{"id":118171,"date":"2019-05-22T12:15:12","date_gmt":"2019-05-22T04:15:12","guid":{"rendered":"http:\/\/www.fagao.me\/?p=118171"},"modified":"2019-05-22T12:15:12","modified_gmt":"2019-05-22T04:15:12","slug":"%e5%ad%a6%e7%95%8c-%e6%9c%ba%e9%81%87%e4%b8%8e%e6%8c%91%e6%88%98-%e7%94%a8%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e8%87%aa%e5%8a%a8%e6%90%9c%e7%b4%a2%e4%bc%98%e5%8c%96%e7%ae%97%e6%b3%95","status":"publish","type":"post","link":"https:\/\/www.fagao.me\/p\/118171.htm","title":{"rendered":"\u5b66\u754c | \u673a\u9047\u4e0e\u6311\u6218 \u7528\u5f3a\u5316\u5b66\u4e60\u81ea\u52a8\u641c\u7d22\u4f18\u5316\u7b97\u6cd5"},"content":{"rendered":"<p><img decoding=\"async\" src=\"https:\/\/www.fagao.me\/images\/uploads\/F027.jpg\" alt=\"\"><\/p>\n<div class=\"focusup\"><\/div>\n<p><span><span>\u9009\u81ea BAIR<\/span><\/span><\/p>\n<p><span><strong>\u53c2\u4e0e\uff1a\u8def\u96ea\u3001\u674e\u6cfd\u5357\u3001\u848b\u601d\u6e90<\/strong><\/span><\/p>\n<blockquote>\n<p><span>\u81ea\u4ece\u53bb\u5e74 UC Berkeley \u8bba\u6587\u300aLearning to Optimize\u300b\u53d1\u8868\u4ee5\u6765\uff0c\u6709\u5173\u4f18\u5316\u5668\u5b66\u4e60\uff08optimizer learning\uff09\u7684\u7814\u7a76\u5c31\u5f15\u8d77\u4e86\u4eba\u4eec\u7684\u91cd\u89c6\u3002\u5728\u672c\u6587\u4e2d\uff0cBAIR \u5728\u8bfb\u535a\u58eb Ke Li \u5c06\u5411\u6211\u4eec\u4ecb\u7ecd\u8fd9\u4e00\u5de5\u4f5c\u7684\u8fdb\u5c55\uff0c\u5e76\u5206\u4eab\u8fd9\u4e00\u9886\u57df\u7684\u673a\u9047\u548c\u6311\u6218\u3002<\/span><\/p>\n<\/blockquote>\n<p><span>\u8fd1\u5e74\u6765\uff0c\u673a\u5668\u5b66\u4e60\u5df2\u7ecf\u53d6\u5f97\u4e86\u5de8\u5927\u7684\u6210\u529f\uff0c\u5b83\u5df2\u88ab\u5e94\u7528\u5728\u4e86\u5f88\u591a\u4e0d\u540c\u9886\u57df\u4e2d\u3002\u8fd9\u79cd\u6210\u529f\u53ef\u4ee5\u5f52\u529f\u4e8e\u7531\u6570\u636e\u9a71\u52a8\u7684\u673a\u5668\u5b66\u4e60\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u80fd\u5728\u4f7f\u7528\u4e13\u4e1a\u77e5\u8bc6\u624b\u52a8\u8bbe\u8ba1\u7684\u7cfb\u7edf\u4e0a\u81ea\u52a8\u6316\u6398\u6570\u636e\u4e2d\u7684\u6a21\u5f0f\u3002<\/span><\/p>\n<p><span>\u7136\u800c\uff0c\u76ee\u524d\u7684\u8303\u5f0f\u4ecd\u5b58\u5728\u4e00\u4e2a\u6096\u8bba\uff1a\u7531\u7b97\u6cd5\u9a71\u52a8\u7684\u673a\u5668\u5b66\u4e60\u4ecd\u7136\u662f\u624b\u52a8\u8bbe\u8ba1\u7684\u3002\u8fd9\u5c31\u5f15\u51fa\u4e86\u4e00\u4e2a\u95ee\u9898\uff1a\u6211\u4eec\u80fd\u5426\u8ba9\u7b97\u6cd5\u81ea\u6211\u5b66\u4e60\uff1f\u8fd9\u5c06\u5f00\u542f\u60ca\u4eba\u7684\u53ef\u80fd\u6027\uff0c\u5373\u81ea\u52a8\u5bfb\u627e\u53ef\u4ee5\u6bd4\u624b\u52a8\u8bbe\u8ba1\u66f4\u52a0\u5b8c\u7f8e\u7684\u65b0\u7b97\u6cd5\uff0c\u5e76\u4e14\u65b0\u7b97\u6cd5\u80fd\u663e\u8457\u5730\u63d0\u5347\u5b66\u4e60\u80fd\u529b\u3002<\/span><\/p>\n<\/p>\n<p><span>\u8981\u60f3\u8fd9\u4e48\u505a\uff0c\u6211\u4eec\u9700\u8981\u514b\u670d\u4e00\u4e2a\u5de8\u5927\u7684\u969c\u788d\uff1a\u5982\u4f55\u53c2\u6570\u5316\u7b97\u6cd5\u7a7a\u95f4\uff0c\u8ba9\u5b83\u4eec\u540c\u65f6\u5177\u6709\u8868\u73b0\u529b\uff08expressive\uff09\u548c\u641c\u7d22\u6709\u6548\u6027\uff1f\u5f88\u591a\u79cd\u5df2\u6709\u7b97\u6cd5\u90fd\u5728\u4e24\u8005\u4e4b\u95f4\u8fdb\u884c\u7740\u53d6\u820d\u3002\u4f8b\u5982\uff0c\u5982\u679c\u7b97\u6cd5\u7a7a\u95f4\u4ee3\u8868\u4e86\u4e00\u5c0f\u7ec4\u5df2\u77e5\u7b97\u6cd5\uff0c\u5219\u5176\u4e2d\u53ef\u80fd\u4e0d\u5305\u542b\u6700\u4f73\u7b97\u6cd5\uff0c\u4f46\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u7b80\u5355\u679a\u4e3e\u96c6\u5408\u4e2d\u7684\u7b97\u6cd5\u6765\u6709\u6548\u5730\u8fdb\u884c\u641c\u7d22\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5982\u679c\u7b97\u6cd5\u7a7a\u95f4\u4ee3\u8868\u4e86\u6240\u6709\u53ef\u80fd\u7684\u7a0b\u5e8f\uff0c\u5176\u4e2d\u5fc5\u7136\u5305\u542b\u4e86\u6700\u4f73\u7b97\u6cd5\uff0c\u4f46\u8fd9\u79cd\u60c5\u51b5\u7f3a\u4e4f\u6709\u6548\u7684\u641c\u7d22\uff0c\u56e0\u4e3a\u679a\u4e3e\u65f6\u95f4\u5c06\u5448\u6307\u6570\u7ea7\u589e\u957f\u3002<\/span><\/p>\n<p><span>\u8fde\u7eed\u4f18\u5316\uff08continuous optimization\uff09\u7b97\u6cd5\u662f\u673a\u5668\u5b66\u4e60\u6700\u4e3a\u5e38\u89c1\u7684\u7b97\u6cd5\u4e4b\u4e00\uff0c\u5176\u4e2d\u5305\u542b\u4e00\u7cfb\u5217\u5df2\u77e5\u6d41\u884c\u7684\u7b97\u6cd5\uff0c\u5305\u62ec\u68af\u5ea6\u4e0b\u964d\u3001\u52a8\u91cf\u6cd5\u3001AdaGrad \u548c ADAM \u65b9\u6cd5\u3002\u6211\u4eec\u8003\u8651\u8fc7\u81ea\u52a8\u8bbe\u8ba1\u8fd9\u4e9b\u4f18\u5316\u7b97\u6cd5\u7684\u95ee\u9898\uff0c\u8fd9\u4e48\u505a\u6709\u4e24\u4e2a\u539f\u56e0\uff1a\u9996\u5148\uff0c\u5f88\u591a\u4f18\u5316\u7b97\u6cd5\u662f\u5728\u51f8\u5047\u8bbe\u4e0b\u8bbe\u8ba1\u7684\uff0c\u4f46\u88ab\u5e94\u7528\u5230\u975e\u51f8\u76ee\u6807\u51fd\u6570\u4e0a\uff1b\u901a\u8fc7\u5728\u5b9e\u9645\u4f7f\u7528\u73af\u5883\u4e0b\u5b66\u4e60\uff0c\u4f18\u5316\u7b97\u6cd5\u6709\u671b\u5b9e\u73b0\u66f4\u597d\u7684\u6027\u80fd\u3002\u7b2c\u4e8c\uff0c\u624b\u52a8\u8bbe\u8ba1\u65b0\u7684\u4f18\u5316\u7b97\u6cd5\u901a\u5e38\u8d39\u65f6\u8d39\u529b\uff0c\u53ef\u80fd\u9700\u8981\u6570\u6708\u6216\u6570\u5e74\u4e4b\u4e45\uff1b\u5b66\u4e60\u4f18\u5316\u7b97\u6cd5\u53ef\u4ee5\u51cf\u5c11\u624b\u52a8\u8bbe\u8ba1\u7684\u52b3\u52a8\u91cf\u3002<\/span><\/p>\n<p><strong>\u5b66\u4e60\u5982\u4f55\u4f18\u5316<\/strong><\/p>\n<p><span>\u5728\u53bb\u5e74\u7684\u8bba\u6587\u300aLearning to Optimize\u300b\u4e2d\uff0cUC Berkeley \u7684\u7814\u7a76\u8005\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u5b66\u4e60\u4f18\u5316\u7b97\u6cd5\u7684\u6846\u67b6\uff0c\u5b83\u88ab\u79f0\u4e4b\u4e3a\u300cLearning to Optimize\u300d\u3002\u5728\u8bba\u6587\u51fa\u73b0\u540e\uff0c\u6211\u4eec\u5f88\u5feb\u6ce8\u610f\u5230 DeepMind \u7b49\u673a\u6784\u7684\u7814\u7a76\u8005\u4e5f\u5728\u8bba\u6587\u300aLearning to learn by gradient descent by gradient descent\u300b\u4e2d\u63d0\u51fa\u4e86\u7c7b\u4f3c\u7684\u60f3\u6cd5\uff08\u8bfb\u8005\u53ef\u4ee5\u6311\u6218\u4e00\u4e0b\u8fd9\u7bc7\u8bba\u6587\u6807\u9898\u5982\u4f55\u7ffb\uff09\u3002<\/span><\/p>\n<p><span>\u8003\u8651\u5230\u73b0\u6709\u7684\u8fde\u7eed\u4f18\u5316\u7b97\u6cd5\u901a\u5e38\u7684\u5de5\u4f5c\u65b9\u5f0f\uff1a\u5b83\u4eec\u901a\u5e38\u4ee5\u8fed\u4ee3\u7684\u65b9\u5f0f\u8fd0\u884c\uff0c\u5e76\u4fdd\u6301\u4e00\u5b9a\u7a0b\u5ea6\u7684\u8fed\u4ee3\u6570\u91cf\uff0c\u5e76\u4e14\u6bcf\u4e00\u6b21\u8fed\u4ee3\u90fd\u9009\u53d6\u76ee\u6807\u51fd\u6570\u57df\u5185\u7684\u4e00\u4e2a\u70b9\u4ee5\u641c\u7d22\u5168\u5c40\u6700\u5c0f\u3002\u5728\u6bcf\u4e2a\u8fed\u4ee3\u4e2d\uff0c\u7b97\u6cd5\u90fd\u4f1a\u4f7f\u7528\u56fa\u5b9a\u7684\u66f4\u65b0\u516c\u5f0f\u6765\u8ba1\u7b97\u6b65\u957f\u5411\u91cf\uff0c\u7136\u540e\u7528\u5b83\u6765\u4fee\u6b63\u8fed\u4ee3\u65b9\u5411\u4e0e\u4e0b\u964d\u6b65\u957f\u3002\u66f4\u65b0\u516c\u5f0f\u901a\u5e38\u662f\u5728\u5f53\u524d\u548c\u8fc7\u53bb\u7684\u8fed\u4ee3\u4e2d\u6240\u8bc4\u4f30\u5386\u53f2\u68af\u5ea6\u7684\u51fd\u6570\uff0c\u8fd9\u4e9b\u68af\u5ea6\u90fd\u662f\u76ee\u6807\u51fd\u6570\u5bf9\u4e0d\u540c\u53c2\u6570\u6240\u6c42\u7684\u504f\u5bfc\u6570\u3002\u4f8b\u5982\uff0c\u5728\u68af\u5ea6\u4e0b\u964d\u4e2d\uff0c\u66f4\u65b0\u516c\u5f0f\u662f\u4e00\u4e9b\u52a0\u6743\u7684\u8d1f\u68af\u5ea6\uff08\u6240\u52a0\u6743\u4e3a\u5b66\u4e60\u7387\uff09\uff1b\u5728\u52a8\u91cf\u6cd5\u4e2d\uff0c\u66f4\u65b0\u516c\u5f0f\u662f\u4e00\u4e9b\u52a0\u6743\u7684\u6307\u6570\u79fb\u52a8\u5747\u503c\u3002<\/span><\/p>\n<\/p>\n<p><span>\u8fd9\u79cd\u66f4\u65b0\u516c\u5f0f\u662f\u533a\u522b\u4e0d\u540c\u7b97\u6cd5\u7684\u57fa\u7840\u3002\u6240\u4ee5\uff0c\u5982\u679c\u6211\u4eec\u53ef\u4ee5\u5b66\u4e60\u66f4\u65b0\u516c\u5f0f\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5b66\u4e60\u4f18\u5316\u7b97\u6cd5\u3002\u6211\u4eec\u5c06\u66f4\u65b0\u516c\u5f0f\u5efa\u6a21\u4e3a\u795e\u7ecf\u7f51\u7edc\u3002\u8fd9\u6837\uff0c\u901a\u8fc7\u5b66\u4e60\u795e\u7ecf\u7f51\u7edc\u7684\u6743\u91cd\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5b66\u4e60\u51fa\u4e00\u4e2a\u4f18\u5316\u7b97\u6cd5\u3002\u5c06\u66f4\u65b0\u516c\u5f0f\u53c2\u6570\u5316\u4e3a\u795e\u7ecf\u7f51\u7edc\u7684\u540c\u65f6\u6ee1\u8db3\u4e86\u524d\u9762\u63d0\u5230\u7684\u4e24\u4e2a\u7279\u8d28\uff1a\u9996\u5148\uff0c\u5b83\u5177\u6709\u8868\u8fbe\u6027\uff0c\u56e0\u4e3a\u795e\u7ecf\u7f51\u7edc\u662f\u901a\u7528\u51fd\u6570\u7684\u8fd1\u4f3c\uff0c\u5b83\u5728\u672c\u8d28\u4e0a\u53ef\u4ee5\u6a21\u62df\u5177\u6709\u8db3\u591f\u5bb9\u91cf\u7684\u4efb\u4f55\u66f4\u65b0\u516c\u5f0f\u3002<\/span><\/p>\n<p><span>\u4e3a\u4e86\u5b66\u4e60\u4f18\u5316\u7b97\u6cd5\uff0c\u6211\u4eec\u9700\u8981\u5148\u5b9a\u4e49\u4e00\u4e2a\u6027\u80fd\u6307\u6807\uff0c\u6211\u4eec\u5c06\u5176\u79f0\u4e4b\u4e3a\u300c\u5143\u635f\u5931\uff08meta-loss\uff09\u300d\uff0c\u5b83\u4f1a\u5956\u52b1\u8868\u73b0\u597d\u7684\u4f18\u5316\u5668\uff0c\u5e76\u60e9\u7f5a\u4e0d\u597d\u7684\u4f18\u5316\u5668\u3002\u56e0\u4e3a\u597d\u7684\u4f18\u5316\u5668\u53ef\u4ee5\u5feb\u901f\u6536\u655b\uff0c\u6b63\u5e38\u7684\u5143\u635f\u5931\u662f\u6240\u6709\u8fed\u4ee3\u4e2d\u7684\u76ee\u6807\u51fd\u6570\u503c\u7684\u603b\u548c\uff08\u5047\u8bbe\u671f\u671b\u662f\u6700\u5c0f\u5316\u76ee\u6807\u51fd\u6570\uff09\uff0c\u6216\u76f8\u5f53\u4e8e\u7f3a\u5931\u503c\u7d2f\u79ef\uff08cumulative regret\uff09\u3002\u76f4\u89c2\u5730\u8bf4\uff0c\u8fd9\u5bf9\u5e94\u4e8e\u66f2\u7ebf\u4e0b\u9762\u79ef\uff0c\u5f53\u4f18\u5316\u5668\u6536\u655b\u901f\u5ea6\u8f83\u6162\u7684\u65f6\u5019\uff0c\u5b83\u7684\u66f2\u7ebf\u4e0b\u9762\u79ef\u5c31\u4f1a\u589e\u5927\u3002<\/span><\/p>\n<p><strong>\u5b66\u4e60\u5982\u4f55\u5b66\u4e60<\/strong><\/p>\n<p><span>\u8003\u8651\u5230\u76ee\u6807\u51fd\u6570\u662f\u8bad\u7ec3\u5176\u4ed6\u6a21\u578b\u7684\u635f\u5931\u51fd\u6570\u8fd9\u4e00\u7279\u6b8a\u60c5\u51b5\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u4f18\u5316\u5668\u7684\u5b66\u4e60\u53ef\u4ee5\u63cf\u8ff0\u4e3a\u300c\u5b66\u4e60\u5982\u4f55\u5b66\u4e60\u300d\u3002\u4e3a\u6e05\u695a\u8d77\u89c1\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u4f18\u5316\u5668\u8bad\u7ec3\u8fc7\u7684\u6a21\u578b\u79f0\u4e3a\u300cbase-model\u300d\uff0c\u5e76\u52a0\u4e0a\u300cbase-\u300d\u548c\u300cmeta-\u300d\u524d\u7f00\u5206\u522b\u6392\u9664\u4e0e base-model \u548c\u4f18\u5316\u5668\u76f8\u5173\u7684\u6982\u5ff5\u6b67\u4e49\u3002<\/span><\/p>\n<p><span>\u300c\u5b66\u4e60\u5982\u4f55\u5b66\u4e60\u300d\u662f\u4ec0\u4e48\u610f\u601d\uff1f\u867d\u7136\u8fd9\u4e2a\u672f\u8bed\u5728\u8fd1\u5e74\u6765\u7684\u8bba\u6587\u4e2d\u4e0d\u65f6\u4f1a\u51fa\u73b0\uff0c\u4f46\u4e0d\u540c\u7684\u4f5c\u8005\u7ed9\u5b83\u8d4b\u4e88\u4e86\u4e0d\u540c\u7684\u610f\u4e49\uff0c\u6211\u4eec\u5e76\u6ca1\u6709\u5bf9\u8fd9\u6bb5\u8bdd\u7684\u7cbe\u786e\u542b\u4e49\u8fbe\u6210\u5171\u8bc6\u3002\u901a\u5e38\uff0c\u5b83\u4e5f\u53ef\u4ee5\u4e0e\u672f\u8bed\u300c\u5143\u5b66\u4e60\u300d\u4e92\u6362\u4f7f\u7528\u3002<\/span><\/p>\n<p><span>\u8fd9\u4e2a\u672f\u8bed\u53ef\u4ee5\u8ffd\u6eaf\u5230\u5143\u8ba4\u77e5\u7684\u6982\u5ff5\uff08\u4e9a\u91cc\u58eb\u591a\u5fb7\uff0c\u516c\u5143\u524d 350 \u5e74\uff09\uff0c\u5b83\u63cf\u8ff0\u4e86\u4eba\u7c7b\u4e0d\u4ec5\u7406\u6027\uff0c\u4e5f\u5bf9\u4e8e\u81ea\u5df1\u4ea7\u751f\u7406\u6027\u7684\u8fc7\u7a0b\u4fdd\u6301\u7406\u6027\u3002\u300c\u5b66\u4e60\u5982\u4f55\u5b66\u4e60\u300d\u4ece\u8fd9\u4e00\u7406\u5ff5\u4e2d\u5bfb\u6c42\u542f\u53d1\uff0c\u5e76\u8bd5\u56fe\u5c06\u5176\u8f6c\u5316\u4e3a\u5177\u4f53\u7684\u7b97\u6cd5\u3002\u901a\u4fd7\u5730\u8bf4\uff0c\u300c\u5b66\u4e60\u5982\u4f55\u5b66\u4e60\u300d\u4ec5\u4ec5\u610f\u5473\u7740\u5b66\u4e60\u4e00\u4e9b\u5173\u4e8e\u5b66\u4e60\u7684\u65b9\u6cd5\u3002\u5728\u5143\u5c42\u9762\u4e0a\uff08meta-level\uff09\u5b58\u5728\u4e0d\u540c\u65b9\u6cd5\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u5143\u77e5\u8bc6\u5c06\u5b83\u5206\u4e3a\u4e09\u7c7b\uff1a<\/span><\/p>\n<ul>\n<li>\n<p><span>\u5b66\u4e60\u8981\u5b66\u4e60\u7684\u76ee\u6807<\/span><\/p>\n<\/li>\n<li>\n<p><span>\u5b66\u4e60\u8981\u5b66\u4e60\u54ea\u4e2a\u6a21\u578b<\/span><\/p>\n<\/li>\n<li>\n<p><span>\u5b66\u4e60\u5982\u4f55\u5b66\u4e60<\/span><\/p>\n<\/li>\n<\/ul>\n<p><strong><span>\u5b66\u4e60\u8981\u5b66\u4e60\u7684\u76ee\u6807<\/span><\/strong><\/p>\n<p><span>\u8fd9\u4e9b\u65b9\u6cd5\u65e8\u5728\u5b66\u4e60 base-model \u53c2\u6570\u4e2d\u7684\u4e00\u4e9b\u7279\u5b9a\u503c\uff0c\u4f7f\u5176\u5bf9\u5e94\u76f8\u5173\u4efb\u52a1\u65f6\u6548\u7387\u66f4\u9ad8\u3002\u5143\u77e5\u8bc6\u6355\u6349\u6574\u4e2a\u65cf\u4e2d\u7684\u5171\u540c\u7279\u5f81\uff0c\u8fd9\u6837\u4e00\u6765\uff0cbase-learning \u5728\u9047\u5230\u65cf\u4e2d\u7684\u65b0\u4efb\u52a1\u65f6\u5c31\u80fd\u8bad\u7ec3\u548c\u63a8\u65ad\u5f97\u66f4\u5feb\u3002\u5176\u4e2d\u7684\u4f8b\u5b50\u5305\u542b\u8fc1\u79fb\u5b66\u4e60\u65b9\u6cd5\u3001\u591a\u4efb\u52a1\u5b66\u4e60\u548c few-shot \u5b66\u4e60\u3002\u65e9\u671f\u65b9\u6cd5\u901a\u8fc7\u5c06 base-model \u7684\u53c2\u6570\u5206\u4e3a\u4e24\u7ec4\uff08\u7528\u4e8e\u7279\u5b9a\u4efb\u52a1\u7684\u548c\u5728\u4efb\u52a1\u4e2d\u901a\u7528\u7684\u53c2\u6570\uff09\u7684\u65b9\u5f0f\u6765\u8fd0\u884c\u3002\u4f8b\u5982\uff0c\u5728\u57fa\u4e8e\u795e\u7ecf\u7f51\u7edc\u7684\u6a21\u578b\u91cc\u6709\u4e00\u79cd\u6d41\u884c\u7684\u65b9\u6cd5\u662f\u5728\u6240\u6709\u4efb\u52a1\u4e2d\u7684\u4f4e\u5c42\u5206\u4eab\u6743\u91cd\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u6355\u83b7\u8de8\u8d8a\u4efb\u52a1\u7684\u5171\u540c\u7279\u5f81\u3002\u53c2\u89c1\u300a\u4e0e\u6a21\u578b\u65e0\u5173\u7684\u5143\u5b66\u4e60\uff0cUC Berkeley \u63d0\u51fa\u4e00\u79cd\u53ef\u63a8\u5e7f\u5230\u5404\u7c7b\u4efb\u52a1\u7684\u5143\u5b66\u4e60\u65b9\u6cd5\u300b\u3002<\/span><\/p>\n<p><strong><span>\u5b66\u4e60\u8981\u5b66\u4e60\u54ea\u4e2a\u6a21\u578b<\/span><\/strong><\/p>\n<p><span>\u8fd9\u4e9b\u65b9\u6cd5\u65e8\u5728\u5b66\u4e60\u54ea\u4e9b base-model \u53ef\u4ee5\u9002\u7528\u4e8e\u7279\u5b9a\u4efb\u52a1\u3002\u5143\u77e5\u8bc6\u6355\u83b7\u4e0d\u540c base-model \u4e4b\u95f4\u7684\u76f8\u5173\u6027\u53ca\u5176\u5728\u4e0d\u540c\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u3002\u5b83\u9762\u4e34\u7684\u6311\u6218\u5728\u4e8e\u901a\u8fc7\u8868\u8fbe\u548c\u6709\u6548\u641c\u7d22\u7684\u65b9\u5f0f\u53c2\u6570\u5316 base-model \u7a7a\u95f4\uff0c\u4ee5\u53ca\u5728\u53c2\u6570\u5316\u4efb\u52a1\u7a7a\u95f4\u4e2d\u6cdb\u5316\u5230\u672a\u89c1\u8fc7\u7684\u65b0\u4efb\u52a1\u4e0a\u3002\u4e0d\u540c\u7684\u65b9\u6cd5\u5728\u8868\u8fbe\u6027\u548c\u53ef\u641c\u7d22\u6027\u95ee\u9898\u4e0a\u6709\u7740\u4e0d\u540c\u7684\u53d6\u820d\u3002\u5728\u300aRanking Learning Algorithms: Using IBL and Meta-Learning on Accuracy and Time Results\u300b\u4e2d\uff0cBrazdil \u7b49\u4eba\u4f7f\u7528\u9884\u5b9a\u4e49\u7684 base-model \u548c\u793a\u4f8b\u4efb\u52a1\u7684\u6570\u636e\u5e93\uff0c\u5e76\u8f93\u51fa\u5728\u6700\u8fd1\u793a\u4f8b\u4efb\u52a1\u4e0a\u6267\u884c\u6700\u4f73\u7684 base-model\u3002\u867d\u7136\u8fd9\u4e2a base-model \u7a7a\u95f4\u662f\u53ef\u641c\u7d22\u7684\uff0c\u4f46\u5b83\u53ef\u80fd\u4e0d\u5305\u542b\u5341\u5206\u4f18\u79c0\u4f46\u5374\u672a\u88ab\u53d1\u73b0\u7684 base-model\u3002J\u00fcrgen Schmidhuber \u7b49\u4eba\u5728\u8bba\u6587\u300aOptimal Ordered Problem Solver\u300b\u4e2d\u63d0\u51fa\u4e86\u8ba9\u6bcf\u4e2a base-model \u90fd\u4f5c\u4e3a\u901a\u7528\u7a0b\u5e8f\u7684\u65b9\u6cd5\u3002\u867d\u7136\u7a7a\u95f4\u975e\u5e38\u5177\u6709\u8868\u8fbe\u6027\uff0c\u4f46\u8fd9\u4e2a\u7a7a\u95f4\u7684\u641c\u7d22\u65f6\u95f4\u548c\u76ee\u6807\u7b97\u6cd5\u7684\u6570\u91cf\u6210\u6307\u6570\u578b\u589e\u957f\u3002Hochreiter \u7b49\u4eba\u7684\u300aLearning to Learn Using Gradient Descent\u300b\u63d0\u51fa\u4e86\u4e00\u79cd\u5c06 base-model \u8bad\u7ec3\u4e3a\u9ed1\u7bb1\u51fd\u6570\u7684\u7b97\u6cd5\uff0c\u8be5\u7b97\u6cd5\u53ef\u4ee5\u5c06\u8bad\u7ec3\u793a\u4f8b\u7684\u5e8f\u5217\u7ed8\u5236\u6210\u9884\u6d4b\u5e8f\u5217\u5e76\u5c06\u5176\u5efa\u6a21\u4e3a\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u3002\u8fd9\u6837\uff0c\u81ea\u8bad\u7ec3\u5c31\u88ab\u7b80\u5316\u6210\u4e3a\u4e86\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u7684\u8fc7\u7a0b\u3002\u56e0\u4e3a base-model \u7f16\u7801\u5230\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u8bb0\u5fc6\u72b6\u6001\u4e2d\uff0c\u6240\u4ee5\u5b83\u7684\u5bb9\u91cf\u53d7\u5230\u5185\u5b58\u7684\u9650\u5236\u3002\u4e0e\u4e4b\u76f8\u5173\u7684\u9886\u57df\u8fd8\u6709\u8d85\u53c2\u6570\u8c03\u4f18\uff0c\u5176\u76ee\u6807\u8f83\u5f31\uff0c\u5373\u641c\u7d22\u7531\u9884\u5b9a\u4e49\u7684\u4e00\u7ec4\u8d85\u53c2\u6570\u53c2\u6570\u5316\u7684 base-model\u3002\u5b83\u662f\u8de8\u8d85\u53c2\u6570\u7684\u8bbe\u7f6e\uff08\u901a\u8fc7\u5ef6\u5c55\u3001base-model\uff09\uff0c\u4f46\u975e\u8de8\u4efb\u52a1\u7684\u6cdb\u5316\uff0c\u56e0\u4e3a\u8d85\u53c2\u6570\u4f18\u5316\u662f\u5728\u4e0d\u540c\u8d85\u53c2\u6570\u4e0b\u5bf9\u540c\u4e00\u4e2a\u4efb\u52a1\u8fdb\u884c\u6d4b\u8bd5\u3002<\/span><\/p>\n<p><strong><span>\u5b66\u4e60\u5982\u4f55\u5b66\u4e60<\/span><\/strong><\/p>\n<p><span>\u524d\u4e24\u4e2a\u65b9\u5411\u96c6\u4e2d\u4e8e\u5b66\u4e60\u4e4b\u5916\uff0c\u7b2c\u4e09\u4e2a\u65b9\u5411\u5219\u6709\u5173\u5b66\u4e60\u7684\u8fc7\u7a0b\uff0c\u5373\u5143\u77e5\u8bc6\u6355\u83b7\u5b66\u4e60\u7b97\u6cd5\u884c\u4e3a\u5171\u540c\u7279\u5f81\u7684\u8fc7\u7a0b\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u4e09\u4e2a\u90e8\u5206\uff1abase-model\u3001\u7528\u4e8e\u8bad\u7ec3 