{"id":9147,"date":"2024-05-11T09:01:01","date_gmt":"2024-05-11T01:01:01","guid":{"rendered":""},"modified":"2024-05-11T09:01:01","modified_gmt":"2024-05-11T01:01:01","slug":"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b SAG\u3001SVRG\uff08\u968f\u673a\u68af\u5ea6\u4e0b\u964d\uff09","status":"publish","type":"post","link":"https:\/\/mushiming.com\/9147.html","title":{"rendered":"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b SAG\u3001SVRG\uff08\u968f\u673a\u68af\u5ea6\u4e0b\u964d\uff09"},"content":{"rendered":"

\u539f\u6587\u51fa\u5904\uff1ahttps:\/\/zhuanlan.zhihu.com\/p\/22402784?utm_source=tuicool&utm_medium=referral<\/p>\n

\u8fd9\u7bc7\u6587\u7ae0\u56de\u987e\u4e86\u57fa\u4e8e\u68af\u5ea6\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u5728\u8fd9\u51e0\u5e74\u7684\u91cd\u8981\u53d1\u5c55 -- SAG\u3001SVRG\u3002<\/span><\/p>\n

\u5f88\u591a\u5e38\u89c1\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\u7684\u76ee\u6807\uff08\u6bd4\u5982\u6700\u5c0f\u4e8c\u4e58\u505a\u7ebf\u6027\u56de\u5f52\u3001\u903b\u8f91\u56de\u5f52\uff09\u90fd\u53ef\u4ee5\u6982\u62ec\u6210\u4ee5\u4e0b\u8fd9\u79cd\u4e00\u822c\u5f62\u5f0f\uff1a<\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b<\/p>\n<\/p>\n

\n <\/p>\n

\u5176\u4e2d \"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b \u4ee3\u8868\u6837\u672c\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u7684\u635f\u5931\u51fd\u6570\uff0c\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u662f\u6a21\u578b\u7684\u53c2\u6570\uff0c\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u4ee3\u8868\u6b63\u5219\u5316\u9879\uff08\u7528\u4e8e\u63a7\u5236\u6a21\u578b\u590d\u6742\u5ea6\u6216\u8005\u6a21\u578b\u7a00\u758f\u5ea6\u7b49\u7b49\uff09\uff0c\u6709\u4e9b\u65f6\u5019\u8fd9\u4e2a\u6b63\u5219\u5316\u9879\u662f\u4e0d\u5e73\u6ed1\u7684\uff0c\u4e5f\u5c31\u662f\u8bf4\u5b83\u53ef\u80fd\u4e0d\u53ef\u5bfc\u3002<\/p>\n

\u6682\u65f6\u5148\u4e0d\u8003\u8651\u8fd9\u4e2a\u6b63\u5219\u5316\u9879\uff0c\u53ea\u8003\u8651\u6837\u672c\u4e0a\u7684\u635f\u5931\uff0c\u5e76\u4e14\u5bf9\u7b26\u53f7\u505a\u4e00\u70b9\u7b80\u5316\uff08\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff09\uff0c\u8003\u8651\u4e0b\u9762\u8fd9\u4e2a\u4f18\u5316\u76ee\u6807\uff1a <\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b
\n <\/p>\n

\u8fd9\u4e2a\u5f62\u5f0f\u975e\u5e38\u7b80\u5355\uff0c\u53ea\u8981\u6bcf\u4e2a\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u90fd\u53ef\u5bfc\uff0c\u5c31\u53ef\u4ee5\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\uff08Gradient Descent\uff09\u8fed\u4ee3\u6c42\u89e3:<\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff0c\u5176\u4e2d\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b \u8868\u793a\u7b2c t+1 \u6b21\u66f4\u65b0\u540e\u7684\u53c2\u6570\u3002 <\/p>\n

