{"id":9034,"date":"2024-05-16T21:01:01","date_gmt":"2024-05-16T13:01:01","guid":{"rendered":""},"modified":"2024-05-16T21:01:01","modified_gmt":"2024-05-16T13:01:01","slug":"\u6307\u6570\u52a0\u6743\u5e73\u5747\u3001\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u3001RMSprop\u3001Adam\u4f18\u5316\u7b97\u6cd5","status":"publish","type":"post","link":"https:\/\/mushiming.com\/9034.html","title":{"rendered":"\u6307\u6570\u52a0\u6743\u5e73\u5747\u3001\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u3001RMSprop\u3001Adam\u4f18\u5316\u7b97\u6cd5"},"content":{"rendered":"
\u76ee\u5f55<\/p>\n
\u8fd9\u91cc\u6709\u4e00\u5e74\u7684\u6e29\u5ea6\u6570\u636e\u3002<\/p>\n
\u5982\u679c\u60f3\u8ba1\u7b97\u6e29\u5ea6\u7684\u8d8b\u52bf<\/strong>\uff0c\u4e5f\u5c31\u662f\u5c40\u90e8\u5e73\u5747\u503c<\/strong>(local average)\uff0c\u6216\u8005\u8bf4\u79fb\u52a8\u5e73\u5747\u503c<\/strong>(moving average)\uff0c\u600e\u4e48\u505a\uff1f<\/p>\n <\/p>\n \uff1a\u5f53\u5929\u7684\u6e29\u5ea6\uff0c\uff1a\u4ece\u5f53\u5929\u5f00\u59cb\u8ba1\u7b97\u524d\u5929\u7684\u5e73\u5747\u6e29\u5ea6\uff0c\uff1a\u4ece\u6628\u5929\u5f00\u59cb\u8ba1\u7b97\u524d\u5929\u7684\u5e73\u5747\u6e29\u5ea6\u3002<\/p>\n \u6bd4\u5982\uff0c\u03b2=0.9\uff0c\u8ba1\u7b97\u7684\u5c31\u662f\u524d10\u5929\u7684\u5e73\u5747\u6e29\u5ea6\uff0c\u5982\u4e0b\u56fe\u7ea2\u7ebf\u6240\u793a\u3002\u03b2=0.98\uff0c\u8ba1\u7b97\u7684\u5c31\u662f\u524d50\u5929\u7684\u5e73\u5747\u6e29\u5ea6\uff0c\u5982\u4e0b\u56fe\u7eff\u7ebf\u6240\u793a\u3002<\/p>\n \u4e3a\u4ec0\u4e48\u662f<\/strong><\/span>\u5929\uff1f<\/strong><\/span><\/p>\n \uff0c\uff0c...<\/p>\n ...<\/p>\n \u8fd9\u4e9b\u9879\u7684\u7cfb\u6570\u5448\u6307\u6570\u7ea7\u51cf\u5c11<\/strong><\/p>\n 0.37\uff0c<\/p>\n \u5f53\u03b2->1\uff0c\u9009\u5b9a\u4e3a\u5206\u754c\u7ebf\uff0c\u4ece\u5f00\u59cb\uff0c\u4ee5\u540e\u7684\u9879\u90fd\u53ef\u5ffd\u7565\u4e0d\u8ba1\uff0c(1-\u03b2)=\u3002<\/p>\n \u4f8b\uff1a<\/strong>\u03b2=0.9\uff0c10\u4e3a\u5206\u754c\u7ebf\uff0c(1-0.9)0.90....+0....\u3002<\/p>\n \u03b2=0.98\uff0c50\u4e3a\u5206\u754c\u7ebf\uff0c(1-0.98)0.0....+0....\u3002<\/p>\n \u504f\u5dee\u4fee\u6b63(Bias correction)<\/strong><\/span><\/p>\n \u03b2=0.98\uff0c\u5982\u679c\u6ca1\u6709\u8fdb\u884c\u504f\u5dee\u4fee\u6b63\uff0c\u5f97\u5230\u7684\u5176\u5b9e\u662f\u4e0b\u56fe\u7684\u7d2b\u7ebf\u3002<\/p>\n \uff0c0.02\uff0c=0.98x0.02+0.02<\/p>\n =5\uff0c\uff0c\u7b97\u5c40\u90e8\u5e73\u5747\u503c\u7684\u65f6\u5019\uff0c\u4e0d\u592a\u5408\u7406\uff0c\u504f\u5c0f<\/strong>\u3002<\/p>\n \u6240\u4ee5\uff0c\u5728\u4f30\u6d4b\u521d\u671f\u8981\u8fdb\u884c\u4e00\u4e2a\u4fee\u6b63\uff0c\u516c\u5f0f\uff1a<\/p>\n \u4f8b<\/strong>\uff1at=2\uff0c\u8981\u4fee\u6b63\u4e3a\uff0c\u03b2=0.98<\/p>\n \u5f53t\u8d8a\u6765\u8d8a\u5927\u7684\u65f6\u5019\uff0c\u5c31\u8fd1\u4f3c\u7b49\u4e8e\uff0c\u4e0a\u56fe\u7684\u7d2b\u7ebf\u5c31\u548c\u7eff\u7ebf\u91cd\u5408\u4e86\uff0c\u4fee\u6b63\u504f\u5dee\u7684\u4f5c\u7528\u4e5f\u5c31\u4e0d\u5927\u4e86\u3002<\/p>\n \u6240\u4ee5\uff0c\u5982\u679c\u4f60\u5173\u5fc3\u521d\u59cb\u65f6\u671f\u7684\u504f\u5dee\uff0c\u504f\u5dee\u4fee\u6b63\u80fd\u8ba9\u4f60\u5728\u65e9\u671f\u83b7\u5f97\u66f4\u597d\u7684\u4f30\u6d4b\u3002\u4e0d\u5728\u4e4e\u65e9\u671f\u7684\u504f\u5dee\uff0c\u4e0d\u505a\u4fee\u6b63\u4e5f\u95ee\u9898\u4e0d\u5927\u3002<\/strong><\/p>\n \u6307\u6570\u52a0\u6743\u5e73\u5747\u4f18\u70b9\u662f\u53ea\u4f1a\u5360\u5f88\u5c11\u7684\u5185\u5b58\uff0c\u8ba1\u7b97\u7684\u65f6\u5019\u53ea\u9700\u8981\u4e00\u884c\u4ee3\u7801\uff0c\u9700\u8981\u77e5\u9053\u7684\u4fe1\u606f\u5f88\u5c11\u3002\u8ba1\u7b97\u7cbe\u5ea6\u80af\u5b9a\u4e0d\u5982\u76f4\u63a5\u8ba1\u7b97\u524dn\u5929\u7684\u5747\u503c\uff0c\u4f46\u662f\u540e\u8005\u8981\u4fdd\u5b58\u524dn\u5929\u7684\u6570\u636e\uff0c\u66f4\u5360\u5185\u5b58\u3002<\/strong><\/span><\/p>\n \u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5(\u7ea2\u7ebf)\u8981\u5feb\u4e8e<\/strong>\u6807\u51c6\u7684\u68af\u5ea6\u4e0b\u964d\u6cd5(\u84dd\u7ebf)\u3002<\/p>\n \u6bcf\u6b21\u8fed\u4ee3\uff0c\u5728\u73b0\u6709\u7684mini-batch\u4e0a\u8ba1\u7b97dW\uff0cdb\uff0c\u66f4\u65b0\u53c2\u6570\u3002\u7b49\u5f0f\u53f3\u8fb9\u7684\u5176\u5b9e\u662f\u4e0a\u4e00\u72b6\u6001\u7684\u503c\u3002<\/p>\n \u4e24\u4e2a\u8d85\u53c2\u6570(hyperparameter)\uff1a\u03b1(\u5b66\u4e60\u7387)\u3001\u03b2(\u63a7\u5236\u7740\u6307\u6570\u52a0\u6743\u5e73\u5747)\uff0c\u03b2\u5e38\u89c1\u503c\u7b49\u4e8e0.9<\/p>\n RMSprop\u7b97\u6cd5\u540c\u6837\u80fd\u52a0\u901f\u68af\u5ea6\u4e0b\u964d<\/strong>\u3002<\/p>\n \u6bcf\u6b21\u8fed\u4ee3\uff0c\u5728\u73b0\u6709\u7684mini-batch\u4e0a\u8ba1\u7b97dW\uff0cdb\uff0c\u66f4\u65b0\u53c2\u6570\u3002<\/p>\n \uff0c<\/p>\n w:=\uff0cb:=<\/p>\n RMS\uff1aRoot Mean Square\uff0c\u5747\u65b9\u6839<\/strong><\/p>\n \u8be5\u7b97\u6cd5\u5b9e\u9645\u4e0a\u662f\u5c06momentum<\/strong>\u548cRMSprop<\/strong>\u8fd9\u4e24\u79cd\u7b97\u6cd5\u7ed3\u5408\u8d77\u6765\uff0c\u4e5f\u662f\u4e3a\u4e86\u52a0\u901f\u68af\u5ea6\u4e0b\u964d<\/strong>\uff0c\u5e94\u7528\u5e7f\u6cdb\u3002<\/strong><\/span><\/p>\n \u6bcf\u6b21\u8fed\u4ee3\uff0c\u5728\u73b0\u6709\u7684mini-batch\u4e0a\u8ba1\u7b97dW\uff0cdb\uff0c\u66f4\u65b0\u53c2\u6570\u3002<\/p>\n \uff0c<\/p>\n \uff0c<\/p>\n \u504f\u5dee\u4fee\u6b63\uff1a\uff0c\uff0c\uff0c<\/p>\n w:=\uff0cb:=<\/p>\n \u8d85\u53c2\u6570<\/strong>\uff1a\u5b66\u4e60\u7387\u03b1\uff1a\u9700\u8981\u8c03\u6574\u3001\u5e38\u7528\u503c\u4e3a0.9\u3001\u5e38\u7528\u503c\u4e3a0.999\u3001\u03b5\u5e38\u7528\u503c\u4e3a<\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"\u6307\u6570\u52a0\u6743\u5e73\u5747\u3001\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u3001RMSprop\u3001Adam\u4f18\u5316\u7b97\u6cd5\u6307\u6570\u52a0\u6743\u5e73\u5747\u3001\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u3001RMSprop\u3001Adam\u4f18\u5316\u7b97\u6cd5_biascorrection...","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\/9034"}],"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=9034"}],"version-history":[{"count":0,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/9034\/revisions"}],"wp:attachment":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/media?parent=9034"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/categories?post=9034"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/tags?post=9034"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}2.\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5(Gradient descent with momentum)<\/strong><\/span><\/h2>\n
3.RMSprop<\/strong><\/span><\/h2>\n
4.Adam\u4f18\u5316\u7b97\u6cd5(optimization algorithm)<\/strong><\/span><\/h2>\n