\u672c\u6587\u8f6c\u8f7d\u81ea\u300c\u673a\u5668\u5b66\u4e60\u70bc\u4e39\u8bb0\u300d\uff0c\u641c\u7d22\u300cjulius-ai\u300d\u5373\u53ef\u5173\u6ce8\u3002 <\/p>\n
\u539f\u6587\u94fe\u63a5\uff1a\u5c0f\u8c61<\/p>\n
\uff08\u4e00\uff09\u4e00\u4e2a\u6846\u67b6\u770b\u61c2\u4f18\u5316\u7b97\u6cd5<\/p>\n
\u673a\u5668\u5b66\u4e60\u754c\u6709\u4e00\u7fa4\u70bc\u4e39\u5e08\uff0c\u4ed6\u4eec\u6bcf\u5929\u7684\u65e5\u5e38\u662f\uff1a<\/p>\n
\u62ff\u6765\u836f\u6750\uff08\u6570\u636e\uff09\uff0c\u67b6\u8d77\u516b\u5366\u7089\uff08\u6a21\u578b\uff09\uff0c\u70b9\u7740\u516d\u5473\u771f\u706b\uff08\u4f18\u5316\u7b97\u6cd5\uff09\uff0c\u5c31\u6447\u7740\u84b2\u6247\u7b49\u7740\u4e39\u836f\u51fa\u7089\u4e86\u3002<\/p>\n
\u4e0d\u8fc7\uff0c\u5f53\u8fc7\u53a8\u5b50\u7684\u90fd\u77e5\u9053\uff0c\u540c\u6837\u7684\u98df\u6750\uff0c\u540c\u6837\u7684\u83dc\u8c31\uff0c\u4f46\u706b\u5019\u4e0d\u4e00\u6837\u4e86\uff0c\u8fd9\u51fa\u6765\u7684\u53e3\u5473\u53ef\u662f\u5343\u5dee\u4e07\u522b\u3002\u706b\u5c0f\u4e86\u5939\u751f\uff0c\u706b\u5927\u4e86\u6613\u7cca\uff0c\u706b\u4e0d\u5300\u5219\u534a\u751f\u534a\u7cca\u3002<\/p>\n
\u673a\u5668\u5b66\u4e60\u4e5f\u662f\u4e00\u6837\uff0c\u6a21\u578b\u4f18\u5316\u7b97\u6cd5\u7684\u9009\u62e9\u76f4\u63a5\u5173\u7cfb\u5230\u6700\u7ec8\u6a21\u578b\u7684\u6027\u80fd\u3002\u6709\u65f6\u5019\u6548\u679c\u4e0d\u597d\uff0c\u672a\u5fc5\u662f\u7279\u5f81\u7684\u95ee\u9898\u6216\u8005\u6a21\u578b\u8bbe\u8ba1\u7684\u95ee\u9898\uff0c\u5f88\u53ef\u80fd\u5c31\u662f\u4f18\u5316\u7b97\u6cd5\u7684\u95ee\u9898\u3002<\/p>\n
\u8bf4\u5230\u4f18\u5316\u7b97\u6cd5\uff0c\u5165\u95e8\u7ea7\u5fc5\u4ece SGD \u5b66\u8d77\uff0c\u8001\u53f8\u673a\u5219\u4f1a\u544a\u8bc9\u4f60\u66f4\u597d\u7684\u8fd8\u6709 AdaGrad \/ AdaDelta\uff0c\u6216\u8005\u76f4\u63a5\u65e0\u8111\u7528 Adam\u3002\u53ef\u662f\u770b\u770b\u5b66\u672f\u754c\u7684\u6700\u65b0 paper\uff0c\u5374\u53d1\u73b0\u4e00\u4f17\u5927\u795e\u8fd8\u5728\u7528\u7740\u5165\u95e8\u7ea7\u7684 SGD\uff0c\u6700\u591a\u52a0\u4e2a Momentum \u6216\u8005 Nesterov\uff0c\u8fd8\u7ecf\u5e38\u4f1a\u9ed1\u4e00\u4e0bAdam\u3002\u6bd4\u5982 UC Berkeley \u7684\u4e00\u7bc7\u8bba\u6587\u5c31\u5728 Conclusion \u4e2d\u5199\u9053\uff1a<\/p>\n
\nDespite the fact that our experimental evidence demonstrates that adaptive methods are not advantageous for machine learning, the Adam algorithm remains incredibly popular. We are not sure exactly as to why \u2026\u2026<\/p>\n<\/blockquote>\n
\u65e0\u5948\u4e0e\u9178\u695a\u4e4b\u60c5\u6ea2\u4e8e\u8a00\u8868\u3002<\/p>\n
\u8fd9\u662f\u4e3a\u4ec0\u4e48\u5462\uff1f\u96be\u9053\u5e73\u5e73\u6de1\u6de1\u624d\u662f\u771f\uff1f<\/p>\n
<\/p>\n
01 \u4e00\u4e2a\u6846\u67b6\u56de\u987e\u4f18\u5316\u7b97\u6cd5<\/p>\n
\u9996\u5148\u6211\u4eec\u6765\u56de\u987e\u4e00\u4e0b\u5404\u7c7b\u4f18\u5316\u7b97\u6cd5\u3002<\/p>\n
\u6df1\u5ea6\u5b66\u4e60\u4f18\u5316\u7b97\u6cd5\u7ecf\u5386\u4e86 SGD -> SGDM -> NAG ->AdaGrad -> AdaDelta -> Adam -> Nadam \u8fd9\u6837\u7684\u53d1\u5c55\u5386\u7a0b\u3002Google\u4e00\u4e0b\u5c31\u53ef\u4ee5\u770b\u5230\u5f88\u591a\u7684\u6559\u7a0b\u6587\u7ae0\uff0c\u8be6\u7ec6\u544a\u8bc9\u4f60\u8fd9\u4e9b\u7b97\u6cd5\u662f\u5982\u4f55\u4e00\u6b65\u4e00\u6b65\u6f14\u53d8\u800c\u6765\u7684\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u6362\u4e00\u4e2a\u601d\u8def\uff0c\u7528\u4e00\u4e2a\u6846\u67b6\u6765\u68b3\u7406\u6240\u6709\u7684\u4f18\u5316\u7b97\u6cd5\uff0c\u505a\u4e00\u4e2a\u66f4\u52a0\u9ad8\u5c4b\u5efa\u74f4\u7684\u5bf9\u6bd4\u3002<\/p>\n
\u4f18\u5316\u7b97\u6cd5\u901a\u7528\u6846\u67b6<\/h3>\n
\u9996\u5148\u5b9a\u4e49\uff1a\u5f85\u4f18\u5316\u53c2\u6570\uff1aw \uff0c\u76ee\u6807\u51fd\u6570\uff1a f(w)\uff0c\u521d\u59cb\u5b66\u4e60\u7387 \u03b1\u3002\u800c\u540e\uff0c\u5f00\u59cb\u8fdb\u884c\u8fed\u4ee3\u4f18\u5316\u3002\u5728\u6bcf\u4e2aepoch t\uff1a<\/p>\n
1. \u8ba1\u7b97\u76ee\u6807\u51fd\u6570\u5173\u4e8e\u5f53\u524d\u53c2\u6570\u7684\u68af\u5ea6\uff1a <\/p>\n
2. \u6839\u636e\u5386\u53f2\u68af\u5ea6\u8ba1\u7b97\u4e00\u9636\u52a8\u91cf\u548c\u4e8c\u9636\u52a8\u91cf\uff1a<\/p>\n
3. \u8ba1\u7b97\u5f53\u524d\u65f6\u523b\u7684\u4e0b\u964d\u68af\u5ea6\uff1a <\/p>\n
4. \u6839\u636e\u4e0b\u964d\u68af\u5ea6\u8fdb\u884c\u66f4\u65b0\uff1a <\/p>\n
\u638c\u63e1\u4e86\u8fd9\u4e2a\u6846\u67b6\uff0c\u4f60\u53ef\u4ee5\u8f7b\u8f7b\u677e\u677e\u8bbe\u8ba1\u81ea\u5df1\u7684\u4f18\u5316\u7b97\u6cd5\u3002<\/p>\n
\u6211\u4eec\u62ff\u7740\u8fd9\u4e2a\u6846\u67b6\uff0c\u6765\u7167\u4e00\u7167\u5404\u79cd\u7384\u4e4e\u5176\u7384\u7684\u4f18\u5316\u7b97\u6cd5\u7684\u771f\u8eab\u3002\u6b65\u9aa43\u30014\u5bf9\u4e8e\u5404\u4e2a\u7b97\u6cd5\u90fd\u662f\u4e00\u81f4\u7684\uff0c\u4e3b\u8981\u7684\u5dee\u522b\u5c31\u4f53\u73b0\u57281\u548c2\u4e0a\u3002<\/p>\n
02 \u56fa\u5b9a\u5b66\u4e60\u7387\u7684\u4f18\u5316\u7b97\u6cd5<\/p>\n
SGD<\/h3>\n
\u5148\u6765\u770bSGD\u3002SGD\u6ca1\u6709\u52a8\u91cf\u7684\u6982\u5ff5\uff0c\u4e5f\u5c31\u662f\u8bf4\uff1a<\/p>\n
\u4ee3\u5165\u6b65\u9aa43\uff0c\u53ef\u4ee5\u770b\u5230\u4e0b\u964d\u68af\u5ea6\u5c31\u662f\u6700\u7b80\u5355\u7684<\/p>\n