\u68af\u5ea6\u4e0b\u964d\u5bf9\u4e8e\u6837\u672c\u6570\u76ee\u6bd4\u8f83\u591a\u7684\u65f6\u5019\u6709\u4e00\u4e2a\u5f88\u5927\u7684\u52a3\u52bf\uff0c\u90a3\u5c31\u662f\u6bcf\u6b21\u9700\u8981\u6c42\u89e3\u6240\u6709\u6837\u672c\u7684\u68af\u5ea6\uff0c\u6837\u672c\u6570\u591a\u7684\u65f6\u5019\uff0c\u5bfc\u81f4\u8ba1\u7b97\u91cf\u5927\u589e\uff0c\u6240\u4ee5\u5b9e\u9645\u751f\u4ea7\u73af\u5883\u4e2d\uff0c\u5f80\u5f80\u91c7\u7528\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u7b97\u6cd5\uff08Stochastic Gradient Descent\uff09\uff0c\u4e00\u822c\u7b80\u5199\u505aSGD\u3002<\/p>\n

SGD\u6bcf\u6b21\u8fed\u4ee3\u7684\u65f6\u5019\u5747\u5300\u968f\u673a\u5f97\u9009\u62e9\u4e00\u4e2a\u6837\u672c\u6216\u8005mini-batch\u505a\u66f4\u65b0\uff1a<\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b <\/p>\n

\u76f8\u5bf9\u4e8e\u68af\u5ea6\u4e0b\u964d\uff0cSGD\u7684\u597d\u5904\u975e\u5e38\u660e\u663e\uff0c\u5c31\u662f\u53ef\u4ee5\u51cf\u5c11\u6bcf\u6b21\u66f4\u65b0\u7684\u8ba1\u7b97\u4ee3\u4ef7\uff0c\u4f46\u662fSGD\u5e26\u6765\u7684\u95ee\u9898\u662f\u6536\u655b\u901f\u5ea6\u4e0d\u5982\u68af\u5ea6\u4e0b\u964d<\/span><\/strong>\uff08\u6536\u655b\u901f\u5ea6\u662f\u8861\u91cf\u4f18\u5316\u7b97\u6cd5\u8ba1\u7b97\u590d\u6742\u5ea6\u7684\u57fa\u672c\u5de5\u5177\uff0c\u5177\u4f53\u5b9a\u4e49\u53ef\u4ee5\u53c2\u8003https:\/\/en.wikipedia.org\/wiki\/Rate_of_convergence \u6216\u8005\u5176\u4ed6\u4f18\u5316\u76f8\u5173\u7684\u6559\u6750\uff09\uff0c\u4e5f\u5c31\u662f\u8bf4\u4e3a\u4e86\u8fbe\u5230\u540c\u6837\u7684\u7cbe\u5ea6\uff0cSGD\u9700\u8981\u7684\u603b\u8fed\u4ee3\u6b21\u6570\u8981\u5927\u4e8e\u68af\u5ea6\u4e0b\u964d\uff0c\u4f46\u662f\uff0c\u5355\u6b21\u8fed\u4ee3\u7684\u8ba1\u7b97\u91cf\u8981\u5c0f\u5f97\u591a\u3002<\/strong><\/span>\u4ece\u6536\u655b\u901f\u5ea6\u5206\u6790\u4e0a\u770b\uff0cSGD\u80fd\u591f\u5728\u76ee\u6807\u51fd\u6570\u5f3a\u51f8\u5e76\u4e14\u9012\u51cf\u6b65\u957f\u7684\u60c5\u51b5\u4e0b\u505a\u5230\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b \u7684\u6b21\u7ebf\u6027\u6536\u655b\uff08sublinear convergence\uff09\uff0c\u800c\u68af\u5ea6\u4e0b\u964d\u5219\u53ef\u4ee5\u5728\u76ee\u6807\u51fd\u6570\u5f3a\u51f8\u7684\u60c5\u51b5\u4e0b\u505a\u5230\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b (\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b) \u7684\u7ebf\u6027\u6536\u655b\uff08linear convergence\uff09\u3002\u603b\u7ed3\u8d77\u6765\u5c31\u662f\uff0c\u5982\u679c\u60f3\u5feb\u901f\u5f97\u5230\u4e00\u4e2a\u53ef\u4ee5\u52c9\u5f3a\u63a5\u53d7\u7684\u89e3\uff0cSGD\u6bd4\u68af\u5ea6\u4e0b\u964d\u66f4\u52a0\u5408\u9002\uff0c\u4f46\u662f\u5982\u679c\u60f3\u5f97\u5230\u4e00\u4e2a\u7cbe\u786e\u5ea6\u9ad8\u7684\u89e3\uff0c\u5e94\u5f53\u9009\u62e9\u68af\u5ea6\u4e0b\u964d\u3002<\/span><\/strong><\/p>\n