base-model \u7684 base-algorithm \u4ee5\u53ca\u5b66\u4e60 base-algorithm \u7684 meta-algorithm\u3002\u5176\u4e2d\u5b66\u5230\u7684\u4e0d\u662f base-model \u672c\u8eab\uff0c\u800c\u662f\u5728\u67d0\u4e2a\u4efb\u52a1\u4e2d\u8bad\u7ec3 base-model \u7684 base-algorithm\u3002\u56e0\u4e3a base-model \u548c\u4efb\u52a1\u90fd\u662f\u7531\u7528\u6237\u63d0\u4f9b\u7684\uff0c\u5b66\u4e60\u5230\u7684 base-algorithm \u5fc5\u987b\u53ef\u4ee5\u5728\u591a\u79cd\u4e0d\u540c\u7684 base-model \u548c\u4efb\u52a1\u4e2d\u4f7f\u7528\u3002\u7531\u4e8e\u5927\u591a\u6570\u5b66\u4e60\u7b97\u6cd5\u4f18\u5316\u4e86\u4e00\u4e9b\u76ee\u6807\u51fd\u6570\uff0c\u56e0\u6b64\u5728\u8bb8\u591a\u60c5\u51b5\u4e0b\u5b66\u4e60 base-algorithm \u96be\u4ee5\u5b66\u5230\u4f18\u5316\u7b97\u6cd5\u3002\u8fd9\u4e2a\u5b66\u4e60\u4f18\u5316\u7b97\u6cd5\u7684\u95ee\u9898\u66fe\u88ab Ke Li \u4e0e Jitendra Malik \u7684\u300aLearning to Optimize\u300b\u548c Marcin Andrychowicz \u7b49\u4eba\u7684\u300aLearning to learn by gradient descent by gradient descent\u300b\u7b49\u8bba\u6587\u7814\u7a76\u8fc7\u3002\u4e0e\u4e4b\u6700\u63a5\u8fd1\u7684\u8bba\u6587\u5219\u662f Bengio \u7b49\u4eba\u5728 1991 \u5e74\u53d1\u8868\u7684\u300aLearning a synaptic learning rule\u300b\uff0c\u5176\u4e2d\u7814\u7a76\u4eba\u5458\u8bd5\u56fe\u5b66\u4e60\u4e86 Hebb \u5f0f\u7684\u89e6\u7a81\u5b66\u4e60\u89c4\u5219\uff08learning rule\uff09\u3002\u5b66\u4e60\u89c4\u5219\u53d6\u51b3\u4e8e\u7f16\u7801\u76f8\u90bb\u795e\u7ecf\u5143\u6d3b\u52a8\u7684\u5f53\u524d\u8fed\u4ee3\u7684\u7ef4\u5ea6\u7684\u5b50\u96c6\uff0c\u4f46\u4e0d\u4f9d\u8d56\u4e8e\u76ee\u6807\u51fd\u6570\uff0c\u56e0\u6b64\u4e0d\u5177\u6709\u5c06\u5176\u63a8\u5e7f\u5230\u4e0d\u540c\u7684\u76ee\u6807\u51fd\u6570\u7684\u80fd\u529b\u3002<\/span><\/p>\n<p><strong>\u6cdb\u5316<\/strong><\/p>\n<p><span>\u4efb\u4f55\u5f62\u5f0f\u7684\u5b66\u4e60\u90fd\u8981\u6c42\u5728\u6709\u9650\u7684\u6837\u672c\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u5e76\u6cdb\u5316\u81f3\u6837\u672c\u6240\u5c5e\u7684\u603b\u4f53\u5206\u5e03\u4e2d\u3002\u56e0\u6b64\uff0c\u8003\u8651\u7528\u4e8e\u8bad\u7ec3 base-model \u7684\u4f18\u5316\u5668\u5b66\u4e60\uff0c\u53ca\u5176\u4f7f\u7528\u7684\u6837\u672c\u548c\u6240\u5c5e\u7c7b\u522b\u662f\u5f88\u6709\u7528\u5904\u7684\u3002\u6bcf\u4e2a\u6837\u672c\u5c31\u4ee3\u8868\u4e00\u4e2a\u76ee\u6807\u51fd\u6570\uff0c\u5373\u4e00\u4e2a\u4efb\u52a1\u4e2d\u7528\u4e8e\u8bad\u7ec3 base-model \u7684\u635f\u5931\u51fd\u6570\u3002\u8be5\u4efb\u52a1\u7684\u7279\u70b9\u662f\u4e00\u7cfb\u5217\u6837\u672c\u548c\u76ee\u6807\u9884\u6d4b\uff0c\u6362\u8a00\u4e4b\uff0c\u76ee\u6807\u51fd\u6570\u5c31\u662f\u7528\u4e8e\u8bad\u7ec3 base-model \u7684\u6570\u636e\u96c6\u3002\u5143\u8bad\u7ec3\u96c6\uff08meta-training set\uff09\u5305\u62ec\u591a\u4e2a\u76ee\u6807\u51fd\u6570\uff0c\u5143\u6d4b\u8bd5\u96c6\u5305\u62ec\u540c\u4e00\u4e2a\u7c7b\u522b\u7684\u4e0d\u540c\u76ee\u6807\u51fd\u6570\u3002\u76ee\u6807\u51fd\u6570\u7684\u4e0d\u540c\u4e4b\u5904\u4f53\u73b0\u5728\u4e24\u4e2a\u65b9\u9762\uff1a\u5373\u5b83\u4eec\u5bf9\u5e94\u4e0d\u540c\u7684 base-model \u6216\u4e0d\u540c\u7684\u4efb\u52a1\u3002\u56e0\u6b64\uff0c\u672c\u6587\u6240\u6307\u6cdb\u5316\u6307\u5728\u4e0d\u540c base-model \u548c\uff0f\u6216\u4e0d\u540c\u4efb\u52a1\u4e0a\u8fd0\u884c\u7684\u5b66\u4e60\u4f18\u5316\u5668\u3002<\/span><\/p>\n<p><strong><span>\u6cdb\u5316\u4e3a\u4f55\u5982\u6b64\u91cd\u8981\uff1f<\/span><\/strong><\/p>\n<p><span>\u5047\u8bbe\u6211\u4eec\u4e0d\u5173\u5fc3\u6cdb\u5316\uff0c\u90a3\u4e48\u6211\u4eec\u8981\u5728\u7528\u4e8e\u8bad\u7ec3\u4f18\u5316\u5668\u7684\u540c\u4e00\u4e2a\u76ee\u6807\u51fd\u6570\u4e0a\u8bc4\u4f30\u4f18\u5316\u5668\u3002\u5982\u679c\u6211\u4eec\u4ec5\u4f7f\u7528\u4e00\u4e2a\u76ee\u6807\u51fd\u6570\uff0c\u90a3\u4e48\u6700\u4f73\u4f18\u5316\u5668\u5c31\u662f\u8bb0\u5fc6\u6700\u4f18\u89e3\u7684\u90a3\u4e2a\u4f18\u5316\u5668\uff1a\u8be5\u4f18\u5316\u5668\u901a\u5e38\u53ef\u5728\u4e00\u4e2a\u8fed\u4ee3\u6b65\u5185\u6536\u655b\u5230\u6700\u4f18\uff0c\u5e76\u4e14\u5728\u4efb\u610f\u521d\u59cb\u5316\u72b6\u6001\u4e0b\u90fd\u80fd\u4e00\u6b65\u5230\u8fbe\u3002\u5728\u672c\u6587\u4e2d\uff0c\u76ee\u6807\u51fd\u6570\u5bf9\u5e94\u5728\u7279\u5b9a\u4efb\u52a1\u4e2d\u8bad\u7ec3\u7279\u5b9a base-model \u7684\u635f\u5931\u51fd\u6570\uff0c\u56e0\u6b64\u672c\u8d28\u4e0a\uff0c\u8be5\u4f18\u5316\u5668\u8bb0\u5fc6\u4e86 base-model \u7684\u6700\u4f18\u6743\u91cd\u3002\u5373\u4f7f\u6211\u4eec\u4f7f\u7528\u4e86\u5f88\u591a\u76ee\u6807\u51fd\u6570\uff0c\u5b66\u4e60\u7684\u4f18\u5316\u5668\u4ecd\u7136\u80fd\u591f\u786e\u5b9a\u5b83\u8fd0\u884c\u4e86\u54ea\u4e2a\u76ee\u6807\u51fd\u6570\uff0c\u5e76\u5c3d\u5feb\u8df3\u8f6c\u81f3\u8bb0\u5fc6\u7684\u6700\u4f18\u89e3\u3002<\/span><\/p>\n<\/p>\n<p><span>\u8fd9\u53c8\u5b58\u5728\u4ec0\u4e48\u95ee\u9898\uff1f\u8bb0\u5fc6\u6700\u4f18\u9700\u8981\u5148\u627e\u5230\u6700\u4f18\uff0c\u56e0\u6b64\u5b66\u4e60\u4e00\u4e2a\u4f18\u5316\u5668\u9700\u8981\u7684\u65f6\u95f4\u6bd4\u8fd0\u884c\u4f20\u7edf\u7684\u4f18\u5316\u5668\u5982\u68af\u5ea6\u4e0b\u964d\u7684\u65f6\u95f4\u8981\u957f\u3002\u90a3\u4e48\uff0c\u4e3a\u4e86\u627e\u5230\u6700\u4f18\u76ee\u6807\u51fd\u6570\uff0c\u6267\u884c\u4f20\u7edf\u7684\u4f18\u5316\u5668\u8981\u66f4\u5feb\u4e00\u4e9b\u3002\u56e0\u6b64\uff0c\u5982\u679c\u6211\u4eec\u4e0d\u5173\u5fc3\u6cdb\u5316\uff0c\u5b66\u4e60\u4f18\u5316\u5668\u5c31\u6ca1\u6709\u610f\u4e49\u3002<\/span><\/p>\n<p><span>\u56e0\u6b64\uff0c\u8981\u4f7f\u5b66\u4e60\u4f18\u5316\u5668\u5177\u5907\u5b9e\u7528\u6027\uff0c\u5fc5\u987b\u4f7f\u5b83\u53ef\u4ee5\u6cdb\u5316\u5230\u4e0d\u540c\u7684\u65b0\u76ee\u6807\u51fd\u6570\u4e0a\uff0c\u4e14\u4e5f\u80fd\u6709\u4f18\u79c0\u7684\u8868\u73b0\u3002<\/span><\/p>\n<p><strong><span>\u6cdb\u5316\u7684\u5185\u5bb9\u5e94\u8be5\u662f\u4ec0\u4e48\uff1f<\/span><\/strong><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.fagao.me\/images\/uploads\/2019_010.jpg\" alt=\"\"><\/p>\n<p><span>\u5982\u679c\u6211\u4eec\u5bf9\u4e8e\u6cdb\u5316\u7684\u8981\u6c42\u662f\u5728\u7c7b\u4f3c\u4efb\u52a1\u4e0a\u8fbe\u5230\u7c7b\u4f3c