SGD\u6700\u5927\u7684\u7f3a\u70b9\u662f\u4e0b\u964d\u901f\u5ea6\u6162\uff0c\u800c\u4e14\u53ef\u80fd\u4f1a\u5728\u6c9f\u58d1\u7684\u4e24\u8fb9\u6301\u7eed\u9707\u8361\uff0c\u505c\u7559\u5728\u4e00\u4e2a\u5c40\u90e8\u6700\u4f18\u70b9\u3002<\/p>\n
SGD with Momentum<\/h3>\n
\u4e3a\u4e86\u6291\u5236SGD\u7684\u9707\u8361\uff0cSGDM\u8ba4\u4e3a\u68af\u5ea6\u4e0b\u964d\u8fc7\u7a0b\u53ef\u4ee5\u52a0\u5165\u60ef\u6027\u3002\u4e0b\u5761\u7684\u65f6\u5019\uff0c\u5982\u679c\u53d1\u73b0\u662f\u9661\u5761\uff0c\u90a3\u5c31\u5229\u7528\u60ef\u6027\u8dd1\u7684\u5feb\u4e00\u4e9b\u3002SGDM\u5168\u79f0\u662fSGD with momentum\uff0c\u5728SGD\u57fa\u7840\u4e0a\u5f15\u5165\u4e86\u4e00\u9636\u52a8\u91cf\uff1a<\/p>\n
\u4e00\u9636\u52a8\u91cf\u662f\u5404\u4e2a\u65f6\u523b\u68af\u5ea6\u65b9\u5411\u7684\u6307\u6570\u79fb\u52a8\u5e73\u5747\u503c\uff0c\u7ea6\u7b49\u4e8e\u6700\u8fd1 1\/(1-\u03b21) \u4e2a\u65f6\u523b\u7684\u68af\u5ea6\u5411\u91cf\u548c\u7684\u5e73\u5747\u503c\u3002<\/p>\n
\u4e5f\u5c31\u662f\u8bf4\uff0ct \u65f6\u523b\u7684\u4e0b\u964d\u65b9\u5411\uff0c\u4e0d\u4ec5\u7531\u5f53\u524d\u70b9\u7684\u68af\u5ea6\u65b9\u5411\u51b3\u5b9a\uff0c\u800c\u4e14\u7531\u6b64\u524d\u7d2f\u79ef\u7684\u4e0b\u964d\u65b9\u5411\u51b3\u5b9a\u3002\u03b21\u7684\u7ecf\u9a8c\u503c\u4e3a0.9\uff0c\u8fd9\u5c31\u610f\u5473\u7740\u4e0b\u964d\u65b9\u5411\u4e3b\u8981\u662f\u6b64\u524d\u7d2f\u79ef\u7684\u4e0b\u964d\u65b9\u5411\uff0c\u5e76\u7565\u5fae\u504f\u5411\u5f53\u524d\u65f6\u523b\u7684\u4e0b\u964d\u65b9\u5411\u3002\u60f3\u8c61\u9ad8\u901f\u516c\u8def\u4e0a\u6c7d\u8f66\u8f6c\u5f2f\uff0c\u5728\u9ad8\u901f\u5411\u524d\u7684\u540c\u65f6\u7565\u5fae\u504f\u5411\uff0c\u6025\u8f6c\u5f2f\u53ef\u662f\u8981\u51fa\u4e8b\u7684\u3002<\/p>\n
SGD with Nesterov Acceleration<\/h3>\n
SGD \u8fd8\u6709\u4e00\u4e2a\u95ee\u9898\u662f\u56f0\u5728\u5c40\u90e8\u6700\u4f18\u7684\u6c9f\u58d1\u91cc\u9762\u9707\u8361\u3002\u60f3\u8c61\u4e00\u4e0b\u4f60\u8d70\u5230\u4e00\u4e2a\u76c6\u5730\uff0c\u56db\u5468\u90fd\u662f\u7565\u9ad8\u7684\u5c0f\u5c71\uff0c\u4f60\u89c9\u5f97\u6ca1\u6709\u4e0b\u5761\u7684\u65b9\u5411\uff0c\u90a3\u5c31\u53ea\u80fd\u5f85\u5728\u8fd9\u91cc\u4e86\u3002\u53ef\u662f\u5982\u679c\u4f60\u722c\u4e0a\u9ad8\u5730\uff0c\u5c31\u4f1a\u53d1\u73b0\u5916\u9762\u7684\u4e16\u754c\u8fd8\u5f88\u5e7f\u9614\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4e0d\u80fd\u505c\u7559\u5728\u5f53\u524d\u4f4d\u7f6e\u53bb\u89c2\u5bdf\u672a\u6765\u7684\u65b9\u5411\uff0c\u800c\u8981\u5411\u524d\u4e00\u6b65\u3001\u591a\u770b\u4e00\u6b65\u3001\u770b\u8fdc\u4e00\u4e9b\u3002<\/p>\n
<\/p>\n
(source: http:\/\/cs231n.github.io\/neural-networks-3)<\/p>\n
\u8fd9\u4e00\u65b9\u6cd5\u4e5f\u79f0\u4e3aNAG\uff0c\u5373 Nesterov Accelerated Gradient\uff0c\u662f\u5728SGD\u3001SGD-M \u7684\u57fa\u7840\u4e0a\u7684\u8fdb\u4e00\u6b65\u6539\u8fdb\uff0c\u6539\u8fdb\u70b9\u5728\u4e8e\u6b65\u9aa4 1\u3002\u6211\u4eec\u77e5\u9053\u5728\u65f6\u523b t \u7684\u4e3b\u8981\u4e0b\u964d\u65b9\u5411\u662f\u7531\u7d2f\u79ef\u52a8\u91cf\u51b3\u5b9a\u7684\uff0c\u81ea\u5df1\u7684\u68af\u5ea6\u65b9\u5411\u8bf4\u4e86\u4e5f\u4e0d\u7b97\uff0c\u90a3\u4e0e\u5176\u770b\u5f53\u524d\u68af\u5ea6\u65b9\u5411\uff0c\u4e0d\u5982\u5148\u770b\u770b\u5982\u679c\u8ddf\u7740\u7d2f\u79ef\u52a8\u91cf\u8d70\u4e86\u4e00\u6b65\uff0c\u90a3\u4e2a\u65f6\u5019\u518d\u600e\u4e48\u8d70\u3002\u56e0\u6b64\uff0cNAG\u5728\u6b65\u9aa4 1\uff0c\u4e0d\u8ba1\u7b97\u5f53\u524d\u4f4d\u7f6e\u7684\u68af\u5ea6\u65b9\u5411\uff0c\u800c\u662f\u8ba1\u7b97\u5982\u679c\u6309\u7167\u7d2f\u79ef\u52a8\u91cf\u8d70\u4e86\u4e00\u6b65\uff0c\u90a3\u4e2a\u65f6\u5019\u7684\u4e0b\u964d\u65b9\u5411\uff1a <\/p>\n
\u7136\u540e\u7528\u4e0b\u4e00\u4e2a\u70b9\u7684\u68af\u5ea6\u65b9\u5411\uff0c\u4e0e\u5386\u53f2\u7d2f\u79ef\u52a8\u91cf\u76f8\u7ed3\u5408\uff0c\u8ba1\u7b97\u6b65\u9aa4 2 \u4e2d\u5f53\u524d\u65f6\u523b\u7684\u7d2f\u79ef\u52a8\u91cf\u3002<\/p>\n
03 \u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7684\u4f18\u5316\u7b97\u6cd5<\/p>\n
\u6b64\u524d\u6211\u4eec\u90fd\u6ca1\u6709\u7528\u5230\u4e8c\u9636\u52a8\u91cf\u3002\u4e8c\u9636\u52a8\u91cf\u7684\u51fa\u73b0\uff0c\u624d\u610f\u5473\u7740\u201c\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u201d\u4f18\u5316\u7b97\u6cd5\u65f6\u4ee3\u7684\u5230\u6765\u3002SGD\u53ca\u5176\u53d8\u79cd\u4ee5\u540c\u6837\u7684\u5b66\u4e60\u7387\u66f4\u65b0\u6bcf\u4e2a\u53c2\u6570\uff0c\u4f46\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u5f80\u5f80\u5305\u542b\u5927\u91cf\u7684\u53c2\u6570\uff0c\u8fd9\u4e9b\u53c2\u6570\u5e76\u4e0d\u662f\u603b\u4f1a\u7528\u5f97\u5230\uff08\u60f3\u60f3\u5927\u89c4\u6a21\u7684embedding\uff09\u3002<\/p>\n
\u5bf9\u4e8e\u7ecf\u5e38\u66f4\u65b0\u7684\u53c2\u6570\uff0c\u6211\u4eec\u5df2\u7ecf\u79ef\u7d2f\u4e86\u5927\u91cf\u5173\u4e8e\u5b83\u7684\u77e5\u8bc6\uff0c\u4e0d\u5e0c\u671b\u88ab\u5355\u4e2a\u6837\u672c\u5f71\u54cd\u592a\u5927\uff0c\u5e0c\u671b\u5b66\u4e60\u901f\u7387\u6162\u4e00\u4e9b\uff1b\u5bf9\u4e8e\u5076\u5c14\u66f4\u65b0\u7684\u53c2\u6570\uff0c\u6211\u4eec\u4e86\u89e3\u7684\u4fe1\u606f\u592a\u5c11\uff0c\u5e0c\u671b\u80fd\u4ece\u6bcf\u4e2a\u5076\u7136\u51fa\u73b0\u7684\u6837\u672c\u8eab\u4e0a\u591a\u5b66\u4e00\u4e9b\uff0c\u5373\u5b66\u4e60\u901f\u7387\u5927\u4e00\u4e9b\u3002<\/p>\n
AdaGrad<\/h3>\n