SGD\u540e\u6765\u540e\u6765\u4e5f\u884d\u751f\u51fa\u4e86\u975e\u5e38\u591a\u7684\u53d8\u79cd\uff0c\u5c24\u5176\u662f\u4e00\u7c7b\u5206\u6790regret\u7684online\u7b97\u6cd5\uff0c\u5305\u62ecAdagrad\u3001Dual Averaging\u3001FTRL\u7b49\u3002\u4f46\u662f\uff0c\u59cb\u7ec8\u5b66\u672f\u754c\u5bf9\u4e8eSGD\u8fd8\u6709\u4e00\u79cd\u671f\u5f85\uff0c\u5c31\u662f\uff1a\u662f\u5426\u53ef\u4ee5\u628aSGD\u505a\u5230\u548c\u68af\u5ea6\u4e0b\u964d\u4e00\u6837\u7684\u7ebf\u6027\u6536\u655b\u3002\u76f4\u52302012\u548c2013\u5e74\uff0cSAG[1]\u4e0eSVRG[2]\u7b97\u6cd5\u53d1\u8868\u5728NIPS\u4e0a\uff0c\u6210\u4e3a\u8fd1\u51e0\u5e74SGD\u7c7b\u7b97\u6cd5\u7684\u6700\u5927\u7a81\u7834<\/strong>\u3002<\/p>\n

SAG\u7b97\u6cd5<\/strong>\uff08\u7b97\u6cd5\u6846\u56fe\u6458\u81ea[4]\uff0c\u8fd9\u91cc\u7684\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u662f\u6307\u68af\u5ea6\u51fd\u6570\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff0c\u800c\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u662f\u6307\u4e0a\u6587\u4e2d\u7684\u4f18\u5316\u53c2\u6570\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff09<\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\n<\/div>\n

\n
\n \u4f5c\u8005\uff1ali Eta
\n
\u94fe\u63a5\uff1ahttps:\/\/zhuanlan.zhihu.com\/p\/22402784
\n
\u6765\u6e90\uff1a\u77e5\u4e4e
\n
\u8457\u4f5c\u6743\u5f52\u4f5c\u8005\u6240\u6709\u3002\u5546\u4e1a\u8f6c\u8f7d\u8bf7\u8054\u7cfb\u4f5c\u8005\u83b7\u5f97\u6388\u6743\uff0c\u975e\u5546\u4e1a\u8f6c\u8f7d\u8bf7\u6ce8\u660e\u51fa\u5904\u3002 <\/p>\n

SAG\u7b97\u6cd5\u5728\u5185\u5b58\u4e2d\u4e3a\u6bcf\u4e2a\u6837\u672c\u90fd\u7ef4\u62a4\u4e00\u4e2a\u65e7\u7684\u68af\u5ea6\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff0c\u968f\u673a\u9009\u62e9\u4e00\u4e2a\u6837\u672c\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u6765\u66f4\u65b0\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff0c\u5e76\u7528\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u6765\u66f4\u65b0\u53c2\u6570\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u3002\u5177\u4f53\u5f97\u8bf4\uff0c\u66f4\u65b0\u7684\u9879\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u6765\u81ea\u4e8e\u7528\u65b0\u7684\u68af\u5ea6\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u66ff\u6362\u6389\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u4e2d\u7684\u65e7\u68af\u5ea6\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff0c\u8fd9\u4e5f\u5c31\u662f\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u8868\u8fbe\u7684\u610f\u601d\u3002\u5982\u6b64\uff0c\u6bcf\u6b21\u66f4\u65b0\u7684\u65f6\u5019\u4ec5\u4ec5\u9700\u8981\u8ba1\u7b97\u4e00\u4e2a\u6837\u672c\u7684\u68af\u5ea6\uff0c\u800c\u4e0d\u662f\u6240\u6709\u6837\u672c\u7684\u68af\u5ea6\u3002\u8ba1\u7b97\u5f00\u9500\u4e0eSGD\u65e0\u5f02\uff0c\u4f46\u662f\u5185\u5b58\u5f00\u9500\u8981\u5927\u5f97\u591a\u3002[1]\u4e2d\u5df2\u7ecf\u8bc1\u660eSAG\u662f\u4e00\u79cd\u7ebf\u6027\u6536\u655b\u7b97\u6cd5\uff0c\u8fd9\u4e2a\u901f\u5ea6\u8fdc\u6bd4SGD\u5feb\u3002<\/p>\n

