{"id":4791,"date":"2024-10-08T11:04:17","date_gmt":"2024-10-08T03:04:17","guid":{"rendered":"https:\/\/nullthought.net\/?p=4791"},"modified":"2025-03-03T12:25:31","modified_gmt":"2025-03-03T04:25:31","slug":"%e4%b8%80%e7%af%87%e8%80%81%e8%ae%ba%e6%96%87%ef%bc%8c%e5%85%b3%e4%ba%8e%e7%9f%a5%e8%af%86%e8%92%b8%e9%a6%8f%ef%bc%88distilling%ef%bc%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=4791","title":{"rendered":"\u8bfa\u5956\u5f97\u4e3bGeoffrey Hinton\u7684\u4e00\u7bc7\u8001\u8bba\u6587\uff0c\u5173\u4e8e\u77e5\u8bc6\u84b8\u998f\uff08Distilling\uff09"},"content":{"rendered":"\n<p><strong><a href=\"https:\/\/arxiv.org\/abs\/1503.02531\" target=\"_blank\" rel=\"noreferrer noopener\">Distilling the Knowledge in a Neural Network<\/a><\/strong>\u662f<strong><a href=\"https:\/\/en.wikipedia.org\/wiki\/Geoffrey_Hinton\" target=\"_blank\" rel=\"noreferrer noopener\">Geoffrey Hinton<\/a>\u3001Oriol Vinyals\u548cJeff Dean<\/strong>\u4e8e2015\u5e74\u53d1\u8868\u7684\u4e00\u7bc7\u201c\u8001\u201d\u8bba\u6587\u3002\u4e09\u4f4d\u4f5c\u8005\u5747\u4e3aAI\u754c\u7684\u5927\u795e\u3002<a href=\"https:\/\/www.nobelprize.org\/prizes\/physics\/2024\/press-release\/\" target=\"_blank\" rel=\"noreferrer noopener\"><strong>Geoffrey Hinton<\/strong>\u4eca\u5929\u521a\u521a\u83b7\u5f97\u4e862024\u5e74\u5ea6\u8bfa\u8d1d\u5c14\u7269\u7406\u5b66\u5956<\/a>\uff0c\u4ed6\u548c<strong>John J. Hopfield<\/strong>\u540c\u65f6\u83b7\u5956\u3002<strong>Geoffrey Hinton<\/strong>\u4e5f\u6210\u4e3a\u8bfa\u5956\u548c\u56fe\u7075\u5956\u7684\u53cc\u6599\u5f97\u4e3b\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"has-small-font-size\">This year\u2019s two Nobel Laureates in Physics have used tools from physics to develop methods that are the foundation of today\u2019s powerful machine learning. John Hopfield created an associative memory that can store and reconstruct images and other types of patterns in data. Geoffrey Hinton invented a method that can autonomously find properties in data, and so perform tasks such as identifying specific elements in pictures.<br>&#8212;<br>\u4eca\u5e74\u7684\u4e24\u4f4d\u8bfa\u8d1d\u5c14\u7269\u7406\u5b66\u5956\u5f97\u4e3b\u4f7f\u7528\u7269\u7406\u5b66\u7684\u5de5\u5177\u5f00\u53d1\u4e86\u5960\u5b9a\u5f53\u4eca\u5f3a\u5927\u673a\u5668\u5b66\u4e60\u57fa\u7840\u7684\u65b9\u6cd5\u3002\u7ea6\u7ff0\u00b7\u970d\u666e\u83f2\u5c14\u5fb7\u521b\u9020\u4e86\u4e00\u79cd\u8054\u60f3\u8bb0\u5fc6\uff0c\u53ef\u4ee5\u5b58\u50a8\u548c\u91cd\u6784\u56fe\u50cf\u53ca\u5176\u4ed6\u7c7b\u578b\u7684\u6570\u636e\u6a21\u5f0f\u3002\u6770\u5f17\u91cc\u00b7\u8f9b\u987f\u53d1\u660e\u4e86\u4e00\u79cd\u80fd\u591f\u81ea\u4e3b\u53d1\u73b0\u6570\u636e\u7279\u6027\u7684\u7b97\u6cd5\uff0c\u4ece\u800c\u6267\u884c\u8bf8\u5982\u8bc6\u522b\u56fe\u7247\u4e2d\u7279\u5b9a\u5143\u7d20\u7684\u4efb\u52a1\u3002<\/p>\n\n\n\n<p><a