\u600e\u4e48\u6837\u53bb\u5ea6\u91cf\u5386\u53f2\u66f4\u65b0\u9891\u7387\u5462\uff1f\u90a3\u5c31\u662f\u4e8c\u9636\u52a8\u91cf\u2014\u2014\u8be5\u7ef4\u5ea6\u4e0a\uff0c\u8fc4\u4eca\u4e3a\u6b62\u6240\u6709\u68af\u5ea6\u503c\u7684\u5e73\u65b9\u548c\uff1a<\/p>\n
\u6211\u4eec\u518d\u56de\u987e\u4e00\u4e0b\u6b65\u9aa43\u4e2d\u7684\u4e0b\u964d\u68af\u5ea6\uff1a<\/p>\n
\u53ef\u4ee5\u770b\u51fa\uff0c\u6b64\u65f6\u5b9e\u8d28\u4e0a\u7684\u5b66\u4e60\u7387\u7531\u53d8\u6210\u4e86\u3002 \u4e00\u822c\u4e3a\u4e86\u907f\u514d\u5206\u6bcd\u4e3a0\uff0c\u4f1a\u5728\u5206\u6bcd\u4e0a\u52a0\u4e00\u4e2a\u5c0f\u7684\u5e73\u6ed1\u9879\u3002\u56e0\u6b64\u662f\u6052\u5927\u4e8e0\u7684\uff0c\u800c\u4e14\u53c2\u6570\u66f4\u65b0\u8d8a\u9891\u7e41\uff0c\u4e8c\u9636\u52a8\u91cf\u8d8a\u5927\uff0c\u5b66\u4e60\u7387\u5c31\u8d8a\u5c0f\u3002<\/p>\n
\u8fd9\u4e00\u65b9\u6cd5\u5728\u7a00\u758f\u6570\u636e\u573a\u666f\u4e0b\u8868\u73b0\u975e\u5e38\u597d\u3002\u4f46\u4e5f\u5b58\u5728\u4e00\u4e9b\u95ee\u9898\uff1a\u56e0\u4e3a\u662f\u5355\u8c03\u9012\u589e\u7684\uff0c\u4f1a\u4f7f\u5f97\u5b66\u4e60\u7387\u5355\u8c03\u9012\u51cf\u81f30\uff0c\u53ef\u80fd\u4f1a\u4f7f\u5f97\u8bad\u7ec3\u8fc7\u7a0b\u63d0\u524d\u7ed3\u675f\uff0c\u5373\u4fbf\u540e\u7eed\u8fd8\u6709\u6570\u636e\u4e5f\u65e0\u6cd5\u5b66\u5230\u5fc5\u8981\u7684\u77e5\u8bc6\u3002<\/p>\n
AdaDelta \/ RMSProp<\/h3>\n
\u7531\u4e8eAdaGrad\u5355\u8c03\u9012\u51cf\u7684\u5b66\u4e60\u7387\u53d8\u5316\u8fc7\u4e8e\u6fc0\u8fdb\uff0c\u6211\u4eec\u8003\u8651\u4e00\u4e2a\u6539\u53d8\u4e8c\u9636\u52a8\u91cf\u8ba1\u7b97\u65b9\u6cd5\u7684\u7b56\u7565\uff1a\u4e0d\u7d2f\u79ef\u5168\u90e8\u5386\u53f2\u68af\u5ea6\uff0c\u800c\u53ea\u5173\u6ce8\u8fc7\u53bb\u4e00\u6bb5\u65f6\u95f4\u7a97\u53e3\u7684\u4e0b\u964d\u68af\u5ea6\u3002\u8fd9\u4e5f\u5c31\u662fAdaDelta\u540d\u79f0\u4e2dDelta\u7684\u6765\u5386\u3002<\/p>\n
\u4fee\u6539\u7684\u601d\u8def\u5f88\u7b80\u5355\u3002\u524d\u9762\u6211\u4eec\u8bb2\u5230\uff0c\u6307\u6570\u79fb\u52a8\u5e73\u5747\u503c\u5927\u7ea6\u5c31\u662f\u8fc7\u53bb\u4e00\u6bb5\u65f6\u95f4\u7684\u5e73\u5747\u503c\uff0c\u56e0\u6b64\u6211\u4eec\u7528\u8fd9\u4e00\u65b9\u6cd5\u6765\u8ba1\u7b97\u4e8c\u9636\u7d2f\u79ef\u52a8\u91cf\uff1a<\/p>\n
\u8fd9\u5c31\u907f\u514d\u4e86\u4e8c\u9636\u52a8\u91cf\u6301\u7eed\u7d2f\u79ef\u3001\u5bfc\u81f4\u8bad\u7ec3\u8fc7\u7a0b\u63d0\u524d\u7ed3\u675f\u7684\u95ee\u9898\u4e86\u3002<\/p>\n
Adam<\/h3>\n
\u8c08\u5230\u8fd9\u91cc\uff0cAdam\u548cNadam\u7684\u51fa\u73b0\u5c31\u5f88\u81ea\u7136\u800c\u7136\u4e86\u2014\u2014\u5b83\u4eec\u662f\u524d\u8ff0\u65b9\u6cd5\u7684\u96c6\u5927\u6210\u8005\u3002\u6211\u4eec\u770b\u5230\uff0cSGD-M\u5728SGD\u57fa\u7840\u4e0a\u589e\u52a0\u4e86\u4e00\u9636\u52a8\u91cf\uff0cAdaGrad\u548cAdaDelta\u5728SGD\u57fa\u7840\u4e0a\u589e\u52a0\u4e86\u4e8c\u9636\u52a8\u91cf\u3002\u628a\u4e00\u9636\u52a8\u91cf\u548c\u4e8c\u9636\u52a8\u91cf\u90fd\u7528\u8d77\u6765\uff0c\u5c31\u662fAdam\u4e86\u2014\u2014Adaptive + Momentum\u3002<\/p>\n
SGD\u7684\u4e00\u9636\u52a8\u91cf\uff1a<\/p>\n
\u52a0\u4e0aAdaDelta\u7684\u4e8c\u9636\u52a8\u91cf\uff1a<\/p>\n
\u4f18\u5316\u7b97\u6cd5\u91cc\u6700\u5e38\u89c1\u7684\u4e24\u4e2a\u8d85\u53c2\u6570\u5c31\u90fd\u5728\u8fd9\u91cc\u4e86\uff0c\u524d\u8005\u63a7\u5236\u4e00\u9636\u52a8\u91cf\uff0c\u540e\u8005\u63a7\u5236\u4e8c\u9636\u52a8\u91cf\u3002<\/p>\n
Nadam<\/h3>\n
\u6700\u540e\u662fNadam\u3002\u6211\u4eec\u8bf4Adam\u662f\u96c6\u5927\u6210\u8005\uff0c\u4f46\u5b83\u5c45\u7136\u9057\u6f0f\u4e86Nesterov\uff0c\u8fd9\u8fd8\u80fd\u5fcd\uff1f\u5fc5\u987b\u7ed9\u5b83\u52a0\u4e0a\u2014\u2014\u53ea\u9700\u8981\u6309\u7167NAG\u7684\u6b65\u9aa41\u6765\u8ba1\u7b97\u68af\u5ea6\uff1a<\/p>\n
\u8fd9\u5c31\u662fNesterov + Adam = Nadam\u4e86\u3002<\/p>\n
\u8bf4\u5230\u8fd9\u91cc\uff0c\u5927\u6982\u53ef\u4ee5\u7406\u89e3\u4e3a\u4ec0\u4e48\u8bf4 Adam \/ Nadam \u76ee\u524d\u6700\u4e3b\u6d41\u3001\u6700\u597d\u7528\u7684\u7b97\u6cd5\u4e86\u3002\u65e0\u8111\u7528Adam\/Nadam\uff0c\u6536\u655b\u901f\u5ea6\u55d6\u55d6\u6ef4\uff0c\u6548\u679c\u4e5f\u662f\u6760\u6760\u6ef4\u3002<\/p>\n
\u90a3\u4e3a\u4ec0\u4e48Adam\u8fd8\u8001\u62db\u4eba\u9ed1\uff0c\u88ab\u5b66\u672f\u754c\u4e00\u987f\u9119\u5937\uff1f\u96be\u9053\u53ea\u662f\u4e3a\u4e86\u53d1paper\u704c\u6c34\u5417\uff1f<\/p>\n
\uff08\u4e8c\uff09Adam\u7684\u4e24\u5b97\u7f6a<\/p>\n
\u4ee5\u4e0a\u5185\u5bb9\u4e2d\uff0c\u6211\u4eec\u7528\u4e00\u4e2a\u6846\u67b6\u6765\u56de\u987e\u4e86\u4e3b\u6d41\u7684\u6df1\u5ea6\u5b66\u4e60\u4f18\u5316\u7b97\u6cd5\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u4e00\u4ee3\u53c8\u4e00\u4ee3\u7684\u7814\u7a76\u8005\u4eec\u4e3a\u4e86\u6211\u4eec\u80fd\u70bc\uff08xun\uff09\u597d\uff08hao\uff09\u91d1\uff08mo\uff09\u4e39\uff08xing\uff09\u53ef\u8c13\u662f\u715e\u8d39\u82e6\u5fc3\u3002\u4ece\u7406\u8bba\u4e0a\u770b\uff0c\u4e00\u4ee3\u66f4\u6bd4\u4e00\u4ee3\u5b8c\u5584\uff0cAdam\/Nadam\u5df2\u7ecf\u767b\u5cf0\u9020\u6781\u4e86\uff0c\u4e3a\u4ec0\u4e48\u5927\u5bb6\u8fd8\u662f\u4e0d\u5fd8\u521d\u5fc3SGD\u5462\uff1f<\/p>\n