SAG\u5b9e\u9a8c\u7ed3\u679c<\/strong>\uff08\u7ed3\u679c\u6458\u81ea[1]\u7684arxiv\u957f\u6587\u7248\uff09<\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b <\/p>\n

\u5b9e\u9a8c\u76ee\u6807\u51fd\u6570\u662fl2-regularized logistic regression\uff0c\u5de6\u4e00\u662f\u8bad\u7ec3\u8bef\u5dee\uff0c\u5de6\u4e8c\u548c\u5de6\u4e09\u5206\u522b\u662f\u4e24\u79cd\u6d4b\u8bd5\u76ee\u6807\u51fd\u6570\u4e0e\u6d4b\u8bd5\u8bef\u5dee\u3002\u6ce8\u610f\u5de6\u4e00\u7684\u7eb5\u5750\u6807\u662f\u5bf9\u6570\u5750\u6807\uff0c\u4e00\u822c\u8861\u91cf\u4f18\u5316\u7b97\u6cd5\u7684\u901f\u5ea6\u90fd\u4f1a\u91c7\u7528\u5bf9\u6570\u5750\u6807\uff0c\u56e0\u4e3a\u5728\u5bf9\u6570\u5750\u6807\u4e2d\u53ef\u4ee5\u660e\u663e\u770b\u51fa\u4e00\u4e2a\u7b97\u6cd5\u662f\u7ebf\u6027\u6536\u655b\uff08\u8fd1\u4e4e\u76f4\u7ebf\u4e0b\u964d\uff09\u8fd8\u662f\u6b21\u7ebf\u6027\u6536\u655b\uff08\u5927\u4f53\u662f\u4e00\u6761\u5411\u4e0b\u51f8\u7684\u66f2\u7ebf\uff09\u3002\u53ef\u4ee5\u770b\u51faSAG\u662f\u4e00\u79cd\u7ebf\u6027\u6536\u655b\u7b97\u6cd5\uff0c\u4e14\u76f8\u5bf9\u4e8e\u5176\u4ed6\u53c2\u4e0e\u6bd4\u8f83\u7684\u7b97\u6cd5\u6709\u5f88\u5927\u7684\u4f18\u52bf\u3002\u5177\u4f53\u5b9e\u9a8c\u914d\u7f6e\u6570\u636e\u96c6\u7b49\u53ef\u4ee5\u53c2\u8003\u539f\u6587\u3002<\/p>\n

SVRG\u7b97\u6cd5<\/strong>\uff08\u7b97\u6cd5\u6458\u81ea[2]\uff0c\u8fd9\u91cc\u7684\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u5c31\u662f\u4e0a\u6587\u4e2d\u7684\u635f\u5931\u51fd\u6570\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff09<\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b <\/p>\n