href=\"https:\/\/www.nobelprize.org\/prizes\/physics\/2024\/press-release\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/www.nobelprize.org\/prizes\/physics\/2024\/press-release<\/a><\/p>\n<\/blockquote>\n\n\n\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/1503.02531\" target=\"_blank\" rel=\"noreferrer noopener\">Distilling the Knowledge in a Neural Network<\/a><\/strong>\u63d0\u51fa\u4e86\u4e00\u79cd\u901a\u8fc7\u201c\u77e5\u8bc6\u84b8\u998f\u201d\uff08Distillation\uff09\u6280\u672f\uff0c\u5c06\u5927\u578b\u795e\u7ecf\u7f51\u7edc\u6216\u6a21\u578b\u96c6\u5408\uff08ensemble models\uff09\u4e2d\u7684\u77e5\u8bc6\u538b\u7f29\u5e76\u4f20\u9012\u7ed9\u8f83\u5c0f\u6a21\u578b\u7684\u65b9\u6cd5\u3002\u8be5\u65b9\u6cd5\u7684\u76ee\u6807\u662f\u901a\u8fc7\u51cf\u5c11\u8ba1\u7b97\u8d44\u6e90\u9700\u6c42\u7684\u540c\u65f6\uff0c\u4fdd\u6301\u539f\u6709\u590d\u6742\u6a21\u578b\u7684\u826f\u597d\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u4ee5\u4e0b\u662f\u5bf9\u8bba\u6587\u7684\u66f4\u8be6\u7ec6\u89e3\u8bfb\uff1a<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u5f15\u8a00<\/h4>\n\n\n\n<p>\u4f20\u7edf\u4e0a\uff0c\u4e3a\u4e86\u63d0\u9ad8\u673a\u5668\u5b66\u4e60\u6a21\u578b\u7684\u6027\u80fd\uff0c\u4eba\u4eec\u4f1a\u4f7f\u7528\u201c\u6a21\u578b\u96c6\u5408\u201d\uff08ensemble models\uff09\uff0c\u5373\u8bad\u7ec3\u591a\u4e2a\u4e0d\u540c\u7684\u6a21\u578b\u5e76\u7ed3\u5408\u5b83\u4eec\u7684\u9884\u6d4b\u7ed3\u679c\u3002\u8fd9\u79cd\u65b9\u6cd5\u80fd\u591f\u63d0\u9ad8\u6cdb\u5316\u80fd\u529b\u548c\u6027\u80fd\u3002\u7136\u800c\uff0c\u6a21\u578b\u96c6\u5408\u901a\u5e38\u975e\u5e38\u5e9e\u5927\uff0c\u8ba1\u7b97\u6210\u672c\u9ad8\uff0c\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u96be\u4ee5\u90e8\u7f72\u3002\u4e3a\u4e86\u514b\u670d\u8fd9\u4e00\u95ee\u9898\uff0cCaruana\u7b49\u4eba\u63d0\u51fa\u4e86\u201c\u6a21\u578b\u538b\u7f29\u201d\uff08Model