\u4e3e\u4e2a\u6817\u5b50\u3002\u5f88\u591a\u5e74\u4ee5\u524d\uff0c\u6444\u5f71\u79bb\u666e\u7f57\u5927\u4f17\u975e\u5e38\u9065\u8fdc\u3002\u5341\u5e74\u524d\uff0c\u50bb\u74dc\u76f8\u673a\u5f00\u59cb\u98ce\u9761\uff0c\u6e38\u5ba2\u51e0\u4e4e\u4eba\u624b\u4e00\u4e2a\u3002\u667a\u80fd\u624b\u673a\u51fa\u73b0\u4ee5\u540e\uff0c\u6444\u5f71\u66f4\u662f\u8d70\u8fdb\u5343\u5bb6\u4e07\u6237\uff0c\u624b\u673a\u968f\u624b\u4e00\u62cd\uff0c\u524d\u540e\u4e24\u5343\u4e07\uff0c\u7167\u4eae\u4f60\u7684\u7f8e\uff08\u54a6\uff0c\u8fd9\u662f\u4ec0\u4e48\u4e71\u4e03\u516b\u7cdf\u7684\uff09\u3002\u4f46\u662f\u4e13\u4e1a\u6444\u5f71\u5e08\u8fd8\u662f\u559c\u6b22\u7528\u5355\u53cd\uff0c\u5b5c\u5b5c\u4e0d\u5026\u5730\u8c03\u5149\u5708\u3001\u5feb\u95e8\u3001ISO\u3001\u767d\u5e73\u8861\u2026\u2026\u4e00\u5806\u81ea\u62cd\u515a\u4ece\u4e0dcare\u7684\u540d\u8bcd\u3002\u6280\u672f\u7684\u8fdb\u6b65\uff0c\u4f7f\u5f97\u50bb\u74dc\u5f0f\u64cd\u4f5c\u5c31\u53ef\u4ee5\u5f97\u5230\u4e0d\u9519\u7684\u6548\u679c\uff0c\u4f46\u662f\u5728\u7279\u5b9a\u7684\u573a\u666f\u4e0b\uff0c\u8981\u62cd\u51fa\u6700\u597d\u7684\u6548\u679c\uff0c\u4f9d\u7136\u9700\u8981\u6df1\u5165\u5730\u7406\u89e3\u5149\u7ebf\u3001\u7406\u89e3\u7ed3\u6784\u3001\u7406\u89e3\u5668\u6750\u3002<\/p>\n
\u4f18\u5316\u7b97\u6cd5\u5927\u62b5\u4e5f\u5982\u6b64\u3002\u5728\u4e0a\u4e00\u7bc7\u4e2d\uff0c\u6211\u4eec\u7528\u540c\u4e00\u4e2a\u6846\u67b6\u8ba9\u5404\u7c7b\u7b97\u6cd5\u5bf9\u53f7\u5165\u5ea7\u3002\u53ef\u4ee5\u770b\u51fa\uff0c\u5927\u5bb6\u90fd\u662f\u6b8a\u9014\u540c\u5f52\uff0c\u53ea\u662f\u76f8\u5f53\u4e8e\u5728SGD\u57fa\u7840\u4e0a\u589e\u52a0\u4e86\u5404\u7c7b\u5b66\u4e60\u7387\u7684\u4e3b\u52a8\u63a7\u5236\u3002\u5982\u679c\u4e0d\u60f3\u505a\u7cbe\u7ec6\u7684\u8c03\u4f18\uff0c\u90a3\u4e48Adam\u663e\u7136\u6700\u4fbf\u4e8e\u76f4\u63a5\u62ff\u6765\u4e0a\u624b\u3002<\/p>\n
\u4f46\u8fd9\u6837\u7684\u50bb\u74dc\u5f0f\u64cd\u4f5c\u5e76\u4e0d\u4e00\u5b9a\u80fd\u591f\u9002\u5e94\u6240\u6709\u7684\u573a\u5408\u3002\u5982\u679c\u80fd\u591f\u6df1\u5165\u4e86\u89e3\u6570\u636e\uff0c\u7814\u7a76\u5458\u4eec\u53ef\u4ee5\u66f4\u52a0\u81ea\u5982\u5730\u63a7\u5236\u4f18\u5316\u8fed\u4ee3\u7684\u5404\u7c7b\u53c2\u6570\uff0c\u5b9e\u73b0\u66f4\u597d\u7684\u6548\u679c\u4e5f\u5e76\u4e0d\u5947\u602a\u3002\u6bd5\u7adf\uff0c\u7cbe\u8c03\u7684\u53c2\u6570\u8fd8\u6bd4\u4e0d\u8fc7\u50bb\u74dc\u5f0f\u7684SGD\uff0c\u65e0\u7591\u662f\u5728\u6311\u6218\u9876\u7ea7\u7814\u7a76\u5458\u4eec\u7684\u70bc\u4e39\u7ecf\u9a8c\uff01<\/p>\n
\u6700\u8fd1\uff0c\u4e0d\u5c11paper\u5f00\u603cAdam\uff0c\u6211\u4eec\u7b80\u5355\u770b\u770b\u90fd\u5728\u8bf4\u4ec0\u4e48\uff1a<\/p>\n
04 Adam\u7f6a\u72b6\u4e00\uff1a\u53ef\u80fd\u4e0d\u6536\u655b<\/p>\n
\u8fd9\u7bc7\u662f\u6b63\u5728\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u9876\u7ea7\u4f1a\u8bae\u4e4b\u4e00 ICLR 2018 \u533f\u540d\u5ba1\u7a3f\u4e2d\u7684\u4e00\u7bc7\u8bba\u6587\u300aOn the Convergence of Adam and Beyond\u300b\uff0c\u63a2\u8ba8\u4e86Adam\u7b97\u6cd5\u7684\u6536\u655b\u6027\uff0c\u901a\u8fc7\u53cd\u4f8b\u8bc1\u660e\u4e86Adam\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u53ef\u80fd\u4f1a\u4e0d\u6536\u655b\u3002<\/p>\n
\u56de\u5fc6\u4e00\u4e0b\u4e0a\u6587\u63d0\u5230\u7684\u5404\u5927\u4f18\u5316\u7b97\u6cd5\u7684\u5b66\u4e60\u7387\uff1a<\/p>\n
<\/p>\n
\u5176\u4e2d\uff0cSGD\u6ca1\u6709\u7528\u5230\u4e8c\u9636\u52a8\u91cf\uff0c\u56e0\u6b64\u5b66\u4e60\u7387\u662f\u6052\u5b9a\u7684\uff08\u5b9e\u9645\u4f7f\u7528\u8fc7\u7a0b\u4e2d\u4f1a\u91c7\u7528\u5b66\u4e60\u7387\u8870\u51cf\u7b56\u7565\uff0c\u56e0\u6b64\u5b66\u4e60\u7387\u9012\u51cf\uff09\u3002AdaGrad\u7684\u4e8c\u9636\u52a8\u91cf\u4e0d\u65ad\u7d2f\u79ef\uff0c\u5355\u8c03\u9012\u589e\uff0c\u56e0\u6b64\u5b66\u4e60\u7387\u662f\u5355\u8c03\u9012\u51cf\u7684\u3002\u56e0\u6b64\uff0c\u8fd9\u4e24\u7c7b\u7b97\u6cd5\u4f1a\u4f7f\u5f97\u5b66\u4e60\u7387\u4e0d\u65ad\u9012\u51cf\uff0c\u6700\u7ec8\u6536\u655b\u52300\uff0c\u6a21\u578b\u4e5f\u5f97\u4ee5\u6536\u655b\u3002<\/p>\n
\u4f46AdaDelta\u548cAdam\u5219\u4e0d\u7136\u3002\u4e8c\u9636\u52a8\u91cf\u662f\u56fa\u5b9a\u65f6\u95f4\u7a97\u53e3\u5185\u7684\u7d2f\u79ef\uff0c\u968f\u7740\u65f6\u95f4\u7a97\u53e3\u7684\u53d8\u5316\uff0c\u9047\u5230\u7684\u6570\u636e\u53ef\u80fd\u53d1\u751f\u5de8\u53d8\uff0c\u4f7f\u5f97 \u53ef\u80fd\u4f1a\u65f6\u5927\u65f6\u5c0f\uff0c\u4e0d\u662f\u5355\u8c03\u53d8\u5316\u3002\u8fd9\u5c31\u53ef\u80fd\u5728\u8bad\u7ec3\u540e\u671f\u5f15\u8d77\u5b66\u4e60\u7387\u7684\u9707\u8361\uff0c\u5bfc\u81f4\u6a21\u578b\u65e0\u6cd5\u6536\u655b\u3002<\/p>\n
\u8fd9\u7bc7\u6587\u7ae0\u4e5f\u7ed9\u51fa\u4e86\u4e00\u4e2a\u4fee\u6b63\u7684\u65b9\u6cd5\u3002\u7531\u4e8eAdam\u4e2d\u7684\u5b66\u4e60\u7387\u4e3b\u8981\u662f\u7531\u4e8c\u9636\u52a8\u91cf\u63a7\u5236\u7684\uff0c\u4e3a\u4e86\u4fdd\u8bc1\u7b97\u6cd5\u7684\u6536\u655b\uff0c\u53ef\u4ee5\u5bf9\u4e8c\u9636\u52a8\u91cf\u7684\u53d8\u5316\u8fdb\u884c\u63a7\u5236\uff0c\u907f\u514d\u4e0a\u4e0b\u6ce2\u52a8\u3002<\/p>\n
\u901a\u8fc7\u8fd9\u6837\u4fee\u6539\uff0c\u5c31\u4fdd\u8bc1\u4e86 <\/p>\n
\u4ece\u800c\u4f7f\u5f97\u5b66\u4e60\u7387\u5355\u8c03\u9012\u51cf\u3002<\/p>\n
05 Adam\u7f6a\u72b6\u4e8c\uff1a\u53ef\u80fd\u9519\u8fc7\u5168\u5c40\u6700\u4f18\u89e3<\/p>\n