SVRG\u7684\u7b97\u6cd5\u601d\u8def\u662f\uff0c\u6bcf\u8fc7\u4e00\u6bb5\u65f6\u95f4\u8ba1\u7b97\u4e00\u6b21\u6240\u6709\u6837\u672c\u7684\u68af\u5ea6\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\uff0c\u6bcf\u4e2a\u9636\u6bb5\u5185\u90e8\u7684\u5355\u6b21\u66f4\u65b0\u91c7\u7528\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u6765\u66f4\u65b0\u5f53\u524d\u53c2\u6570\uff0c\u6bcf\u6b21\u66f4\u65b0\u6700\u591a\u8ba1\u7b97\u4e24\u6b21\u68af\u5ea6\u3002\u76f8\u5bf9\u4e8eSAG\u6765\u8bf4\uff0c\u4e0d\u9700\u8981\u5728\u5185\u5b58\u4e2d\u4e3a\u6bcf\u4e2a\u6837\u672c\u90fd\u7ef4\u62a4\u4e00\u4e2a\u68af\u5ea6\uff0c\u4e5f\u5c31\u662f\u8bf4\u8282\u7701\u4e86\u5185\u5b58\u8d44\u6e90\u3002\u6b64\u5916\uff0cSVRG\u4e2d\u63d0\u51fa\u4e86\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u6982\u5ff5\u53eb\u505avariance reduction\uff08\u65b9\u5dee\u7f29\u51cf\uff09\uff0c\u8fd9\u4e2a\u6982\u5ff5\u9700\u8981\u8054\u7cfbSGD\u7684\u6536\u655b\u6027\u5206\u6790\u6765\u7406\u89e3\uff0c\u5728SGD\u7684\u6536\u655b\u6027\u5206\u6790\u4e2d\u9700\u8981\u5047\u8bbe\u6837\u672c\u68af\u5ea6\u7684\u7684\u65b9\u5dee\u662f\u6709\u5e38\u6570\u4e0a\u754c\u7684\uff0c\u7136\u800c\u6b63\u662f\u56e0\u4e3a\u8fd9\u4e2a\u5e38\u6570\u4e0a\u754c\u5bfc\u81f4\u4e86SGD\u65e0\u6cd5\u7ebf\u6027\u6536\u655b\uff0c\u56e0\u6b64SVRG\u7684\u6536\u655b\u6027\u5206\u6790\u4e2d\u5229\u7528\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b\u8fd9\u79cd\u7279\u6b8a\u7684\u66f4\u65b0\u9879\u6765\u8ba9\u65b9\u5dee\u6709\u4e00\u4e2a\u53ef\u4ee5\u4e0d\u65ad\u51cf\u5c11\u7684\u4e0a\u754c\uff0c\u56e0\u6b64\u4e5f\u5c31\u505a\u5230\u4e86\u7ebf\u6027\u6536\u655b\uff0c\u8fd9\u4e00\u70b9\u5c31\u662fSVRG\u7684\u6838\u5fc3\uff0cSAG\u7684\u7b56\u7565\u5176\u5b9e\u4e5f\u4e0e\u6b64\u7c7b\u4f3c\uff08\u867d\u7136\u8bc1\u660e\u8fc7\u7a0b\u4e0d\u540c\uff09\u3002<\/p>\n

<\/p>\n

SVRG\u5b9e\u9a8c\u7ed3\u679c<\/strong>\uff08\u7ed3\u679c\u6458\u81ea[2]\uff09<\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b <\/p>\n

\u4e0a\u56fe\u4e3aSVRG\u5728\u51f8\u7684logistic regression\u4e0a\u7684\u8868\u73b0\uff0c\u6ce8\u610f\u5de6\u4e00\u7eb5\u5750\u6807\u662f\u8bad\u7ec3\u8bef\u5dee\uff0c\u5de6\u4e8c\u5de6\u4e09\u7eb5\u5750\u6807\u662f\u5bf9\u6570\u5750\u6807\uff0c\u5b9e\u9a8c\u4e2d\u53ef\u4ee5\u770b\u51faSVRG\u663e\u7136\u662f\u7ebf\u6027\u6536\u655b\u7b97\u6cd5\uff0c\u76f8\u5bf9\u4e8eSGD\u6709\u975e\u5e38\u5927\u7684\u4f18\u52bf\uff0c\u548cSDCA\u5177\u5907\u540c\u9636\u7684\u901f\u5ea6\u3002<\/p>\n