base-model \u7684\u6548\u679c\uff0c\u90a3\u4e48\u5b66\u4e60\u7684\u4f18\u5316\u5668\u80fd\u591f\u8bb0\u5fc6 base-model \u548c\u4e00\u822c\u4efb\u52a1\u7684\u90e8\u5206\u6700\u4f18\u6743\u91cd\uff0c\u5982\u795e\u7ecf\u7f51\u7edc\u8f83\u4f4e\u5c42\u7ea7\u7684\u6743\u91cd\u3002\u8fd9\u5728\u672c\u8d28\u4e0a\u4e0e\u5b66\u4e60\u8981\u5b66\u4e60\u7684\u5185\u5bb9\u6784\u60f3\u76f8\u540c\uff0c\u6b63\u5982\u8fc1\u79fb\u5b66\u4e60\u90a3\u6837\u3002<\/span><\/p>\n<p><span>\u4e0e\u5b66\u4e60\u8981\u5b66\u4e60\u7684\u5185\u5bb9\u4e0d\u540c\u7684\u662f\uff0c\u5b66\u4e60\u5982\u4f55\u5b66\u4e60\u7684\u76ee\u6807\u4e0d\u662f\u5b66\u4e60\u4ec0\u4e48\u662f\u6700\u4f18\uff0c\u800c\u662f\u5982\u4f55\u627e\u5230\u6700\u4f18\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5fc5\u987b\u628a\u76ee\u6807\u8bbe\u7f6e\u4e3a\u66f4\u5f3a\u5927\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u5373\u5728\u4e0d\u7c7b\u4f3c\u4efb\u52a1\u4e0a\u6cdb\u5316\u5230\u7c7b\u4f3c\u7684 base-model \u4e2d\u3002\u80fd\u591f\u6cdb\u5316\u81f3\u4e0d\u7c7b\u4f3c\u4efb\u52a1\u7684\u4f18\u5316\u5668\u65e0\u6cd5\u90e8\u5206\u5730\u8bb0\u5fc6\u6700\u4f18\u6743\u91cd\uff0c\u56e0\u4e3a\u4e0d\u7c7b\u4f3c\u4efb\u52a1\u7684\u6700\u4f18\u6743\u91cd\u53ef\u80fd\u4f1a\u5b8c\u5168\u4e0d\u540c\u3002\u5982\u5728 MNIST\uff08\u5305\u62ec\u624b\u5199\u4f53\u6570\u5b57\u7684\u9ed1\u767d\u56fe\u50cf\u6570\u636e\u96c6\uff09\u548c CIFAR-10\uff08\u5305\u62ec\u81ea\u7136\u573a\u666f\u4e2d\u5e38\u89c1\u7269\u4f53\u7684\u5f69\u8272\u56fe\u50cf\u6570\u636e\u96c6\uff09\u4e0a\u8bad\u7ec3\u7684\u795e\u7ecf\u7f51\u7edc\u8f83\u4f4e\u5c42\u7ea7\u7684\u6743\u91cd\u751a\u81f3\u6709\u53ef\u80fd\u5b8c\u5168\u4e0d\u540c\u3002<\/span><\/p>\n<p><span>\u6211\u4eec\u5e94\u8be5\u628a\u76ee\u6807\u5b9a\u4e3a\u6bd4\u8fd9\u66f4\u5f3a\u7684\u6cdb\u5316\uff0c\u5373\u662f\u5426\u80fd\u5728\u4e0d\u7c7b\u4f3c\u4efb\u52a1\u4e0a\u6cdb\u5316\u5230\u4e0d\u7c7b\u4f3c\u7684 base-model \u4e2d\uff1f\u7531\u4e8e\u8fd9\u4e9b\u5bf9\u5e94\u7684\u76ee\u6807\u51fd\u6570\u4e0e\u8bad\u7ec3\u4f18\u5316\u5668\u6240\u7528\u7684\u76ee\u6807\u51fd\u6570\u5b8c\u5168\u4e0d\u540c\uff0c\u6240\u4ee5\u8fd9\u5b9e\u9645\u4e0a\u662f\u5728\u95ee\u5b66\u4e60\u4f18\u5316\u5668\u662f\u5426\u53ef\u6cdb\u5316\u81f3\u5b8c\u5168\u4e0d\u540c\u7684\u76ee\u6807\u51fd\u6570\u3002<\/span><\/p>\n<p><span>\u7b54\u6848\u662f\u4e0d\u53ef\u4ee5\u3002\u7ed9\u5b9a\u4efb\u610f\u4f18\u5316\u5668\uff0c\u6211\u4eec\u8003\u8651\u5728\u7279\u5b9a\u76ee\u6807\u51fd\u6570\u4e0a\u4f18\u5316\u5668\u540e\u7684\u8f68\u8ff9\u3002\u7531\u4e8e\u4f18\u5316\u5668\u4ec5\u4f9d\u8d56\u524d\u4e00\u6b21\u8fed\u4ee3\u7684\u4fe1\u606f\uff0c\u6240\u4ee5\u6211\u4eec\u53ef\u4ee5\u8c03\u6574\u6700\u540e\u4e00\u6b21\u8fed\u4ee3\u7684\u76ee\u6807\u51fd\u6570\uff0c\u4f7f\u4e4b\u53d1\u6563\uff08arbitrarily bad\uff09\uff0c\u540c\u65f6\u7ef4\u6301\u524d\u9762\u6240\u6709\u8fed\u4ee3\u4e2d\u76ee\u6807\u51fd\u6570\u7684\u51e0\u4f55\u7ed3\u6784\u3002\u56e0\u6b64\uff0c\u4efb\u4f55\u4f18\u5316\u5668\u90fd\u6709\u8868\u73b0\u5f88\u5dee\u7684\u76ee\u6807\u51fd\u6570\uff0c\u6ca1\u6709\u4f18\u5316\u5668\u80fd\u591f\u6cdb\u5316\u81f3\u6240\u6709\u76ee\u6807\u51fd\u6570\u3002<\/span><\/p>\n<\/p>\n<p><span>\u5982\u679c\u6ca1\u6709\u8868\u73b0\u666e\u904d\u826f\u597d\u7684\u4f18\u5316\u5668\uff0c\u6211\u4eec\u8fd8\u80fd\u591f\u671f\u5f85\u5b66\u4e60\u6709\u7528\u7684\u4f18\u5316\u5668\u5417\uff1f\u7b54\u6848\u662f yes\uff1a\u7531\u4e8e\u6211\u4eec\u5728\u5b9e\u8df5\u4e2d\u901a\u5e38\u5bf9\u4f18\u5316\u7279\u5b9a\u7c7b\u522b\u7684\u51fd\u6570\u611f\u5174\u8da3\uff0c\u6240\u4ee5\u5b66\u4e60\u5728\u8fd9\u4e9b\u7c7b\u522b\u4e2d\u6027\u80fd\u826f\u597d\u7684\u4f18\u5316\u5668\u8fd8\u662f\u53ef\u4ee5\u7684\u3002\u4e00\u4e2a\u7c7b\u522b\u4e2d\u7684\u76ee\u6807\u51fd\u6570\u53ef\u4ee5\u5171\u4eab\u51e0\u4f55\u7ed3\u6784\u4e2d\u7684\u89c4\u5219\uff0c\u5982\uff1a\u5b83\u4eec\u53ef\u4ee5\u5177\u6709\u5171\u540c\u7684\u51e0\u4f55\u7279\u6027\uff0c\u5982\u51f8\u6027\uff08convexity\uff09\u3001\u5206\u6bb5\u7ebf\u6027\u3001Lipschitz \u8fde\u7eed\u6216\u5176\u4ed6\u672a\u547d\u540d\u7684\u7279\u6027\u3002\u5728\u5b66\u4e60\u5982\u4f55\u5b66\u4e60\u7684\u80cc\u666f\u4e0b\uff0c\u6bcf\u4e2a\u7c7b\u522b\u80fd\u591f\u5bf9\u5e94\u4e00\u79cd base-model\u3002\u4f8b\u5982\uff0c\u5e26 ReLU \u6fc0\u6d3b\u5355\u5143\u7684\u795e\u7ecf\u7f51\u7edc\u662f\u4e00\u4e2a\u7c7b\u522b\uff0c\u56e0\u4e3a\u5b83\u4eec\u90fd\u662f\u5206\u6bb5\u7ebf\u6027\u3002\u6ce8\u610f\uff0c\u5728\u5b66\u4e60\u4f18\u5316\u5668\u7684\u65f6\u5019\uff0c\u6ca1\u5fc5\u8981\u660e\u786e\u5730\u63cf\u8ff0\u51e0\u4f55\u89c4\u5219\u7684\u5f62\u5f0f\uff0c\u56e0\u4e3a\u4f18\u5316\u5668\u53ef\u4ee5\u5728\u8be5\u7c7b\u522b\u7684\u76ee\u6807\u51fd\u6570\u4e0a\u8fdb\u884c\u8bad\u7ec3\u65f6\uff0c\u81ea\u52a8\u83b7\u53d6\u3002<\/span><\/p>\n<p><strong>\u5982\u4f55\u5b66\u4e60\u4f18\u5316\u5668<\/strong><\/p>\n<p><span>\u6211\u4eec\u5c1d\u8bd5\u7684\u7b2c\u4e00\u4e2a\u65b9\u6cd5\u662f\u5c06\u8be5\u95ee\u9898\u5f53\u4f5c\u4e00\u4e2a\u6807\u51c6\u7684\u76d1\u7763\u5b66\u4e60\u95ee\u9898\uff1a\u6211\u4eec\u4ec5\u533a\u5206\u4e0e\u66f4\u65b0\u516c\u5f0f\u76f8\u5173\u7684\u5143\u635f\u5931\uff08meta-loss\uff09\uff0c\u4f7f\u7528\u6807\u51c6\u7684\u57fa\u4e8e\u68af\u5ea6\u7684\u4f18\u5316\u65b9\u5f0f\u5b66\u4e60\u8fd9\u4e9b\u53c2\u6570\u3002\uff08\u6211\u4eec\u4e0d\u662f\u552f\u4e00\u8fd9\u4e48\u60f3\u7684\u4eba\uff1b\uff08Andrychowicz