\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u5f80\u5f80\u5305\u542b\u5927\u91cf\u7684\u53c2\u6570\uff0c\u5728\u8fd9\u6837\u4e00\u4e2a\u7ef4\u5ea6\u6781\u9ad8\u7684\u7a7a\u95f4\u5185\uff0c\u975e\u51f8\u7684\u76ee\u6807\u51fd\u6570\u5f80\u5f80\u8d77\u8d77\u4f0f\u4f0f\uff0c\u62e5\u6709\u65e0\u6570\u4e2a\u9ad8\u5730\u548c\u6d3c\u5730\u3002\u6709\u7684\u662f\u9ad8\u5cf0\uff0c\u901a\u8fc7\u5f15\u5165\u52a8\u91cf\u53ef\u80fd\u5f88\u5bb9\u6613\u8d8a\u8fc7\uff1b\u4f46\u6709\u4e9b\u662f\u9ad8\u539f\uff0c\u53ef\u80fd\u63a2\u7d22\u5f88\u591a\u6b21\u90fd\u51fa\u4e0d\u6765\uff0c\u4e8e\u662f\u505c\u6b62\u4e86\u8bad\u7ec3\u3002<\/p>\n
\u4e4b\u524d\uff0cArxiv\u4e0a\u7684\u4e24\u7bc7\u6587\u7ae0\u8c08\u5230\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n
\u7b2c\u4e00\u7bc7\u5c31\u662f\u524d\u6587\u63d0\u5230\u7684\u5410\u69fdAdam\u6700\u72e0\u7684UC Berkeley\u7684\u6587\u7ae0\u300aThe Marginal Value of Adaptive Gradient Methods in Machine Learning\u300b\u3002\u6587\u4e2d\u8bf4\u5230\uff0c\u540c\u6837\u7684\u4e00\u4e2a\u4f18\u5316\u95ee\u9898\uff0c\u4e0d\u540c\u7684\u4f18\u5316\u7b97\u6cd5\u53ef\u80fd\u4f1a\u627e\u5230\u4e0d\u540c\u7684\u7b54\u6848\uff0c\u4f46\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7684\u7b97\u6cd5\u5f80\u5f80\u627e\u5230\u975e\u5e38\u5dee\u7684\u7b54\u6848\uff08very poor solution\uff09\u3002\u4ed6\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u7279\u5b9a\u7684\u6570\u636e\u4f8b\u5b50\uff0c\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7b97\u6cd5\u53ef\u80fd\u4f1a\u5bf9\u524d\u671f\u51fa\u73b0\u7684\u7279\u5f81\u8fc7\u62df\u5408\uff0c\u540e\u671f\u624d\u51fa\u73b0\u7684\u7279\u5f81\u5f88\u96be\u7ea0\u6b63\u524d\u671f\u7684\u62df\u5408\u6548\u679c\u3002\u4f46\u8fd9\u4e2a\u6587\u7ae0\u7ed9\u7684\u4f8b\u5b50\u5f88\u6781\u7aef\uff0c\u5728\u5b9e\u9645\u60c5\u51b5\u4e2d\u672a\u5fc5\u4f1a\u51fa\u73b0\u3002<\/p>\n
\u53e6\u5916\u4e00\u7bc7\u662f\u300aImproving Generalization Performance by Switching from Adam to SGD\u300b\uff0c\u8fdb\u884c\u4e86\u5b9e\u9a8c\u9a8c\u8bc1\u3002\u4ed6\u4eecCIFAR-10\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u6d4b\u8bd5\uff0cAdam\u7684\u6536\u655b\u901f\u5ea6\u6bd4SGD\u8981\u5feb\uff0c\u4f46\u6700\u7ec8\u6536\u655b\u7684\u7ed3\u679c\u5e76\u6ca1\u6709SGD\u597d\u3002\u4ed6\u4eec\u8fdb\u4e00\u6b65\u5b9e\u9a8c\u53d1\u73b0\uff0c\u4e3b\u8981\u662f\u540e\u671fAdam\u7684\u5b66\u4e60\u7387\u592a\u4f4e\uff0c\u5f71\u54cd\u4e86\u6709\u6548\u7684\u6536\u655b\u3002\u4ed6\u4eec\u8bd5\u7740\u5bf9Adam\u7684\u5b66\u4e60\u7387\u7684\u4e0b\u754c\u8fdb\u884c\u63a7\u5236\uff0c\u53d1\u73b0\u6548\u679c\u597d\u4e86\u5f88\u591a\u3002<\/p>\n
\u4e8e\u662f\u4ed6\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u7528\u6765\u6539\u8fdbAdam\u7684\u65b9\u6cd5\uff1a\u524d\u671f\u7528Adam\uff0c\u4eab\u53d7Adam\u5feb\u901f\u6536\u655b\u7684\u4f18\u52bf\uff1b\u540e\u671f\u5207\u6362\u5230SGD\uff0c\u6162\u6162\u5bfb\u627e\u6700\u4f18\u89e3\u3002\u8fd9\u4e00\u65b9\u6cd5\u4ee5\u524d\u4e5f\u88ab\u7814\u7a76\u8005\u4eec\u7528\u5230\uff0c\u4e0d\u8fc7\u4e3b\u8981\u662f\u6839\u636e\u7ecf\u9a8c\u6765\u9009\u62e9\u5207\u6362\u7684\u65f6\u673a\u548c\u5207\u6362\u540e\u7684\u5b66\u4e60\u7387\u3002\u8fd9\u7bc7\u6587\u7ae0\u628a\u8fd9\u4e00\u5207\u6362\u8fc7\u7a0b\u50bb\u74dc\u5316\uff0c\u7ed9\u51fa\u4e86\u5207\u6362SGD\u7684\u65f6\u673a\u9009\u62e9\u65b9\u6cd5\uff0c\u4ee5\u53ca\u5b66\u4e60\u7387\u7684\u8ba1\u7b97\u65b9\u6cd5\uff0c\u6548\u679c\u770b\u8d77\u6765\u4e5f\u4e0d\u9519\u3002<\/p>\n
\u8fd9\u4e2a\u7b97\u6cd5\u633a\u6709\u8da3\uff0c\u4e0b\u4e00\u7bc7\u6211\u4eec\u53ef\u4ee5\u6765\u8c08\u8c08\uff0c\u8fd9\u91cc\u5148\u8d34\u4e2a\u7b97\u6cd5\u6846\u67b6\u56fe\uff1a<\/p>\n
<\/p>\n
06 \u5230\u5e95\u8be5\u7528Adam\u8fd8\u662fSGD\uff1f<\/p>\n
\u6240\u4ee5\uff0c\u8c08\u5230\u73b0\u5728\uff0c\u5230\u5e95Adam\u597d\u8fd8\u662fSGD\u597d\uff1f\u8fd9\u53ef\u80fd\u662f\u5f88\u96be\u4e00\u53e5\u8bdd\u8bf4\u6e05\u695a\u7684\u4e8b\u60c5\u3002\u53bb\u770b\u5b66\u672f\u4f1a\u8bae\u4e2d\u7684\u5404\u79cdpaper\uff0c\u7528SGD\u7684\u5f88\u591a\uff0cAdam\u7684\u4e5f\u4e0d\u5c11\uff0c\u8fd8\u6709\u5f88\u591a\u504f\u7231AdaGrad\u6216\u8005AdaDelta\u3002\u53ef\u80fd\u7814\u7a76\u5458\u628a\u6bcf\u4e2a\u7b97\u6cd5\u90fd\u8bd5\u4e86\u4e00\u904d\uff0c\u54ea\u4e2a\u51fa\u6765\u7684\u6548\u679c\u597d\u5c31\u7528\u54ea\u4e2a\u4e86\u3002\u6bd5\u7adfpaper\u7684\u91cd\u70b9\u662f\u7a81\u51fa\u81ea\u5df1\u67d0\u65b9\u9762\u7684\u8d21\u732e\uff0c\u5176\u4ed6\u65b9\u9762\u5f53\u7136\u662f\u65e0\u6240\u4e0d\u7528\u5176\u6781\uff0c\u600e\u4e48\u80fd\u8f93\u5728\u7ec6\u8282\u4e0a\u5462\uff1f<\/p>\n
\u800c\u4ece\u8fd9\u51e0\u7bc7\u6012\u603cAdam\u7684paper\u6765\u770b\uff0c\u591a\u6570\u90fd\u6784\u9020\u4e86\u4e00\u4e9b\u6bd4\u8f83\u6781\u7aef\u7684\u4f8b\u5b50\u6765\u6f14\u793a\u4e86Adam\u5931\u6548\u7684\u53ef\u80fd\u6027\u3002\u8fd9\u4e9b\u4f8b\u5b50\u4e00\u822c\u8fc7\u4e8e\u6781\u7aef\uff0c\u5b9e\u9645\u60c5\u51b5\u4e2d\u53ef\u80fd\u672a\u5fc5\u4f1a\u8fd9\u6837\uff0c\u4f46\u8fd9\u63d0\u9192\u4e86\u6211\u4eec\uff0c\u7406\u89e3\u6570\u636e\u5bf9\u4e8e\u8bbe\u8ba1\u7b97\u6cd5\u7684\u5fc5\u8981\u6027\u3002\u4f18\u5316\u7b97\u6cd5\u7684\u6f14\u53d8\u5386\u53f2\uff0c\u90fd\u662f\u57fa\u4e8e\u5bf9\u6570\u636e\u7684\u67d0\u79cd\u5047\u8bbe\u800c\u8fdb\u884c\u7684\u4f18\u5316\uff0c\u90a3\u4e48\u67d0\u79cd\u7b97\u6cd5\u662f\u5426\u6709\u6548\uff0c\u5c31\u8981\u770b\u4f60\u7684\u6570\u636e\u662f\u5426\u7b26\u5408\u8be5\u7b97\u6cd5\u7684\u80c3\u53e3\u4e86\u3002<\/p>\n