\"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b <\/p>\n

\u4e0a\u56fe\u4e3aSVRG\u5728\u975e\u51f8\u7684\u795e\u7ecf\u7f51\u7edc\uff08Neural Network\u6216\u79f0\u4f5cDeep Learning\uff09\u4e0a\u7684\u8868\u73b0\uff08\u539f\u6587\u4e2d\u662f\u5728\u5355\u9690\u5c42\u795e\u7ecf\u7f51\u7edc\u4e0a\u505a\u7684\u5b9e\u9a8c\uff09\u3002\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u8bf4\u660e\uff0cSVRG\u5728NN\u4e0a\u4e5f\u53ef\u4ee5\u53d1\u6325\u5f88\u597d\u7684\u4f5c\u7528\u3002<\/p>\n

\u540e\u6765\u8fd9\u7c7b\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u9646\u7eed\u51fa\u73b0\u4e86\u5f88\u591a\u53d8\u79cd\uff0c\u6bd4\u5982SAGA\u7b97\u6cd5[3]\u3002<\/p>\n

<\/p>\n

\u53c2\u8003\u6587\u732e\uff1a<\/strong><\/p>\n

    \n
  1. Roux, Nicolas L., Mark Schmidt, and Francis R. Bach. \"A stochastic gradient method with an exponential convergence rate for finite training sets.\"Advances in Neural Information Processing Systems<\/em>. 2012. <\/li>\n
  2. Johnson, Rie, and Tong Zhang. \"Accelerating stochastic gradient descent using predictive variance reduction.\"Advances in Neural Information Processing Systems<\/em>. 2013. <\/li>\n
  3. Defazio, Aaron, Francis Bach, and Simon Lacoste-Julien. \"Saga: A fast incremental gradient method with support for non-strongly convex composite objectives.\"Advances in Neural Information Processing Systems<\/em>. 2014.<\/li>\n
  4. Schmidt, Mark, Nicolas Le Roux, and Francis Bach. \"Minimizing finite sums with the stochastic average gradient.\"arXiv preprint arXiv:1309.2388<\/em> (2013).<\/li>\n<\/ol>\n
    \n
    \n <\/div>\n<\/p><\/div>\n
    \n \u8865\u5145\uff1a<\/strong><\/span>\n <\/div>\n
    \n

    \u68af\u5ea6\u4e0b\u964d\u6cd5\u5927\u5bb6\u65cf\uff08BGD\uff0cSGD\uff0cMBGD\uff09<\/h2>\n

    \u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\uff08Batch Gradient Descent\uff09<\/h3>\n

    \u3000\u3000\u3000\u3000\u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\uff0c\u662f\u68af\u5ea6\u4e0b\u964d\u6cd5\u6700\u5e38\u7528\u7684\u5f62\u5f0f\uff0c\u5177\u4f53\u505a\u6cd5\u4e5f\u5c31\u662f\u5728\u66f4\u65b0\u53c2\u6570\u65f6\u4f7f\u7528\u6240\u6709\u7684\u6837\u672c\u6765\u8fdb\u884c\u66f4\u65b0<\/p>\n

    \u968f\u673a\u68af\u5ea6\u4e0b\u964d\u6cd5\uff08Stochastic Gradient Descent\uff09<\/h3>\n

    \u3000\u3000\u3000\u3000\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u6cd5\uff0c\u5176\u5b9e\u548c\u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u539f\u7406\u7c7b\u4f3c\uff0c\u533a\u522b\u5728\u4e0e\u6c42\u68af\u5ea6\u65f6\u6ca1\u6709\u7528\u6240\u6709\u7684m\u4e2a\u6837\u672c\u7684\u6570\u636e\uff0c\u800c\u662f\u4ec5\u4ec5\u9009\u53d6\u4e00\u4e2a\u6837\u672cj\u6765\u6c42\u68af\u5ea6\u3002<\/p>\n

