff0cLSTM \u662f RNN \u7684\u6539\u826f\u7248\u672c\uff0c\u4e3a\u89e3\u51b3 RNN \u7684\u68af\u5ea6\u8870\u51cf\u95ee\u9898\u800c\u8bbe\u8ba1\u3002<\/p>\n\n\n\n

\u73b0\u5728\u4e3b\u8981\u90fd\u5728\u7528 LSTM \u548c\u5b83\u7684\u53d8\u4f53\uff0c\u6027\u80fd\u786e\u5b9e\u6ca1\u5f97\u8bf4\u3002<\/p>\n\n\n\n

\u5b66\u4e60\u5b8c\u57fa\u7840\u7684\u65f6\u5e8f\u6a21\u578b\uff0c\u4e0b\u4e00\u8282\u5c31\u662f\u7ecf\u5178\u7684\u6ce8\u610f\u529b\u673a\u5236\u4e86\u3002<\/p>\n\n\n\n

 <\/p>\n\n\n\n

\u4e0d\u89c4\u8303\u7684\u53c2\u8003\u51fa\u5904\uff1a<\/p>\n\n\n\n

[1] \u53f2\u4e0a\u6700\u8be6\u7ec6\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u8bb2\u89e3\uff08RNN\/LSTM\/GRU\uff09 – \u77e5\u4e4e<\/a><\/p>\n\n\n\n

[2] https:\/\/colah.github.io\/posts\/2015-08-Understanding-LSTMs\/<\/a><\/p>\n\n\n\n

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

\u4e0b\u4e00\u7bc7\uff1a\u4ece 0 \u5f00\u59cb\u7684 Transformer \u68b3\u7406\uff083\uff09\u6ce8\u610f\u529b\u673a\u5236\u4e0e Transformer \u539f\u7406<\/a><\/p>\n\n\n\n

 <\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08Rerrent Neural Network, RNN\uff09 \u4f20\u7edf\u7684\u795e\u7ecf\u7f51\u7edc\u5728\u5904\u7406\u4e0a\u4e00\u6761\u6570\u636e\u548c\u4e0b\u4e00\u6761 […]<\/p>\n","protected":false},"author":311,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_bbp_topic_count":0,"_bbp_reply_count":0,"_bbp_total_topic_count":0,"_bbp_total_reply_count":0,"_bbp_voice_count":0,"_bbp_anonymous_reply_count":0,"_bbp_topic_count_hidden":0,"_bbp_reply_count_hidden":0,"_bbp_forum_subforum_count":0,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2614","post","type-post","status-publish","format-standard","hentry","category-common"],"_links":{"self":[{"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/posts\/2614","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/users\/311"}],"replies":[{"embeddable":true,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2614"}],"version-history":[{"count":12,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/posts\/2614\/revisions"}],"predecessor-version":[{"id":2707,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/posts\/2614\/revisions\/2707"}],"wp:attachment":[{"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2614"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2614"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2614"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}