\u7b97\u6cd5\u56fa\u7136\u7f8e\u597d\uff0c\u6570\u636e\u624d\u662f\u6839\u672c\u3002<\/p>\n
\u53e6\u4e00\u65b9\u9762\uff0cAdam\u4e4b\u6d41\u867d\u7136\u8bf4\u5df2\u7ecf\u7b80\u5316\u4e86\u8c03\u53c2\uff0c\u4f46\u662f\u5e76\u6ca1\u6709\u4e00\u52b3\u6c38\u9038\u5730\u89e3\u51b3\u95ee\u9898\uff0c\u9ed8\u8ba4\u7684\u53c2\u6570\u867d\u7136\u597d\uff0c\u4f46\u4e5f\u4e0d\u662f\u653e\u4e4b\u56db\u6d77\u800c\u7686\u51c6\u3002\u56e0\u6b64\uff0c\u5728\u5145\u5206\u7406\u89e3\u6570\u636e\u7684\u57fa\u7840\u4e0a\uff0c\u4f9d\u7136\u9700\u8981\u6839\u636e\u6570\u636e\u7279\u6027\u3001\u7b97\u6cd5\u7279\u6027\u8fdb\u884c\u5145\u5206\u7684\u8c03\u53c2\u5b9e\u9a8c\u3002<\/p>\n
\u5c11\u5e74\uff0c\u597d\u597d\u70bc\u4e39\u5427\u3002<\/p>\n
\uff08\u4e09\uff09\u4f18\u5316\u7b97\u6cd5\u7684\u9009\u62e9\u4e0e\u4f7f\u7528\u7b56\u7565<\/p>\n
\u201c \u5728\u524d\u9762\u4e24\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u7528\u4e00\u4e2a\u6846\u67b6\u68b3\u7406\u4e86\u5404\u5927\u4f18\u5316\u7b97\u6cd5\uff0c\u5e76\u4e14\u6307\u51fa\u4e86\u4ee5 Adam \u4e3a\u4ee3\u8868\u7684\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u4f18\u5316\u7b97\u6cd5\u53ef\u80fd\u5b58\u5728\u7684\u95ee\u9898\u3002\u90a3\u4e48\uff0c\u5728\u5b9e\u8df5\u4e2d\u6211\u4eec\u5e94\u8be5\u5982\u4f55\u9009\u62e9\u5462\uff1f\u672c\u6587\u4ecb\u7ecd Adam + SGD \u7684\u7ec4\u5408\u7b56\u7565\uff0c\u4ee5\u53ca\u4e00\u4e9b\u6bd4\u8f83\u6709\u7528\u7684 tricks.\u201d<\/p>\n
07 \u4e0d\u540c\u7b97\u6cd5\u7684\u6838\u5fc3\u5dee\u5f02\uff1a\u4e0b\u964d\u65b9\u5411<\/p>\n
\u4ece\u7b2c\u4e00\u7bc7\u7684\u6846\u67b6\u4e2d\u6211\u4eec\u770b\u5230\uff0c\u4e0d\u540c\u4f18\u5316\u7b97\u6cd5\u6700\u6838\u5fc3\u7684\u533a\u522b\uff0c\u5c31\u662f\u7b2c\u4e09\u6b65\u6240\u6267\u884c\u7684\u4e0b\u964d\u65b9\u5411\uff1a<\/p>\n
<\/p>\n
\u8fd9\u4e2a\u5f0f\u5b50\u4e2d\uff0c\u524d\u534a\u90e8\u5206\u662f\u5b9e\u9645\u7684\u5b66\u4e60\u7387\uff08\u4e5f\u5373\u4e0b\u964d\u6b65\u957f\uff09\uff0c\u540e\u534a\u90e8\u5206\u662f\u5b9e\u9645\u7684\u4e0b\u964d\u65b9\u5411\u3002SGD\u7b97\u6cd5\u7684\u4e0b\u964d\u65b9\u5411\u5c31\u662f\u8be5\u4f4d\u7f6e\u7684\u68af\u5ea6\u65b9\u5411\u7684\u53cd\u65b9\u5411\uff0c\u5e26\u4e00\u9636\u52a8\u91cf\u7684SGD\u7684\u4e0b\u964d\u65b9\u5411\u5219\u662f\u8be5\u4f4d\u7f6e\u7684\u4e00\u9636\u52a8\u91cf\u65b9\u5411\u3002\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7c7b\u4f18\u5316\u7b97\u6cd5\u4e3a\u6bcf\u4e2a\u53c2\u6570\u8bbe\u5b9a\u4e86\u4e0d\u540c\u7684\u5b66\u4e60\u7387\uff0c\u5728\u4e0d\u540c\u7ef4\u5ea6\u4e0a\u8bbe\u5b9a\u4e0d\u540c\u6b65\u957f\uff0c\u56e0\u6b64\u5176\u4e0b\u964d\u65b9\u5411\u662f\u7f29\u653e\u8fc7\uff08scaled\uff09\u7684\u4e00\u9636\u52a8\u91cf\u65b9\u5411\u3002<\/p>\n
\u7531\u4e8e\u4e0b\u964d\u65b9\u5411\u7684\u4e0d\u540c\uff0c\u53ef\u80fd\u5bfc\u81f4\u4e0d\u540c\u7b97\u6cd5\u5230\u8fbe\u5b8c\u5168\u4e0d\u540c\u7684\u5c40\u90e8\u6700\u4f18\u70b9\u3002\u300aAn empirical analysis of the optimization of deep network loss surfaces\u300b \u8fd9\u7bc7\u8bba\u6587\u4e2d\u505a\u4e86\u4e00\u4e2a\u6709\u8da3\u7684\u5b9e\u9a8c\uff0c\u4ed6\u4eec\u628a\u76ee\u6807\u51fd\u6570\u503c\u548c\u76f8\u5e94\u7684\u53c2\u6570\u5f62\u6210\u7684\u8d85\u5e73\u9762\u6620\u5c04\u5230\u4e00\u4e2a\u4e09\u7ef4\u7a7a\u95f4\uff0c\u8fd9\u6837\u6211\u4eec\u53ef\u4ee5\u76f4\u89c2\u5730\u770b\u5230\u5404\u4e2a\u7b97\u6cd5\u662f\u5982\u4f55\u5bfb\u627e\u8d85\u5e73\u9762\u4e0a\u7684\u6700\u4f4e\u70b9\u7684\u3002<\/p>\n
<\/p>\n
\u4e0a\u56fe\u662f\u8bba\u6587\u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u6a2a\u7eb5\u5750\u6807\u8868\u793a\u964d\u7ef4\u540e\u7684\u7279\u5f81\u7a7a\u95f4\uff0c\u533a\u57df\u989c\u8272\u5219\u8868\u793a\u76ee\u6807\u51fd\u6570\u503c\u7684\u53d8\u5316\uff0c\u7ea2\u8272\u662f\u9ad8\u539f\uff0c\u84dd\u8272\u662f\u6d3c\u5730\u3002\u4ed6\u4eec\u505a\u7684\u662f\u914d\u5bf9\u513f\u5b9e\u9a8c\uff0c\u8ba9\u4e24\u4e2a\u7b97\u6cd5\u4ece\u540c\u4e00\u4e2a\u521d\u59cb\u5316\u4f4d\u7f6e\u5f00\u59cb\u51fa\u53d1\uff0c\u7136\u540e\u5bf9\u6bd4\u4f18\u5316\u7684\u7ed3\u679c\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u51e0\u4e4e\u4efb\u4f55\u4e24\u4e2a\u7b97\u6cd5\u90fd\u8d70\u5230\u4e86\u4e0d\u540c\u7684\u6d3c\u5730\uff0c\u4ed6\u4eec\u4e2d\u95f4\u5f80\u5f80\u9694\u4e86\u4e00\u4e2a\u5f88\u9ad8\u7684\u9ad8\u539f\u3002\u8fd9\u5c31\u8bf4\u660e\uff0c\u4e0d\u540c\u7b97\u6cd5\u5728\u9ad8\u539f\u7684\u65f6\u5019\uff0c\u9009\u62e9\u4e86\u4e0d\u540c\u7684\u4e0b\u964d\u65b9\u5411\u3002<\/p>\n
08 Adam+SGD \u7ec4\u5408\u7b56\u7565<\/p>\n
\u6b63\u662f\u5728\u6bcf\u4e00\u4e2a\u5341\u5b57\u8def\u53e3\u7684\u9009\u62e9\uff0c\u51b3\u5b9a\u4e86\u4f60\u7684\u5f52\u5bbf\u3002\u5982\u679c\u4e0a\u5929\u80fd\u591f\u7ed9\u6211\u4e00\u4e2a\u518d\u6765\u4e00\u6b21\u7684\u673a\u4f1a\uff0c\u6211\u4f1a\u5bf9\u90a3\u4e2a\u5973\u5b69\u5b50\u8bf4\uff1aSGD\uff01<\/p>\n