    \u3000\u3000\u3000\u3000\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u6cd5\uff0c\u548c\u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u662f\u4e24\u4e2a\u6781\u7aef\uff0c\u4e00\u4e2a\u91c7\u7528\u6240\u6709\u6570\u636e\u6765\u68af\u5ea6\u4e0b\u964d\uff0c\u4e00\u4e2a\u7528\u4e00\u4e2a\u6837\u672c\u6765\u68af\u5ea6\u4e0b\u964d\u3002\u81ea\u7136\u5404\u81ea\u7684\u4f18\u7f3a\u70b9\u90fd\u975e\u5e38\u7a81\u51fa\u3002\u5bf9\u4e8e\u8bad\u7ec3\u901f\u5ea6\u6765\u8bf4\uff0c\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u6cd5\u7531\u4e8e\u6bcf\u6b21\u4ec5\u4ec5\u91c7\u7528\u4e00\u4e2a\u6837\u672c\u6765\u8fed\u4ee3\uff0c\u8bad\u7ec3\u901f\u5ea6\u5f88\u5feb\uff0c\u800c\u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u5728\u6837\u672c\u91cf\u5f88\u5927\u7684\u65f6\u5019\uff0c\u8bad\u7ec3\u901f\u5ea6\u4e0d\u80fd\u8ba9\u4eba\u6ee1\u610f\u3002\u5bf9\u4e8e\u51c6\u786e\u5ea6\u6765\u8bf4\uff0c\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u6cd5\u7528\u4e8e\u4ec5\u4ec5\u7528\u4e00\u4e2a\u6837\u672c\u51b3\u5b9a\u68af\u5ea6\u65b9\u5411\uff0c\u5bfc\u81f4\u89e3\u5f88\u6709\u53ef\u80fd\u4e0d\u662f\u6700\u4f18\u3002\u5bf9\u4e8e\u6536\u655b\u901f\u5ea6\u6765\u8bf4\uff0c\u7531\u4e8e\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u6cd5\u4e00\u6b21\u8fed\u4ee3\u4e00\u4e2a\u6837\u672c\uff0c\u5bfc\u81f4\u8fed\u4ee3\u65b9\u5411\u53d8\u5316\u5f88\u5927\uff0c\u4e0d\u80fd\u5f88\u5feb\u7684\u6536\u655b\u5230\u5c40\u90e8\u6700\u4f18\u89e3\u3002<\/span><\/strong><\/p>\n

    \u5c0f\u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\uff08Mini-batch Gradient Descent\uff09<\/h3>\n

    \u3000\u3000\u5c0f\u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u662f\u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u548c\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u6cd5\u7684\u6298\u8877\uff0c\u4e5f\u5c31\u662f\u5bf9\u4e8em\u4e2a\u6837\u672c\uff0c\u6211\u4eec\u91c7\u7528x\u4e2a\u6837\u5b50\u6765\u8fed\u4ee3\uff0c1<x<m\u3002\u4e00\u822c\u53ef\u4ee5\u53d6x=10\uff0c\u5f53\u7136\u6839\u636e\u6837\u672c\u7684\u6570\u636e\uff0c\u53ef\u4ee5\u8c03\u6574\u8fd9\u4e2ax\u7684\u503c\u3002<\/p>\n

    \u68af\u5ea6\u4e0b\u964d\u6cd5\u548c\u5176\u4ed6\u65e0\u7ea6\u675f\u4f18\u5316\u7b97\u6cd5\u7684\u6bd4\u8f83<\/h2>\n

    \u3000\u3000\u3000\u3000\u5728\u673a\u5668\u5b66\u4e60\u4e2d\u7684\u65e0\u7ea6\u675f\u4f18\u5316\u7b97\u6cd5\uff0c\u9664\u4e86\u68af\u5ea6\u4e0b\u964d\u4ee5\u5916\uff0c\u8fd8\u6709\u524d\u9762\u63d0\u5230\u7684\u6700\u5c0f\u4e8c\u4e58\u6cd5\uff0c\u6b64\u5916\u8fd8\u6709\u725b\u987f\u6cd5\u548c\u62df\u725b\u987f\u6cd5\u3002<\/p>\n