et al., 2016\uff09\u4e5f\u4f7f\u7528\u4e86\u7c7b\u4f3c\u7684\u65b9\u6cd5\u3002\uff09<\/span><\/p>\n<p><span>\u8fd9\u770b\u8d77\u6765\u662f\u975e\u5e38\u81ea\u7136\u7684\u65b9\u6cd5\uff0c\u4f46\u5e76\u4e0d\u594f\u6548\uff1a\u5c3d\u7ba1\u6211\u4eec\u5c3d\u4e86\u6700\u5927\u7684\u52aa\u529b\uff0c\u8fd8\u662f\u65e0\u6cd5\u7528\u8fd9\u79cd\u65b9\u5f0f\u8bad\u7ec3\u51fa\u53ef\u6cdb\u5316\u81f3\u672a\u77e5\u76ee\u6807\u51fd\u6570\u7684\u4f18\u5316\u5668\uff0c\u5373\u4f7f\u672a\u77e5\u76ee\u6807\u51fd\u6570\u662f\u4ece\u8bad\u7ec3\u4f18\u5316\u5668\u6240\u7528\u76ee\u6807\u51fd\u6570\u7684\u540c\u5206\u5e03\u4e2d\u9009\u53d6\u7684\u3002\u5728\u51e0\u4e4e\u6240\u6709\u672a\u77e5\u76ee\u6807\u51fd\u6570\u4e0a\uff0c\u5b66\u4e60\u7684\u4f18\u5316\u5668\u5f00\u59cb\u7684\u65f6\u5019\u4fdd\u6301\u6b63\u786e\u7684\u6536\u655b\u8d8b\u52bf\uff0c\u4f46\u5f88\u5feb\u5c31\u53d1\u6563\u4e86\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5728\u8bad\u7ec3\u7684\u76ee\u6807\u51fd\u6570\u4e0a\u6ca1\u6709\u8fd9\u6837\u7684\u95ee\u9898\uff0c\u6548\u679c\u5f88\u597d\u3002\u8fd9\u662f\u4ec0\u4e48\u539f\u56e0\u9020\u6210\u7684\u5462\uff1f<\/span><\/p>\n<\/p>\n<p><span>\u5b66\u4e60\u4f18\u5316\u5668\u5e76\u4e0d\u662f\u770b\u4e0a\u53bb\u90a3\u4e48\u7b80\u5355\u7684\u5b66\u4e60\u95ee\u9898\u3002\u6807\u51c6\u7684\u76d1\u7763\u5b66\u4e60\u5047\u8bbe\u6240\u6709\u7684\u8bad\u7ec3\u6837\u672c\u90fd\u662f\u72ec\u7acb\u540c\u5206\u5e03\u7684\uff08i.i.d.\uff09\uff1b\u5728\u6211\u4eec\u7684\u8bbe\u7f6e\u4e2d\uff0c\u4f18\u5316\u5668\u5728\u6bcf\u6b21\u8fed\u4ee3\u4e2d\u91c7\u7528\u7684\u8fed\u4ee3\u6b65\u5411\u91cf\u90fd\u4f1a\u5f71\u54cd\u5230\u4f18\u5316\u5668\u5728\u6240\u6709\u540e\u7eed\u8fed\u4ee3\u4e2d\u7ecf\u5386\u7684\u68af\u5ea6\u3002\u8fed\u4ee3\u6b65\u5411\u91cf\u5f71\u54cd\u540e\u7eed\u8fed\u4ee3\u4e2d\u68af\u5ea6\u7684\u65b9\u5f0f\u5c1a\u4e0d\u53ef\u77e5\uff0c\u56e0\u4e3a\u8fd9\u4f9d\u8d56\u4e8e\u76ee\u6807\u51fd\u6570\u7684\u5c40\u90e8\u51e0\u4f55\u7ed3\u6784\uff0c\u800c\u5728\u5143\u6d4b\u8bd5\u65f6\uff0c\u8fd9\u79cd\u51e0\u4f55\u7ed3\u6784\u4ecd\u662f\u672a\u77e5\u7684\u3002\u76d1\u7763\u5b66\u4e60\u4e0d\u80fd\u5728\u8be5\u8bbe\u7f6e\u4e2d\u8fd0\u884c\uff0c\u5fc5\u987b\u5047\u8bbe\u672a\u77e5\u76ee\u6807\u51fd\u6570\u7684\u5c40\u90e8\u51e0\u4f55\u7ed3\u6784\u4e0e\u8bad\u7ec3\u76ee\u6807\u51fd\u6570\u5728\u6240\u6709\u8fed\u4ee3\u4e2d\u7684\u5c40\u90e8\u51e0\u4f55\u7ed3\u6784\u76f8\u540c\u3002<\/span><\/p>\n<p><span>\u8003\u8651\u4e00\u4e0b\u5982\u679c\u4f7f\u7528\u76d1\u7763\u5b66\u4e60\u65b9\u5f0f\u8bad\u7ec3\u7684\u4f18\u5316\u5668\uff0c\u5e76\u7528\u4e8e\u672a\u77e5\u7684\u76ee\u6807\u51fd\u6570\u4e0a\u4f1a\u53d1\u751f\u4ec0\u4e48\u3002\u4f18\u5316\u5668\u8fed\u4ee3\u4e00\u6b65\uff0c\u5c31\u80fd\u53d1\u73b0\u4e0b\u4e00\u4e2a\u8fed\u4ee3\u4e2d\u7684\u68af\u5ea6\u4e0e\u5b83\u671f\u671b\u7684\u4e0d\u540c\u3002\u7136\u540e\u5f53\u4f18\u5316\u5668\u9047\u5230\u8fd9\u6837\u7684\u68af\u5ea6\u65f6\uff0c\u5b83\u5c31\u4f1a\u8bb0\u5fc6\u5728\u8bad\u7ec3\u76ee\u6807\u51fd\u6570\u65f6\u9047\u5230\u8fd9\u79cd\u68af\u5ea6\u8be5\u5982\u4f55\u5904\u7406\uff0c\u8be5\u8bb0\u5fc6\u53ef\u80fd\u5728\u7a7a\u95f4\u4e2d\u5b8c\u5168\u4e0d\u540c\u7684\u533a\u57df\u53d1\u751f\uff0c\u4f18\u5316\u5668\u4f1a\u636e\u6b64\u91c7\u53d6\u4e0b\u964d\u6b65\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u4f18\u5316\u5668\u4f1a\u53d1\u73b0\u4e0b\u4e00\u4e2a\u8fed\u4ee3\u7684\u68af\u5ea6\u4e0e\u5b83\u671f\u671b\u7684\u66f4\u52a0\u4e0d\u540c\uff0c\u56e0\u6b64\u8fd9\u79cd\u5faa\u73af\u4e0d\u65ad\u91cd\u590d\uff0c\u4f18\u5316\u5668\u9020\u6210\u7684\u8bef\u5dee\u968f\u7740\u65f6\u95f4\u8d8a\u6765\u8d8a\u5927\uff0c\u4e5f\u5c31\u5bfc\u81f4\u4e86\u5feb\u901f\u53d1\u6563\u3002<\/span><\/p>\n<p><span>\u6587\u732e\u4e2d\u5c06\u8fd9\u4e00\u73b0\u8c61\u79f0\u4f5c\u300c\u590d\u5408\u8bef\u5dee\u300d\uff08compounding error\uff09\u3002\u6211\u4eec\u90fd\u77e5\u9053\u76d1\u7763\u5b66\u4e60\u5668\u7684\u8bef\u5dee\u603b\u91cf\u662f\u8fed\u4ee3\u6570\u91cf\u7684\u4e8c\u500d\uff0c\u800c\u4e0d\u662f\u72ec\u7acb\u540c\u5206\u5e03\u8bbe\u7f6e\u4e2d\u7684\u7ebf\u6027\u60c5\u51b5\uff08Ross and Bagnell, 2010\uff09\u3002\u672c\u8d28\u4e0a\uff0c\u4f7f\u7528\u76d1\u7763\u5b66\u4e60\u8bad\u7ec3\u7684\u4f18\u5316\u5668\u52bf\u5fc5\u4f1a\u4e0e\u8bad\u7ec3\u76ee\u6807\u51fd\u6570\u7684\u51e0\u4f55\u7ed3\u6784\u4ea7\u751f\u8fc7\u62df\u5408\u3002\u89e3\u51b3\u8be5\u95ee\u9898\u7684\u4e00\u79cd\u65b9\u5f0f\u662f\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60\u3002<\/span><\/p>\n<\/p>\n<p><strong>\u5f3a\u5316\u5b66\u4e60\u7684\u80cc\u666f<\/strong><\/p>\n<p><span>\u8003\u8651\u4e00\u79cd\u73af\u5883\uff08environment\uff09\uff0c\u8be5\u73af\u5883\u7684\u72b6\u6001\uff08state\uff09\u6839\u636e\u91c7\u53d6\u7684\u52a8\u4f5c\uff08action\uff09\u4ee5\u4e00\u79cd\u672a\u77e5\u7684\u65b9\u5f0f\u53d1\u5c55\u3002\u6211\u4eec\u6709\u4e00\u4e2a\u53ef\u4ee5\u4e0e\u8fd9\u79cd\u73af\u5883\u4e92\u52a8\u7684\u667a\u80fd\u4f53\uff08agent\uff09\uff0c\u5b83\u8fde\u7eed\u9009\u62e9\u52a8\u4f5c\uff0c\u5e76\u5728\u6bcf\u4e2a\u52a8\u4f5c\u4e4b\u540e\u6536\u5230