\u4e0d\u540c\u4f18\u5316\u7b97\u6cd5\u7684\u4f18\u52a3\u4f9d\u7136\u662f\u672a\u6709\u5b9a\u8bba\u7684\u4e89\u8bae\u8bdd\u9898\u3002\u636e\u6211\u5728paper\u548c\u5404\u7c7b\u793e\u533a\u770b\u5230\u7684\u53cd\u9988\uff0c\u4e3b\u6d41\u7684\u89c2\u70b9\u8ba4\u4e3a\uff1aAdam\u7b49\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7b97\u6cd5\u5bf9\u4e8e\u7a00\u758f\u6570\u636e\u5177\u6709\u4f18\u52bf\uff0c\u4e14\u6536\u655b\u901f\u5ea6\u5f88\u5feb\uff1b\u4f46\u7cbe\u8c03\u53c2\u6570\u7684SGD\uff08+Momentum\uff09\u5f80\u5f80\u80fd\u591f\u53d6\u5f97\u66f4\u597d\u7684\u6700\u7ec8\u7ed3\u679c\u3002<\/p>\n
\u90a3\u4e48\u6211\u4eec\u5c31\u4f1a\u60f3\u5230\uff0c\u53ef\u4e0d\u53ef\u4ee5\u628a\u8fd9\u4e24\u8005\u7ed3\u5408\u8d77\u6765\uff0c\u5148\u7528Adam\u5feb\u901f\u4e0b\u964d\uff0c\u518d\u7528SGD\u8c03\u4f18\uff0c\u4e00\u4e3e\u4e24\u5f97\uff1f\u601d\u8def\u7b80\u5355\uff0c\u4f46\u91cc\u9762\u6709\u4e24\u4e2a\u6280\u672f\u95ee\u9898\uff1a<\/p>\n
\n
- \n
\u4ec0\u4e48\u65f6\u5019\u5207\u6362\u4f18\u5316\u7b97\u6cd5\uff1f\u2014\u2014\u5982\u679c\u5207\u6362\u592a\u665a\uff0cAdam\u53ef\u80fd\u5df2\u7ecf\u8dd1\u5230\u81ea\u5df1\u7684\u76c6\u5730\u91cc\u53bb\u4e86\uff0cSGD\u518d\u600e\u4e48\u597d\u4e5f\u8dd1\u4e0d\u51fa\u6765\u4e86\u3002<\/p>\n<\/li>\n
- \n
\u5207\u6362\u7b97\u6cd5\u4ee5\u540e\u7528\u4ec0\u4e48\u6837\u7684\u5b66\u4e60\u7387\uff1f\u2014\u2014Adam\u7528\u7684\u662f\u81ea\u9002\u5e94\u5b66\u4e60\u7387\uff0c\u4f9d\u8d56\u7684\u662f\u4e8c\u9636\u52a8\u91cf\u7684\u7d2f\u79ef\uff0cSGD\u63a5\u7740\u8bad\u7ec3\u7684\u8bdd\uff0c\u7528\u4ec0\u4e48\u6837\u7684\u5b66\u4e60\u7387\uff1f<\/p>\n<\/li>\n<\/ol>\n
\u4e0a\u4e00\u7bc7\u4e2d\u63d0\u5230\u7684\u8bba\u6587 Improving Generalization Performance by Switching from Adam to SGD \u63d0\u51fa\u4e86\u89e3\u51b3\u8fd9\u4e24\u4e2a\u95ee\u9898\u7684\u601d\u8def\u3002<\/p>\n
\u9996\u5148\u6765\u770b\u7b2c\u4e8c\u4e2a\u95ee\u9898\uff0c\u5207\u6362\u4e4b\u540e\u7684\u5b66\u4e60\u7387\u3002<\/p>\n
Adam\u7684\u4e0b\u964d\u65b9\u5411\u662f<\/p>\n
<\/p>\n
\u800cSGD\u7684\u4e0b\u964d\u65b9\u5411\u662f<\/p>\n
<\/p>\n
SGD\u4e0b\u964d\u65b9\u5411\u5fc5\u5b9a\u53ef\u4ee5\u5206\u89e3\u4e3aAdam\u4e0b\u964d\u65b9\u5411\u53ca\u5176\u6b63\u4ea4\u65b9\u5411\u4e0a\u7684\u4e24\u4e2a\u65b9\u5411\u4e4b\u548c\uff0c\u90a3\u4e48\u5176\u5728Adam\u4e0b\u964d\u65b9\u5411\u4e0a\u7684\u6295\u5f71\u5c31\u610f\u5473\u7740SGD\u5728Adam\u7b97\u6cd5\u51b3\u5b9a\u7684\u4e0b\u964d\u65b9\u5411\u4e0a\u524d\u8fdb\u7684\u8ddd\u79bb\uff0c\u800c\u5728Adam\u4e0b\u964d\u65b9\u5411\u7684\u6b63\u4ea4\u65b9\u5411\u4e0a\u7684\u6295\u5f71\u662f SGD \u5728\u81ea\u5df1\u9009\u62e9\u7684\u4fee\u6b63\u65b9\u5411\u4e0a\u524d\u8fdb\u7684\u8ddd\u79bb\u3002<\/p>\n
<\/p>\n
(\u56fe\u7247\u6765\u81ea\u539f\u6587\uff0c\u8fd9\u91ccp\u4e3aAdam\u4e0b\u964d\u65b9\u5411\uff0cg\u4e3a\u68af\u5ea6\u65b9\u5411\uff0cr\u4e3aSGD\u7684\u5b66\u4e60\u7387\u3002)<\/p>\n
\u5982\u679cSGD\u8981\u8d70\u5b8cAdam\u672a\u8d70\u5b8c\u7684\u8def\uff0c\u90a3\u5c31\u9996\u5148\u8981\u63a5\u8fc7Adam\u7684\u5927\u65d7\u2014\u2014\u6cbf\u7740 \u65b9\u5411\u8d70\u4e00\u6b65\uff0c\u800c\u540e\u5728\u6cbf\u7740\u5176\u6b63\u4ea4\u65b9\u5411\u8d70\u76f8\u5e94\u7684\u4e00\u6b65\u3002<\/p>\n
\u8fd9\u6837\u6211\u4eec\u5c31\u77e5\u9053\u8be5\u5982\u4f55\u786e\u5b9aSGD\u7684\u6b65\u957f\uff08\u5b66\u4e60\u7387\uff09\u4e86\u2014\u2014SGD\u5728Adam\u4e0b\u964d\u65b9\u5411\u4e0a\u7684\u6b63\u4ea4\u6295\u5f71\uff0c\u5e94\u8be5\u6b63\u597d\u7b49\u4e8eAdam\u7684\u4e0b\u964d\u65b9\u5411\uff08\u542b\u6b65\u957f\uff09\u3002\u4e5f\u5373\uff1a<\/p>\n
<\/p>\n
\u89e3\u8fd9\u4e2a\u65b9\u7a0b\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5f97\u5230\u63a5\u7eed\u8fdb\u884cSGD\u7684\u5b66\u4e60\u7387\uff1a<\/p>\n
<\/p>\n
\u4e3a\u4e86\u51cf\u5c11\u566a\u58f0\u5f71\u54cd\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u79fb\u52a8\u5e73\u5747\u503c\u6765\u4fee\u6b63\u5bf9\u5b66\u4e60\u7387\u7684\u4f30\u8ba1\uff1a<\/p>\n
<\/p>\n
\u8fd9\u91cc\u76f4\u63a5\u590d\u7528\u4e86Adam\u7684 beta \u53c2\u6570\u3002<\/p>\n
\u7136\u540e\u6765\u770b\u7b2c\u4e00\u4e2a\u95ee\u9898\uff0c\u4f55\u65f6\u8fdb\u884c\u7b97\u6cd5\u7684\u5207\u6362\u3002<\/p>\n
\u4f5c\u8005\u63d0\u51fa\u7684\u65b9\u6cd5\u5f88\u7b80\u5355\uff0c\u90a3\u5c31\u662f\u5f53 SGD\u7684\u76f8\u5e94\u5b66\u4e60\u7387\u7684\u79fb\u52a8\u5e73\u5747\u503c\u57fa\u672c\u4e0d\u53d8\u7684\u65f6\u5019\uff0c\u5373\uff1a<\/p>\n
<\/p>\n
\u6bcf\u6b21\u8fed\u4ee3\u5b8c\u90fd\u8ba1\u7b97\u4e00\u4e0bSGD\u63a5\u73ed\u4eba\u7684\u76f8\u5e94\u5b66\u4e60\u7387\uff0c\u5982\u679c\u53d1\u73b0\u57fa\u672c\u7a33\u5b9a\u4e86\uff0c\u90a3\u5c31SGD\u4ee5\u4e3a\u5b66\u4e60\u7387\u63a5\u73ed\u524d\u8fdb\u3002\u4e0d\u8fc7\uff0c\u8fd9\u4e00\u65f6\u673a\u662f\u4e0d\u662f\u6700\u4f18\u7684\u5207\u6362\u65f6\u673a\uff0c\u4f5c\u8005\u5e76\u6ca1\u6709\u7ed9\u51fa\u6570\u5b66\u8bc1\u660e\uff0c\u53ea\u662f\u901a\u8fc7\u5b9e\u9a8c\u9a8c\u8bc1\u4e86\u6548\u679c\uff0c\u5207\u6362\u65f6\u673a\u8fd8\u662f\u4e00\u4e2a\u5f88\u503c\u5f97\u6df1\u5165\u7814\u7a76\u7684\u8bdd\u9898\u3002<\/p>\n
09 \u4f18\u5316\u7b97\u6cd5\u7684\u5e38\u7528tricks<\/p>\n
\u6700\u540e\uff0c\u5206\u4eab\u4e00\u4e9b\u5728\u4f18\u5316\u7b97\u6cd5\u7684\u9009\u62e9\u548c\u4f7f\u7528\u65b9\u9762\u7684\u4e00\u4e9btricks\u3002<\/p>\n