    \u3000\u3000\u3000\u3000\u68af\u5ea6\u4e0b\u964d\u6cd5\u548c\u6700\u5c0f\u4e8c\u4e58\u6cd5\u76f8\u6bd4\uff0c\u68af\u5ea6\u4e0b\u964d\u6cd5\u9700\u8981\u9009\u62e9\u6b65\u957f\uff0c\u800c\u6700\u5c0f\u4e8c\u4e58\u6cd5\u4e0d\u9700\u8981\u3002\u68af\u5ea6\u4e0b\u964d\u6cd5\u662f\u8fed\u4ee3\u6c42\u89e3\uff0c\u6700\u5c0f\u4e8c\u4e58\u6cd5\u662f\u8ba1\u7b97\u89e3\u6790\u89e3\u3002\u5982\u679c\u6837\u672c\u91cf\u4e0d\u7b97\u5f88\u5927\uff0c\u4e14\u5b58\u5728\u89e3\u6790\u89e3\uff0c\u6700\u5c0f\u4e8c\u4e58\u6cd5\u6bd4\u8d77\u68af\u5ea6\u4e0b\u964d\u6cd5\u8981\u6709\u4f18\u52bf\uff0c\u8ba1\u7b97\u901f\u5ea6\u5f88\u5feb\u3002\u4f46\u662f\u5982\u679c\u6837\u672c\u91cf\u5f88\u5927\uff0c\u7528\u6700\u5c0f\u4e8c\u4e58\u6cd5\u7531\u4e8e\u9700\u8981\u6c42\u4e00\u4e2a\u8d85\u7ea7\u5927\u7684\u9006\u77e9\u9635\uff0c\u8fd9\u65f6\u5c31\u5f88\u96be\u6216\u8005\u5f88\u6162\u624d\u80fd\u6c42\u89e3\u89e3\u6790\u89e3\u4e86\uff0c\u4f7f\u7528\u8fed\u4ee3\u7684\u68af\u5ea6\u4e0b\u964d\u6cd5\u6bd4\u8f83\u6709\u4f18\u52bf\u3002<\/p>\n

    \u3000\u3000\u3000\u3000\u68af\u5ea6\u4e0b\u964d\u6cd5\u548c\u725b\u987f\u6cd5\/\u62df\u725b\u987f\u6cd5\u76f8\u6bd4\uff0c\u4e24\u8005\u90fd\u662f\u8fed\u4ee3\u6c42\u89e3\uff0c\u4e0d\u8fc7\u68af\u5ea6\u4e0b\u964d\u6cd5\u662f\u68af\u5ea6\u6c42\u89e3\uff0c\u800c\u725b\u987f\u6cd5\/\u62df\u725b\u987f\u6cd5\u662f\u7528\u4e8c\u9636\u7684\u6d77\u68ee\u77e9\u9635\u7684\u9006\u77e9\u9635\u6216\u4f2a\u9006\u77e9\u9635\u6c42\u89e3\u3002\u76f8\u5bf9\u800c\u8a00\uff0c\u4f7f\u7528\u725b\u987f\u6cd5\/\u62df\u725b\u987f\u6cd5\u6536\u655b\u66f4\u5feb\u3002\u4f46\u662f\u6bcf\u6b21\u8fed\u4ee3\u7684\u65f6\u95f4\u6bd4\u68af\u5ea6\u4e0b\u964d\u6cd5\u957f\u3002<\/p>\n

    \n <\/div>\n

    \n<\/div>\n

    <\/p>\n","protected":false},"excerpt":{"rendered":"\u7ebf\u6027\u6536\u655b\u7684\u968f\u673a\u4f18\u5316\u7b97\u6cd5\u4e4b SAG\u3001SVRG\uff08\u968f\u673a\u68af\u5ea6\u4e0b\u964d\uff09\u539f\u6587\u51fa\u5904\uff1ahttps:\/\/zhuanlan.zhihu.com\/p\/22402784?utm_source=tuicool&...","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\/9147"}],"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=9147"}],"version-history":[{"count":0,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/9147\/revisions"}],"wp:attachment":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/media?parent=9147"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/categories?post=9147"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/tags?post=9147"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}