\u5173\u4e8e\u65b0\u72b6\u6001\u662f\u597d\u662f\u574f\u7684\u53cd\u9988\u3002\u5f3a\u5316\u5b66\u4e60\u7684\u76ee\u6807\u662f\u627e\u5230\u4e00\u79cd\u65b9\u6cd5\uff0c\u4f7f\u667a\u80fd\u4f53\u6839\u636e\u5f53\u524d\u72b6\u6001\u9009\u62e9\u52a8\u4f5c\uff0c\u4f7f\u72b6\u6001\u5728\u671f\u671b\u4e0a\u4fdd\u6301\u4f18\u79c0\u3002\u66f4\u51c6\u786e\u5730\u8bf4\uff0c\u5f3a\u5316\u5b66\u4e60\u95ee\u9898\u5177\u5907\u4ee5\u4e0b\u7ec4\u4ef6\uff1a<\/span><\/p>\n<ul>\n<li>\n<p><span>\u4e00\u4e2a\u72b6\u6001\u7a7a\u95f4\uff1a\u6240\u6709\u53ef\u80fd\u72b6\u6001\u7684\u96c6\u5408\uff1b<\/span><\/p>\n<\/li>\n<li>\n<p><span>\u4e00\u4e2a\u52a8\u4f5c\u7a7a\u95f4\uff0c\u6240\u6709\u53ef\u80fd\u52a8\u4f5c\u7684\u96c6\u5408\uff1b<\/span><\/p>\n<\/li>\n<li>\n<p><span>\u4e00\u4e2a\u6210\u672c\u51fd\u6570\uff0c\u7528\u6765\u8bc4\u4f30\u72b6\u6001\u6709\u591a\u7cdf\u7cd5\uff1b<\/span><\/p>\n<\/li>\n<li>\n<p><span>\u4e00\u4e2a\u65f6\u95f4\u8303\u56f4\uff0c\u4ee3\u8868\u65f6\u95f4\u6b65\u7684\u6570\u91cf\uff1b<\/span><\/p>\n<\/li>\n<li>\n<p><span>\u4e00\u4e2a\u521d\u59cb\u72b6\u6001\u6982\u7387\u5206\u5e03\uff0c\u6307\u5b9a\u6700\u521d\u91c7\u53d6\u4efb\u4f55\u52a8\u4f5c\u4e4b\u524d\u4e0d\u540c\u72b6\u6001\u7684\u51fa\u73b0\u9891\u7387\uff1b<\/span><\/p>\n<\/li>\n<li>\n<p><span>\u4e00\u4e2a\u72b6\u6001\u8f6c\u6362\u6982\u7387\u5206\u5e03\uff0c\u6307\u5b9a\u7279\u5b9a\u52a8\u4f5c\u4e4b\u540e\u72b6\u6001\u7684\uff08\u53ef\u80fd\uff09\u6539\u53d8\u3002<\/span><\/p>\n<\/li>\n<\/ul>\n<p><span>\u5373\u4f7f\u5b66\u4e60\u7b97\u6cd5\u610f\u8bc6\u5230\u4e0a\u9762 5 \u4e2a\u7ec4\u4ef6\u7684\u5185\u5bb9\uff0c\u5b83\u4e5f\u4e0d\u77e5\u9053\u6700\u540e\u4e00\u4e2a\u7ec4\u4ef6\uff1a\u72b6\u6001\u5982\u4f55\u6839\u636e\u9009\u62e9\u7684\u52a8\u4f5c\u53d1\u751f\u53d8\u5316\u3002\u5728\u8bad\u7ec3\u4e2d\uff0c\u5b66\u4e60\u7b97\u6cd5\u53ef\u4ee5\u4e0e\u73af\u5883\u4e92\u52a8\u3002\u5c24\u5176\u662f\uff0c\u5728\u6bcf\u4e2a\u65f6\u95f4\u6b65\u4e0a\uff0c\u5b83\u53ef\u4ee5\u6839\u636e\u5f53\u524d\u72b6\u6001\u9009\u62e9\u52a8\u4f5c\u3002\u7136\u540e\uff0c\u6839\u636e\u9009\u62e9\u7684\u52a8\u4f5c\u4ee5\u53ca\u5f53\u524d\u7684\u72b6\u6001\uff0c\u73af\u5883\u91c7\u6837\u4e86\u4e00\u79cd\u65b0\u72b6\u6001\uff0c\u8be5\u72b6\u6001\u7531\u5b66\u4e60\u7b97\u6cd5\u5728\u540e\u7eed\u7684\u65f6\u95f4\u6b65\u4e0a\u89c2\u5bdf\u5230\u3002\u91c7\u6837\u7684\u72b6\u6001\u5e8f\u5217\u548c\u52a8\u4f5c\u5e8f\u5217\u5c31\u662f\u8f68\u8ff9\u3002\u91c7\u6837\u8fc7\u7a0b\u5f15\u5165\u4e86\u4e00\u79cd\u4f9d\u8d56\u4e8e\u521d\u59cb\u72b6\u6001\u548c\u8f6c\u6362\u6982\u7387\u7684\u8f68\u8ff9\u5206\u5e03\uff0c\u4ee5\u53ca\u6839\u636e\u5f53\u524d\u72b6\u6001\u9009\u62e9\u52a8\u4f5c\u7684\u65b9\u5f0f\uff0c\u540e\u8005\u53eb\u4f5c\u7b56\u7565\uff08policy\uff09\u3002\u8be5\u7b56\u7565\u901a\u5e38\u5efa\u6a21\u4e3a\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\uff0c\u5728\u5f53\u524d\u72b6\u6001\u4e2d\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u8f93\u51fa\u52a8\u4f5c\u3002\u8be5\u5b66\u4e60\u7b97\u6cd5\u7684\u76ee\u6807\u662f\u627e\u5230\u4e00\u4e2a\u7b56\u7565\uff0c\u4f7f\u6240\u6709\u65f6\u95f4\u6b65\u4e0a\u72b6\u6001\u7684\u671f\u671b\u7d2f\u8ba1\u6210\u672c\u6700\u5c0f\u5316\uff0c\u671f\u671b\u4e0e\u8f68\u8ff9\u5206\u5e03\u6709\u5173\u3002<\/span><\/p>\n<p><strong>\u5f62\u5f0f\u5316\u4e3a\u5f3a\u5316\u5b66\u4e60\u95ee\u9898<\/strong><\/p>\n<p><span>\u56de\u5230\u6211\u4eec\u4e0a\u6587\u4e2d\u4ecb\u7ecd\u7684\u5b66\u4e60\u6846\u67b6\uff0c\u5176\u76ee\u6807\u662f\u627e\u5230\u4f7f\u5143\u635f\u5931\u6700\u5c0f\u5316\u7684\u66f4\u65b0\u516c\u5f0f\u3002\u76f4\u89c2\u6765\u770b\uff0c\u6211\u4eec\u5c06\u667a\u80fd\u4f53\u770b\u4f5c\u4e00\u4e2a\u4f18\u5316\u7b97\u6cd5\uff0c\u5c06\u73af\u5883\u770b\u4f5c\u7531\u6211\u4eec\u60f3\u8ba9\u5b66\u4e60\u7684\u4f18\u5316\u5668\u9762\u5bf9\u7684\u76ee\u6807\u51fd\u6570\u96c6\u5408\u3002\u72b6\u6001\u5305\u62ec\u5f53\u524d\u8fed\u4ee3\u548c\u76ee\u524d\u4f18\u5316\u8f68\u8ff9\u7684\u4e00\u4e9b\u7279\u5f81\uff0c\u8fd9\u53ef\u4ee5\u662f\u5386\u53f2\u68af\u5ea6\u3001\u8fed\u4ee3\u6570\u548c\u76ee\u6807\u503c\u7684\u7edf\u8ba1\u3002\u52a8\u4f5c\u662f\u7528\u4e8e\u66f4\u65b0\u8fed\u4ee3\u7684\u8fed\u4ee3\u6b65\u5411\u91cf\u3002<\/span><\/p>\n<\/p>\n<p><span>\u5728\u8be5\u516c\u5f0f\u4e2d\uff0c\u7b56\u7565\u5b9e\u9645\u4e0a\u662f\u8ba1\u7b97\u52a8\u4f5c\uff08\u5373\u8fed\u4ee3\u6b65\u5411\u91cf\uff09\u7684\u6b65\u9aa4\uff0c\u4f9d\u8d56\u4e8e\u5f53\u524d\u8fed\u4ee3\u548c\u5386\u53f2\u68af\u5ea6\u3001\u8fed\u4ee3\u6570\u548c\u76ee\u6807\u503c\u3002\u6362\u8a00\u4e4b\uff0c\u7279\u5b9a\u7684\u7b56\u7565\u4ee3\u8868\u7279\u5b9a\u7684\u66f4\u65b0\u516c\u5f0f\u3002<\/span><\/p>\n<p><span>\u56e0\u6b64\uff0c\u5b66\u4e60\u8be5\u7b56\u7565\u7b49\u4e8e\u5b66\u4e60\u66f4\u65b0\u516c\u5f0f\u4ee5\u53ca\u4f18\u5316\u7b97\u6cd5\u3002\u521d\u59cb\u72b6\u6001\u6982\u7387\u5206\u5e03\u662f\u521d\u59cb\u8fed\u4ee3\u3001\u68af\u5ea6\u548c\u76ee\u6807\u503c\u7684\u8054\u5408\u5206\u5e03\u3002\u72b6\u6001\u8f6c\u6362\u6982\u7387\u5206\u5e03\u8bb0\u5f55\u4e86\u5728\u5f53\u524d\u72b6\u6001\u548c\u52a8\u4f5c\u4e0b\u4e00\u4e2a\u53ef\u80fd\u7684\u72b6\u6001\u3002\u7531\u4e8e\u8be5\u72b6\u6001\u5305\u62ec\u68af\u5ea6\u548c\u76ee\u6807\u503c\uff0c\u56e0\u6b64\u72b6\u6001\u8f6c\u6362\u6982\u7387\u5206\u5e03\u80fd\u591f\u6355\u6349\u5230\u68af\u5ea6\u548c\u76ee\u6807\u503c\u5728\u4efb\u610f\u7ed9\u5b9a\u7684\u8fed\u4ee3\u6b65\u5411\u91cf\u4e0b\u7684\u53d8\u5316\u8d8b\u52bf\u3002\u6362\u8a00\u4e4b\uff0c\u5b83\u5bf9\u611f\u5174\u8da3\u7684\u76ee\u6807\u51fd\u6570\u7684\u5c40\u90e8\u51e0\u4f55\u5173\u7cfb\u8fdb\u884c\u7f16\u7801\u3002\u5173\u952e\u662f\uff0c\u8fd