Compression\uff09\uff0c\u901a\u8fc7\u5c06\u6a21\u578b\u96c6\u5408\u7684\u77e5\u8bc6\u538b\u7f29\u5230\u5355\u4e00\u6a21\u578b\u4e2d\uff0c\u51cf\u5c11\u8ba1\u7b97\u8d1f\u62c5\u3002<\/p>\n\n\n\n<p>\u672c\u6587\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5229\u7528\u4e00\u79cd\u66f4\u4e3a\u9ad8\u6548\u7684\u201c\u77e5\u8bc6\u84b8\u998f\u201d\u6280\u672f\uff0c\u5c06\u6a21\u578b\u96c6\u5408\u4e2d\u7684\u77e5\u8bc6\u8f6c\u79fb\u5230\u4e00\u4e2a\u66f4\u5c0f\u7684\u6a21\u578b\u4e2d\uff0c\u8fd9\u6837\u5c0f\u6a21\u578b\u53ef\u4ee5\u5728\u4e0d\u663e\u8457\u727a\u7272\u6027\u80fd\u7684\u60c5\u51b5\u4e0b\uff0c\u66f4\u9002\u5408\u5b9e\u9645\u90e8\u7f72\u3002\u4f5c\u8005\u901a\u8fc7\u5728MNIST\u624b\u5199\u6570\u5b57\u8bc6\u522b\u4efb\u52a1\u548c\u8bed\u97f3\u8bc6\u522b\u4efb\u52a1\u4e0a\u8fdb\u884c\u5b9e\u9a8c\uff0c\u9a8c\u8bc1\u4e86\u8be5\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. \u77e5\u8bc6\u84b8\u998f\u7684\u57fa\u672c\u539f\u7406<\/h4>\n\n\n\n<p>\u5728\u8bad\u7ec3\u590d\u6742\u7684\u795e\u7ecf\u7f51\u7edc\u65f6\uff0c\u6a21\u578b\u5e76\u4e0d\u4ec5\u4ec5\u5b66\u4e60\u5982\u4f55\u5bf9\u8f93\u5165\u8fdb\u884c\u5206\u7c7b\uff0c\u5b83\u8fd8\u5b66\u4e60\u4e86\u5927\u91cf\u5173\u4e8e\u8f93\u5165\u6837\u672c\u7684\u7ec6\u5fae\u4fe1\u606f\uff0c\u4f8b\u5982\u7c7b\u522b\u4e4b\u95f4\u7684\u76f8\u4f3c\u6027\u3002\u4f20\u7edf\u8bad\u7ec3\u65b9\u6cd5\u901a\u5e38\u5173\u6ce8\u6b63\u786e\u6807\u7b7e\uff08hard targets\uff09\uff0c\u5373\u5c06\u8f93\u5165\u5f52\u4e3a\u67d0\u4e2a\u56fa\u5b9a\u7684\u7c7b\u522b\uff0c\u4f46\u8fd9\u79cd\u65b9\u6cd5\u5ffd\u7565\u4e86\u6a21\u578b\u5728\u9519\u8bef\u5206\u7c7b\u4e0a\u7684\u6982\u7387\u5206\u5e03\uff08soft targets\uff09\u3002\u8fd9\u4e9b\u201c\u9519\u8bef\u201d\u7684\u9884\u6d4b\u4fe1\u606f\u5b9e\u9645\u4e0a\u53cd\u6620\u4e86\u6a21\u578b\u5bf9\u7c7b\u522b\u4e4b\u95f4\u5173\u7cfb\u7684\u7406\u89e3\u3002\u77e5\u8bc6\u84b8\u998f\u7684\u6838\u5fc3\u601d\u60f3\u6b63\u662f\u5229\u7528\u8fd9\u79cd\u989d\u5916\u7684\u4fe1\u606f\u3002<\/p>\n\n\n\n<p><strong>\u8f6f\u76ee\u6807\uff08Soft Targets\uff09<\/strong>\uff1a\u5927\u578b\u795e\u7ecf\u7f51\u7edc\u901a\u8fc7softmax\u5c42\u8f93\u51fa\u6982\u7387\u5206\u5e03\uff0c\u8fd9\u4e0d\u4ec5\u4ec5\u544a\u8bc9\u6211\u4eec\u6b63\u786e\u7c7b\u522b\u7684\u6982\u7387\uff0c\u8fd8\u63ed\u793a\u4e86\u9519\u8bef\u7c7b\u522b\u4e4b\u95f4\u7684\u6982\u7387\u5173\u7cfb\u3002\u84b8\u998f\u65b9\u6cd5\u901a\u8fc7\u4f7f\u7528\u8f6f\u76ee\u6807\uff0c\u5c06\u8fd9\u79cd\u4fe1\u606f\u4ece\u5927\u6a21\u578b\u4f20\u9012\u7ed9\u5c0f\u6a21\u578b\u3002\u5c0f\u6a21\u578b\u4e0d\u4ec5\u4ec5\u5b66\u4e60\u6b63\u786e\u7b54\u6848\uff0c\u8fd8\u5b66\u4e60\u5230\u5927\u6a21\u578b\u5982\u4f55\u5728\u9519\u8bef\u7c7b\u522b\u4e4b\u95f4\u8fdb\u884c\u63a8\u65ad\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u6e29\u5ea6\u53c2\u6570\uff08Temperature\uff09<\/h5>\n\n\n\n<p>\u4e3a\u4e86\u8ba9\u5927