1.\u9996\u5148\uff0c\u5404\u5927\u7b97\u6cd5\u5b70\u4f18\u5b70\u52a3\u5e76\u65e0\u5b9a\u8bba\u3002<\/p>\n
\u5982\u679c\u662f\u521a\u5165\u95e8\uff0c\u4f18\u5148\u8003\u8651 SGD+Nesterov Momentum\u6216\u8005Adam.\uff08Standford 231n : The two recommended updates to use are either SGD+Nesterov Momentum or Adam\uff09<\/p>\n
2. \u9009\u62e9\u4f60\u719f\u6089\u7684\u7b97\u6cd5\u2014\u2014\u8fd9\u6837\u4f60\u53ef\u4ee5\u66f4\u52a0\u719f\u7ec3\u5730\u5229\u7528\u4f60\u7684\u7ecf\u9a8c\u8fdb\u884c\u8c03\u53c2\u3002<\/p>\n
3.\u5145\u5206\u4e86\u89e3\u4f60\u7684\u6570\u636e\u2014\u2014\u5982\u679c\u6a21\u578b\u662f\u975e\u5e38\u7a00\u758f\u7684\uff0c\u90a3\u4e48\u4f18\u5148\u8003\u8651\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7684\u7b97\u6cd5\u3002<\/p>\n
4. \u6839\u636e\u4f60\u7684\u9700\u6c42\u6765\u9009\u62e9\u2014\u2014\u5728\u6a21\u578b\u8bbe\u8ba1\u5b9e\u9a8c\u8fc7\u7a0b\u4e2d\uff0c\u8981\u5feb\u901f\u9a8c\u8bc1\u65b0\u6a21\u578b\u7684\u6548\u679c\uff0c\u53ef\u4ee5\u5148\u7528Adam\u8fdb\u884c\u5feb\u901f\u5b9e\u9a8c\u4f18\u5316\uff1b\u5728\u6a21\u578b\u4e0a\u7ebf\u6216\u8005\u7ed3\u679c\u53d1\u5e03\u524d\uff0c\u53ef\u4ee5\u7528\u7cbe\u8c03\u7684SGD\u8fdb\u884c\u6a21\u578b\u7684\u6781\u81f4\u4f18\u5316\u3002<\/p>\n
5. \u5148\u7528\u5c0f\u6570\u636e\u96c6\u8fdb\u884c\u5b9e\u9a8c\u3002\u6709\u8bba\u6587\u7814\u7a76\u6307\u51fa\uff0c\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u7b97\u6cd5\u7684\u6536\u655b\u901f\u5ea6\u548c\u6570\u636e\u96c6\u7684\u5927\u5c0f\u7684\u5173\u7cfb\u4e0d\u5927\u3002\uff08The mathematics of stochastic gradient descent are amazingly independent of the training set size. In particular, the asymptotic SGD convergence rates are independent from the sample size. [2]\uff09\u56e0\u6b64\u53ef\u4ee5\u5148\u7528\u4e00\u4e2a\u5177\u6709\u4ee3\u8868\u6027\u7684\u5c0f\u6570\u636e\u96c6\u8fdb\u884c\u5b9e\u9a8c\uff0c\u6d4b\u8bd5\u4e00\u4e0b\u6700\u597d\u7684\u4f18\u5316\u7b97\u6cd5\uff0c\u5e76\u901a\u8fc7\u53c2\u6570\u641c\u7d22\u6765\u5bfb\u627e\u6700\u4f18\u7684\u8bad\u7ec3\u53c2\u6570\u3002<\/p>\n
6. \u8003\u8651\u4e0d\u540c\u7b97\u6cd5\u7684\u7ec4\u5408\u3002\u5148\u7528Adam\u8fdb\u884c\u5feb\u901f\u4e0b\u964d\uff0c\u800c\u540e\u518d\u6362\u5230SGD\u8fdb\u884c\u5145\u5206\u7684\u8c03\u4f18\u3002\u5207\u6362\u7b56\u7565\u53ef\u4ee5\u53c2\u8003\u672c\u6587\u4ecb\u7ecd\u7684\u65b9\u6cd5\u3002<\/p>\n
7. \u6570\u636e\u96c6\u4e00\u5b9a\u8981\u5145\u5206\u7684\u6253\u6563\uff08shuffle\uff09\u3002\u8fd9\u6837\u5728\u4f7f\u7528\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7b97\u6cd5\u7684\u65f6\u5019\uff0c\u53ef\u4ee5\u907f\u514d\u67d0\u4e9b\u7279\u5f81\u96c6\u4e2d\u51fa\u73b0\uff0c\u800c\u5bfc\u81f4\u7684\u6709\u65f6\u5b66\u4e60\u8fc7\u5ea6\u3001\u6709\u65f6\u5b66\u4e60\u4e0d\u8db3\uff0c\u4f7f\u5f97\u4e0b\u964d\u65b9\u5411\u51fa\u73b0\u504f\u5dee\u7684\u95ee\u9898\u3002<\/p>\n
8. \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6301\u7eed\u76d1\u63a7\u8bad\u7ec3\u6570\u636e\u548c\u9a8c\u8bc1\u6570\u636e\u4e0a\u7684\u76ee\u6807\u51fd\u6570\u503c\u4ee5\u53ca\u7cbe\u5ea6\u6216\u8005AUC\u7b49\u6307\u6807\u7684\u53d8\u5316\u60c5\u51b5\u3002\u5bf9\u8bad\u7ec3\u6570\u636e\u7684\u76d1\u63a7\u662f\u8981\u4fdd\u8bc1\u6a21\u578b\u8fdb\u884c\u4e86\u5145\u5206\u7684\u8bad\u7ec3\u2014\u2014\u4e0b\u964d\u65b9\u5411\u6b63\u786e\uff0c\u4e14\u5b66\u4e60\u7387\u8db3\u591f\u9ad8\uff1b\u5bf9\u9a8c\u8bc1\u6570\u636e\u7684\u76d1\u63a7\u662f\u4e3a\u4e86\u907f\u514d\u51fa\u73b0\u8fc7\u62df\u5408\u3002<\/p>\n
9. \u5236\u5b9a\u4e00\u4e2a\u5408\u9002\u7684\u5b66\u4e60\u7387\u8870\u51cf\u7b56\u7565\u3002\u53ef\u4ee5\u4f7f\u7528\u5b9a\u671f\u8870\u51cf\u7b56\u7565\uff0c\u6bd4\u5982\u6bcf\u8fc7\u591a\u5c11\u4e2aepoch\u5c31\u8870\u51cf\u4e00\u6b21\uff1b\u6216\u8005\u5229\u7528\u7cbe\u5ea6\u6216\u8005AUC\u7b49\u6027\u80fd\u6307\u6807\u6765\u76d1\u63a7\uff0c\u5f53\u6d4b\u8bd5\u96c6\u4e0a\u7684\u6307\u6807\u4e0d\u53d8\u6216\u8005\u4e0b\u8dcc\u65f6\uff0c\u5c31\u964d\u4f4e\u5b66\u4e60\u7387\u3002<\/p>\n
\u8fd9\u91cc\u53ea\u5217\u4e3e\u51fa\u4e00\u4e9b\u5728\u4f18\u5316\u7b97\u6cd5\u65b9\u9762\u7684trick\uff0c\u5982\u6709\u9057\u6f0f\uff0c\u6b22\u8fce\u5404\u4f4d\u5728\u8bc4\u8bba\u4e2d\u8865\u5145\u3002\u63d0\u524d\u81f4\u8c22\uff01<\/p>\n","protected":false},"excerpt":{"rendered":"\u4f18\u5316\u65b9\u6cd5\u603b\u7ed3\uff1aAdam\u90a3\u4e48\u68d2\uff0c\u4e3a\u4ec0\u4e48\u8fd8\u5bf9SGD\u5ff5\u5ff5\u4e0d\u5fd8? (SGD\uff0cAdagrad\uff0cAdadelta\uff0cAdam\uff0cAdamax\uff0cNadam)\u672c\u6587\u8f6c\u8f7d\u81ea\u300c\u673a\u5668\u5b66\u4e60\u70bc...","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"_links":{"self":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/5172"}],"collection":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/comments?post=5172"}],"version-history":[{"count":0,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/5172\/revisions"}],"wp:attachment":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/media?parent=5172"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/categories?post=5172"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/tags?post=5172"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}