9\u79cd\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u4e0d\u80fd\u76f4\u63a5\u8bbf\u95ee\u72b6\u6001\u8f6c\u6362\u6982\u7387\u5206\u5e03\uff0c\u56e0\u6b64\u5b83\u5b66\u4e60\u7684\u7b56\u7565\u4e0d\u4f1a\u4e0e\u8bad\u7ec3\u76ee\u6807\u51fd\u6570\u7684\u51e0\u4f55\u7ed3\u6784\u4ea7\u751f\u8fc7\u62df\u5408\u3002<\/span><\/p>\n<p><span>\u6211\u4eec\u9009\u62e9\u4e00\u4e2a\u72b6\u6001\u7684\u4ee3\u4ef7\u51fd\u6570\u4f5c\u4e3a\u5f53\u524d\u8fed\u4ee3\u8bc4\u4f30\u7684\u76ee\u6807\u51fd\u6570\u503c\u3002\u7531\u4e8e\u5f3a\u5316\u5b66\u4e60\u4f7f\u6240\u6709\u65f6\u95f4\u6b65\u4e0a\u7684\u7d2f\u8ba1\u6210\u672c\u6700\u5c0f\u5316\uff0c\u56e0\u6b64\u5b83\u672c\u8d28\u4e0a\u4f7f\u6240\u6709\u8fed\u4ee3\u4e2d\u7684\u76ee\u6807\u503c\u603b\u6570\uff08\u5373\u5143\u635f\u5931\uff09\u6700\u5c0f\u5316\u3002<\/span><\/p>\n<p><strong>\u7ed3\u679c<\/strong><\/p>\n<p><span>\u6211\u4eec\u5c31\u5982\u540c\u5728 MNIST \u4e0a\u8bad\u7ec3\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u800c\u8bad\u7ec3\u4f18\u5316\u7b97\u6cd5\uff0c\u5e76\u5c31\u5728 Toronto Faces Dataset\uff08TFD\uff09\u3001CIFAR-10 \u548c CIFAR-100 \u4e0a\u8bad\u7ec3\u4e0d\u540c\u7684\u795e\u7ecf\u7f51\u7edc\u7684\u95ee\u9898\u5bf9\u4f18\u5316\u7b97\u6cd5\u8fdb\u884c\u6d4b\u8bd5\u3002\u8fd9\u4e9b\u6570\u636e\u96c6\u4e92\u76f8\u6ca1\u6709\u4efb\u4f55\u76f8\u4f3c\u6027\uff1aMNIST \u5305\u62ec\u624b\u5199\u4f53\u6570\u5b57\u7684\u9ed1\u767d\u56fe\u50cf\uff0cTFD \u5305\u62ec\u4eba\u8138\u7684\u7070\u5ea6\u56fe\u50cf\uff0cCIFAR-10\/100 \u5305\u62ec\u81ea\u7136\u573a\u666f\u5e38\u89c1\u7269\u4f53\u7684\u5f69\u8272\u56fe\u50cf\u3002\u56e0\u6b64\u5b66\u4e60\u7684\u4f18\u5316\u7b97\u6cd5\u4e0d\u53ef\u80fd\u9003\u907f\u4f7f\u7528\u8bb0\u5fc6\uff0c\u6bd4\u5982\u8bf4\uff0cMNIST \u4e0a\u8f83\u4f4e\u5c42\u7684\u6743\u91cd\uff0c\u5e76\u4e14\u5728 TFD \u548c CIFAR-10\/100 \u4e0a\u4ecd\u7136\u8868\u73b0\u4e0d\u9519\u3002<\/span><\/p>\n<\/p>\n<p><span>\u5982\u56fe\u6240\u793a\uff0c\u4f7f\u7528\u6211\u4eec\u7684\u7b97\u6cd5\u5728 MNIST\uff08\u6d45\u7ea2\u8272\uff09\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u7684\u4f18\u5316\u7b97\u6cd5\u53ef\u6cdb\u5316\u81f3 TFD\u3001CIFAR-10 \u548c CIFAR-100\uff0c\u4e14\u8868\u73b0\u4f18\u4e8e\u5176\u4ed6\u4f18\u5316\u7b97\u6cd5\u3002<\/span><\/p>\n<p><span>\u4e3a\u4e86\u7406\u89e3\u4f7f\u7528\u6211\u4eec\u7684\u65b9\u6cd5\u5b66\u4e60\u7684\u4f18\u5316\u7b97\u6cd5\u7684\u884c\u4e3a\uff0c\u6211\u4eec\u5728\u4e8c\u7ef4 logistic \u56de\u5f52\u95ee\u9898\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u4f18\u5316\u7b97\u6cd5\uff0c\u5e76\u5728\u53c2\u6570\u7a7a\u95f4\u4e2d\u53ef\u89c6\u5316\u5176\u8f68\u8ff9\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4f18\u5316\u7b97\u6cd5\u5728\u4f4e\u7ef4\u548c\u9ad8\u7ef4\u4e2d\u7684\u884c\u4e3a\u4e0d\u540c\uff0c\u56e0\u6b64\u4e0b\u56fe\u4e2d\u7684\u53ef\u89c6\u5316\u672a\u5fc5\u8868\u793a\u4f18\u5316\u7b97\u6cd5\u5728\u9ad8\u7ef4\u4e2d\u7684\u884c\u4e3a\u3002\u4f46\u662f\uff0c\u5b83\u4eec\u63d0\u4f9b\u4e86\u4e00\u4e9b\u6709\u7528\u7684\u5173\u4e8e\u53ef\u4ee5\u5b66\u5230\u7684\u884c\u4e3a\u7c7b\u578b\u7684\u7ecf\u9a8c\u3002<\/span><\/p>\n<\/p>\n<p><span>\u4e0a\u56fe\u663e\u793a\u51fa\u4e24\u4e2a\u4e0d\u540c\u7684\u672a\u77e5 logistic \u56de\u5f52\u95ee\u9898\u4e0a\u4e0d\u540c\u7b97\u6cd5\u7684\u4f18\u5316\u8f68\u8ff9\u3002\u6bcf\u4e2a\u7bad\u5934\u4ee3\u8868\u4f18\u5316\u7b97\u6cd5\u7684\u4e00\u6b21\u8fed\u4ee3\u3002\u5982\u56fe\u6240\u793a\uff0c\u4f7f\u7528\u6211\u4eec\u7684\u65b9\u6cd5\u5b66\u5230\u7684\u7b97\u6cd5\uff08\u6d45\u7ea2\u8272\uff09\u6bd4\u5176\u4ed6\u7b97\u6cd5\u7684\u6b65\u957f\u66f4\u5927\uff0c\u5728\u4e24\u6b21\u8fed\u4ee3\u540e\u8d85\u8fc7\u5176\u4ed6\u7b97\u6cd5\uff0c\u4f46\u662f\u6ca1\u6709\u632f\u8361\uff0c\u800c\u662f\u91c7\u7528\u66f4\u5c0f\u7684\u6b65\u957f\u6765\u9010\u6e10\u56de\u6536\u3002\u5728\u7b2c\u4e8c\u4e2a\u793a\u4f8b\u4e2d\uff0c\u7531\u4e8e\u68af\u5ea6\u6d88\u5931\uff0c\u4f20\u7edf\u7684\u4f18\u5316\u7b97\u6cd5\u6b65\u957f\u8f83\u5c0f\uff0c\u5e76\u56e0\u6b64\u6536\u655b\u8f83\u6162\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5b66\u5f97\u7684\u7b97\u6cd5\u6b65\u957f\u66f4\u5927\uff0c\u6536\u655b\u66f4\u5feb\u3002<\/span><\/p>\n<p><span><em><span>\u539f\u6587\u94fe\u63a5\uff1ahttp:\/\/bair.berkeley.edu\/blog\/2017\/09\/12\/learning-to-optimize-with-rl\/<\/span><\/em><\/span><\/p>\n<p><strong><em><span><strong><em><span>\u672c\u6587\u4e3a\u673a\u5668\u4e4b\u5fc3\u7f16\u8bd1\uff0c<strong><em><span>\u8f6c\u8f7d\u8bf7\u8054\u7cfb\u672c\u516c\u4f17\u53f7\u83b7\u5f97\u6388\u6743<\/span><\/em><\/strong><\/span><\/em><\/strong>\u3002<\/span><\/em><\/strong><\/p>\n<div class=\"articleend\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u9009\u81ea BAIR \u53c2\u4e0e\uff1a\u8def\u96ea\u3001\u674e\u6cfd\u5357\u3001\u848b\u601d\u6e90 \u81ea\u4ece\u53bb\u5e74 UC Berkeley \u8bba\u6587\u300aLearning to O [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[177],"class_list":["post-118171","post","type-post","status-publish","format-standard","hentry","category-news","tag-177"],"_links":{"self":[{"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/posts\/118171","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/comments?post=118171"}],"version-history":[{"count":0,"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/posts\/118171\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/media?parent=118171"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/categories?post=118171"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.fagao.me\/p\/wp-json\/wp\/v2\/tags?post=118171"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}