\u6a21\u578b\u8f93\u51fa\u7684\u6982\u7387\u5206\u5e03\u66f4\u5e73\u6ed1\u3001\u66f4\u5bb9\u6613\u5b66\u4e60\uff0c\u4f5c\u8005\u5f15\u5165\u4e86\u201c\u6e29\u5ea6\u53c2\u6570\u201d\uff08T\uff09\u3002\u901a\u8fc7\u5c06softmax\u51fd\u6570\u4e2d\u7684\u6e29\u5ea6\u53c2\u6570\u8bbe\u7f6e\u4e3a\u4e00\u4e2a\u8f83\u9ad8\u7684\u503c\uff0c\u5927\u6a21\u578b\u7684\u8f93\u51fa\u4f1a\u66f4\u52a0\u5e73\u6ed1\uff0c\u4ece\u800c\u63d0\u4f9b\u66f4\u4e30\u5bcc\u7684\u7c7b\u522b\u4fe1\u606f\u3002\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"270\" height=\"95\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-3.png\" alt=\"\" class=\"wp-image-4792\"\/><\/figure>\n\n\n\n<p>\u8fd9\u91cc\uff0cz<sub>i<\/sub>\u662f\u7c7bi\u7684logit\u503c\uff0cT\u4e3a\u6e29\u5ea6\u3002\u5f53T\u8f83\u9ad8\u65f6\uff0c\u7c7b\u522b\u4e4b\u95f4\u7684\u6982\u7387\u5206\u5e03\u66f4\u5747\u5300\uff1b\u5f53T=1\u65f6\uff0csoftmax\u51fd\u6570\u6062\u590d\u6b63\u5e38\u3002<\/p>\n\n\n\n<p><strong>\u5c0f\u6a21\u578b\u7684\u8bad\u7ec3<\/strong>\uff1a\u5c0f\u6a21\u578b\u901a\u8fc7\u5b66\u4e60\u5927\u6a21\u578b\u5728\u9ad8\u6e29\u5ea6\u4e0b\u751f\u6210\u7684\u8f6f\u76ee\u6807\u8fdb\u884c\u8bad\u7ec3\u3002\u5728\u8bad\u7ec3\u9636\u6bb5\uff0c\u5c0f\u6a21\u578b\u7684softmax\u4e5f\u4f7f\u7528\u76f8\u540c\u7684\u9ad8\u6e29\u5ea6\uff0c\u4f46\u5728\u6700\u7ec8\u90e8\u7f72\u65f6\u4f7f\u7528\u6807\u51c6\u7684T=1\u3002\u9664\u4e86\u8f6f\u76ee\u6807\u5916\uff0c\u4f5c\u8005\u8fd8\u5efa\u8bae\u5c06\u6b63\u786e\u6807\u7b7e\u4f5c\u4e3a\u786c\u76ee\u6807\u8fdb\u884c\u8bad\u7ec3\uff0c\u7ed3\u5408\u4e24\u8005\uff0c\u53ef\u4ee5\u63d0\u5347\u6a21\u578b\u7684\u51c6\u786e\u6027\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u77e5\u8bc6\u84b8\u998f\u7684\u4f18\u52bf<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4fe1\u606f\u5bc6\u5ea6\u9ad8<\/strong>\uff1a\u8f6f\u76ee\u6807\u4f20\u9012\u4e86\u6bd4\u786c\u76ee\u6807\u66f4\u591a\u7684\u4fe1\u606f\u3002\u5373\u4f7f\u662f\u9519\u8bef\u7684\u9884\u6d4b\uff0c\u8f6f\u76ee\u6807\u4e2d\u5305\u542b\u7684\u7c7b\u522b\u4e4b\u95f4\u7684\u76f8\u4f3c\u6027\u4e5f\u80fd\u5e2e\u52a9\u5c0f\u6a21\u578b\u66f4\u597d\u5730\u7406\u89e3\u6570\u636e\u3002<\/li>\n\n\n\n<li><strong>\u51cf\u5c11\u6570\u636e\u91cf\u9700\u6c42<\/strong>\uff1a\u901a\u8fc7\u84b8\u998f\u6280\u672f\uff0c\u5c0f\u6a21\u578b\u53ef\u4ee5\u5728\u8f83\u5c11\u7684\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u56e0\u4e3a\u8f6f\u76ee\u6807\u63d0\u4f9b\u4e86\u66f4\u4e30\u5bcc\u7684\u7c7b\u522b\u5173\u7cfb\u4fe1\u606f\u3002<\/li>\n\n\n\n<li><strong>\u63d0\u5347\u6cdb\u5316\u80fd\u529b<\/strong>\uff1a\u5c0f\u6a21\u578b\u901a\u8fc7\u5b66\u4e60\u5927\u6a21\u578b\u7684\u8f6f\u76ee\u6807\uff0c\u4e0d\u4ec5\u80fd\u5728\u8bad\u7ec3\u6570\u636e\u4e0a\u8868\u73b0\u826f\u597d\uff0c\u4e5f\u80fd\u5728\u6d4b\u8bd5\u6570\u636e\u4e0a\u6709\u66f4\u597d\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">3. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\">3.1 MNIST\u624b\u5199\u6570\u5b57\u8bc6\u522b\u4efb\u52a1<\/h5>\n\n\n\n<p>MNIST\u6570\u636e\u96c6\u5305\u542b60000\u5f20\u8bad\u7ec3\u56fe\u50cf\u548c10000\u5f20\u6d4b\u8bd5\u56fe\u50cf\uff0c\u4efb\u52a1\u662f\u8bc6\u522b0\u52309\u7684\u624b\u5199\u6570\u5b57\u3002\u4f5c\u8005\u5728MNIST\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u5b9e\u9a8c\uff0c\u4ee5\u9a8c\u8bc1\u77e5\u8bc6\u84b8\u998f\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5927\u6a21\u578b<\/strong>\uff1a\u4f5c\u8005\u8bad\u7ec3\u4e86\u4e00\u4e2a\u5927\u578b\u795e\u7ecf\u7f51\u7edc\uff0c\u5305\u542b\u4e24\u4e2a\u9690\u85cf\u5c42\uff0c\u6bcf\u5c42\u67091200\u4e2aReLU\u6fc0\u6d3b\u7684\u9690\u85cf\u5355\u5143\u3002\u8be5\u6a21\u578b\u4f7f\u7528dropout\u8fdb\u884c\u5f3a\u6b63\u5219\u5316\uff0c\u5e76\u5728\u8bad\u7ec3\u65f6\u5bf9\u8f93\u5165\u56fe\u50cf\u8fdb\u884c\u968f\u673a\u6296\u52a8\uff08translation jittering\uff09\u3002\u8be5\u5927\u6a21\u578b\u5728\u6d4b\u8bd5\u96c6\u4e0a\u53d6\u5f97\u4e8667\u4e2a\u9519\u8bef\u3002<\/li>\n\n\n\n<li><strong>\u5c0f\u6a21\u578b<\/strong>\uff1a\u4e00\u4e2a\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u5305\u542b\u4e24\u4e2a\u9690\u85cf\u5c42\uff0c\u6bcf\u5c42800\u4e2a\u5355\u5143\uff0c\u5728\u6ca1\u6709\u6b63\u5219\u5316\u7684\u60c5\u51b5\u4e0b\uff0c\u6d4b\u8bd5\u96c6\u4e0a\u7684\u9519\u8bef\u6570\u4e3a146\u4e2a\u3002<\/li>\n\n\n\n<li><strong>\u84b8\u998f\u540e\u7684\u5c0f\u6a21\u578b<\/strong>\uff1a\u5c06\u5927\u6a21\u578b\u7684\u8f6f\u76ee\u6807\u7528\u4f5c\u8bad\u7ec3\u5c0f\u6a21\u578b\u7684\u6307\u5bfc\uff0c\u5373\u4f7f\u6ca1\u6709\u8fdb\u884c\u6b63\u5219\u5316\uff0c\u5c0f\u6a21\u578b\u4e5f\u80fd\u5c06\u6d4b\u8bd5\u9519\u8bef\u6570\u51cf\u5c11\u523074\u4e2a\u3002\u8fd9\u8868\u660e\u8f6f\u76ee\u6807\u80fd\u591f\u6709\u6548\u4f20\u9012\u5927\u6a21\u578b\u4e2d\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6b64\u5916\uff0c\u4f5c\u8005\u8fd8\u8fdb\u884c\u4e86\u4e0d\u540c\u6e29\u5ea6\u53c2\u6570\u4e0b\u7684\u5b9e\u9a8c\uff0c\u53d1\u73b0\u5f53\u5c0f\u6a21\u578b\u7684\u9690\u85cf\u5355\u5143\u6570\u91cf\u51cf\u5c11\u65f6\uff0c\u8f83\u4f4e\u7684\u6e29\u5ea6\uff08\u5982T=2.5\u52304\uff09\u6548\u679c\u66f4\u597d\u3002\u5373\u4f7f\u5728\u8bad\u7ec3\u96c6\u4e2d\u5b8c\u5168\u53bb\u9664\u67d0\u4e9b\u6570\u5b57\uff08\u5982\u6570\u5b573\uff09\uff0c\u84b8\u998f\u540e\u7684\u5c0f\u6a21\u578b\u4ecd\u80fd\u8f83\u597d\u5730\u8bc6\u522b\u8fd9\u4e9b\u6570\u5b57\uff0c\u901a\u8fc7\u8c03\u6574\u504f\u7f6e\uff0c\u5b83\u751a\u81f3\u80fd\u591f\u5728\u6ca1\u6709\u89c1\u8fc7\u8bad\u7ec3\u6837\u672c\u7684\u60c5\u51b5\u4e0b\u8fbe\u523098.6%\u7684\u6b63\u786e\u7387\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">3.2 \u8bed\u97f3\u8bc6\u522b\u4efb\u52a1<\/h5>\n\n\n\n<p>\u8bed\u97f3\u8bc6\u522b\u4efb\u52a1\u4e2d\uff0c\u4f5c\u8005\u4f7f\u7528\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\uff08DNN\uff09\u8fdb\u884c\u58f0\u5b66\u6a21\u578b\u7684\u8bad\u7ec3\u3002DNN\u7684\u8f93\u5165\u4e3a\u7ecf\u8fc7\u6885\u5c14\u6ee4\u6ce2\u7684\u58f0\u5b66\u7279\u5f81\uff0c\u8f93\u51fa\u4e3a\u9690\u9a6c\u5c14\u53ef\u592b\u6a21\u578b\uff08HMM\uff09\u7684\u72b6\u6001\u6982\u7387\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5927\u6a21\u578b<\/strong>\uff1a\u75318\u5c42\uff0c\u6bcf\u5c422560\u4e2aReLU\u5355\u5143\u7ec4\u6210\uff0c\u6700\u7ec8\u8f93\u51fa\u5c42\u4e3a14,000\u4e2a\u6807\u7b7e\u3002\u8be5\u6a21\u578b\u4f7f\u7528\u5927\u89c4\u6a21\u8bed\u97f3\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u6d4b\u8bd5\u96c6\u4e0a\u7684\u8bcd\u9519\u8bef\u7387\uff08WER\uff09\u4e3a10.9%\u3002<\/li>\n\n\n\n<li><strong>\u84b8\u998f\u6a21\u578b<\/strong>\uff1a\u5c06\u5927\u6a21\u578b\u7684\u8f6f\u76ee\u6807\u7528\u4e8e\u8bad\u7ec3\u4e00\u4e2a\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u5b9e\u9a8c\u8868\u660e\uff0c\u84b8\u998f\u540e\u7684\u5c0f\u6a21\u578b\u5728\u5e27\u7ea7\u522b\u5206\u7c7b\u51c6\u786e\u7387\u4e0a\u8fbe\u5230\u4e8660.8%\uff0c\u4e0e10\u4e2a\u5927\u6a21\u578b\u96c6\u5408\u7684\u5e73\u5747\u9884\u6d4b\u7ed3\u679c\uff0861.1%\uff09\u76f8\u8fd1\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">3.3 \u4e13\u5bb6\u6a21\u578b\uff08Specialist Models\uff09<\/h5>\n\n\n\n<p>\u5bf9\u4e8e\u8d85\u5927\u89c4\u6a21\u6570\u636e\u96c6\uff08\u5982\u8c37\u6b4c\u7684JFT\u6570\u636e\u96c6\uff0c\u5305\u542b1\u4ebf\u5f20\u56fe\u50cf\uff0c15,000\u4e2a\u7c7b\u522b\uff09\uff0c\u8bad\u7ec3\u5355\u4e00\u5927\u6a21\u578b\u7684\u65f6\u95f4\u6210\u672c\u6781\u9ad8\uff0c\u751a\u81f3\u65e0\u6cd5\u5b8c\u6210\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u201c\u4e13\u5bb6\u6a21\u578b\u201d\u6982\u5ff5\uff0c\u5373\u901a\u8fc7\u591a\u4e2a\u5c0f\u6a21\u578b\u6765\u5206\u522b\u5904\u7406\u67d0\u4e9b\u5bb9\u6613\u6df7\u6dc6\u7684\u7c7b\u522b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6a21\u578b\u8bad\u7ec3<\/strong>\uff1a\u6bcf\u4e2a\u4e13\u5bb6\u6a21\u578b\u4e13\u6ce8\u4e8e\u67d0\u4e2a\u7c7b\u522b\u7684\u5b50\u96c6\uff0c\u5e76\u4e0e\u603b\u6a21\u578b\uff08generalist model\uff09\u5171\u4eab\u90e8\u5206\u6743\u91cd\u3002\u901a\u8fc7\u5bf9\u8fd9\u4e9b\u5b50\u96c6\u8fdb\u884c\u8f6f\u76ee\u6807\u8bad\u7ec3\uff0c\u53ef\u4ee5\u907f\u514d\u8fc7\u62df\u5408\uff0c\u5e76\u5927\u5e45\u51cf\u5c11\u8bad\u7ec3\u65f6\u95f4\u3002<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u8fc7\u7a0b<\/strong>\uff1a\u63a8\u7406\u9636\u6bb5\uff0c\u9996\u5148\u7531\u603b\u6a21\u578b\u5224\u65ad\u54ea\u4e9b\u7c7b\u522b\u76f8\u5173\uff0c\u7136\u540e\u8c03\u7528\u76f8\u5e94\u7684\u4e13\u5bb6\u6a21\u578b\u8fdb\u884c\u7ec6\u7c92\u5ea6\u7684\u5206\u7c7b\u3002\u8fd9\u79cd\u65b9\u6cd5\u80fd\u591f\u63d0\u9ad8\u63a8\u7406\u7684\u51c6\u786e\u6027\u548c\u6548\u7387\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">4. \u8f6f\u76ee\u6807\u4f5c\u4e3a\u6b63\u5219\u5316\u5668<\/h4>\n\n\n\n<p>\u8bba\u6587\u8fdb\u4e00\u6b65\u8ba8\u8bba\u4e86\u8f6f\u76ee\u6807\u5728\u5c0f\u6570\u636e\u96c6\u4e0a\u4f5c\u4e3a\u6b63\u5219\u5316\u5668\u7684\u4f5c\u7528\u3002\u4f5c\u8005\u53d1\u73b0\uff0c\u5f53\u4ec5\u4f7f\u75283%\u7684\u8bad\u7ec3\u6570\u636e\u65f6\uff0c\u76f4\u63a5\u4f7f\u7528\u786c\u6807\u7b7e\u8bad\u7ec3\u4f1a\u5bfc\u81f4\u4e25\u91cd\u7684\u8fc7\u62df\u5408\uff0c\u800c\u4f7f\u7528\u8f6f\u76ee\u6807\u8bad\u7ec3\u7684\u5c0f\u6a21\u578b\u80fd\u591f\u6709\u6548\u907f\u514d\u8fd9\u79cd\u60c5\u51b5\u3002\u5373\u4f7f\u53ea\u4f7f\u7528\u5c11\u91cf\u6570\u636e\uff0c\u8f6f\u76ee\u6807\u4e5f\u80fd\u591f\u5e2e\u52a9\u6a21\u578b\u4ece\u4e2d\u63d0\u53d6\u5230\u4e30\u5bcc\u7684\u4fe1\u606f\uff0c\u63d0\u5347\u6cdb\u5316\u6027\u80fd\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">5. \u4e0e\u4e13\u5bb6\u6a21\u578b\uff08Mixture of Experts\uff09\u7684\u5173\u7cfb<\/h4>\n\n\n\n<p>\u8bba\u6587\u8fd8\u63a2\u8ba8\u4e86\u77e5\u8bc6\u84b8\u998f\u4e0e\u4f20\u7edf\u7684\u4e13\u5bb6\u6a21\u578b\u65b9\u6cd5\u7684\u5173\u7cfb\u3002\u4f20\u7edf\u7684\u4e13\u5bb6\u6a21\u578b\u901a\u8fc7\u4e00\u4e2a\u95e8\u63a7\u7f51\u7edc\uff08gating network\uff09\u6765\u4e3a\u6bcf\u4e2a\u6837\u672c\u9009\u62e9\u5408\u9002\u7684\u4e13\u5bb6\uff0c\u800c\u77e5\u8bc6\u84b8\u998f\u4e2d\u7684\u4e13\u5bb6\u6a21\u578b\u5219\u4e0d\u9700\u8981\u8fd9\u6837\u7684\u590d\u6742\u7ed3\u6784\uff0c\u4e13\u5bb6\u53ef\u4ee5\u72ec\u7acb\u8bad\u7ec3\uff0c\u5e76\u901a\u8fc7\u603b\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u6765\u51b3\u5b9a\u54ea\u4e9b\u4e13\u5bb6\u53c2\u4e0e\u63a8\u7406\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">6. \u603b\u7ed3\u4e0e\u8ba8\u8bba<\/h4>\n\n\n\n<p>\u4f5c\u8005\u603b\u7ed3\u4e86\u77e5\u8bc6\u84b8\u998f\u7684\u4f18\u52bf\uff0c\u5f3a\u8c03\u4e86\u8be5\u65b9\u6cd5\u5728\u5c06\u5927\u6a21\u578b\u6216\u6a21\u578b\u96c6\u5408\u7684\u77e5\u8bc6\u538b\u7f29\u5230\u5c0f\u6a21\u578b\u4e2d\u7684\u6f5c\u529b\u3002\u901a\u8fc7\u8fd9\u79cd\u6280\u672f\uff0c\u6a21\u578b\u5728\u4e0d\u663e\u8457\u727a\u7272\u6027\u80fd\u7684\u524d\u63d0\u4e0b\uff0c\u80fd\u591f\u5927\u5e45\u51cf\u5c11\u8ba1\u7b97\u5f00\u9500\uff0c\u9002\u5408\u5b9e\u9645\u90e8\u7f72\u573a\u666f\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u8bba\u6587\u8fd8\u5c55\u671b\u4e86\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\uff0c\u4f8b\u5982\u5982\u4f55\u5c06\u4e13\u5bb6\u6a21\u578b\u7684\u77e5\u8bc6\u8fdb\u4e00\u6b65\u84b8\u998f\u5230\u5355\u4e2a\u5927\u6a21\u578b\u4e2d\uff0c\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6a21\u578b\u7684\u6027\u80fd\u548c\u6548\u7387\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u6392\u7248\u6781\u4e3a\u6734\u7d20\u7684\u8bba\u6587\u8bf4\u660eSlides\uff1a<a href=\"https:\/\/www.ttic.edu\/dl\/dark14.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/www.ttic.edu\/dl\/dark14.pdf<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Distilling the Knowledge in a Neural Network\u662fGeoffrey H [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8,36],"tags":[39,86,97,66],"class_list":["post-4791","post","type-post","status-publish","format-standard","hentry","category-tech","category-36","tag-ai","tag-llm","tag-distillation","tag-google"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a> <a href=\"https:\/\/nullthought.net\/?cat=36\" rel=\"category\">\u79d1\u5b66<\/a>","rttpg_excerpt":"Distilling the Knowledge in a Neural Network\u662fGeoffrey H&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4791","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4791"}],"version-history":[{"count":6,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4791\/revisions"}],"predecessor-version":[{"id":4800,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4791\/revisions\/4800"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4791"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4791"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4791"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}