{"id":38631,"date":"2025-02-19T23:00:56","date_gmt":"2025-02-19T15:00:56","guid":{"rendered":"https:\/\/17aitech.com\/?p=38631"},"modified":"2025-02-19T23:00:56","modified_gmt":"2025-02-19T15:00:56","slug":"springer%e7%9f%a5%e8%af%86%e8%92%b8%e9%a6%8f%e4%b8%93%e8%91%97%e8%a7%a3%e8%af%bb-%e9%9d%a2%e5%90%91%e5%9b%be%e5%83%8f%e8%af%86%e5%88%ab%e7%9a%84%e7%9f%a5%e8%af%86%e8%92%b8%e9%a6%8f%e7%bb%bc%e8%bf%b0","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=38631","title":{"rendered":"Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-02-19-12\" target=\"_blank\">Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0<\/a><\/p>\n<p data-first-child=\"\" data-pid=\"HdeW2Yz-\">\u672c\u6b21\u6587\u7ae0\u4ecb\u7ecd\u6211\u4eec\u53d1\u8868\u4e8e\u7531Springer\u51fa\u7248\u7684\u4e13\u8457\u300a<a data-za-detail-view-id=\"1043\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/link.springer.com\/book\/10.1007\/978-3-031-32095-8\" rel=\"nofollow noreferrer\" target=\"_blank\">Advancements in Knowledge Distillation: Towards New Horizons of Intelligent Systems<\/a> \u300b\u4e2d\u7684\u7b2c\u4e00\u7ae0\u201cCategories of Response-Based, Feature-Based, and Relation-Based Knowledge Distillation\u201d\u3002\u8be5\u7bc7\u6587\u7ae0\u7684\u4e3b\u8981\u5185\u5bb9\u662f\u6574\u7406\u4e86\u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7684\u76f8\u5173\u5de5\u4f5c\uff0c\u9996\u5148\u5728<strong>response-based\u3001feature-based\u548crelation-based<\/strong>\u4e09\u79cd\u77e5\u8bc6\u5f62\u5f0f\u6765\u4ecb\u7ecd\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\u7684\u76f8\u5173\u5de5\u4f5c\uff0c\u7136\u540e\u6574\u7406\u4e86<strong>\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f<\/strong>\u548c<strong>\u81ea\u77e5\u8bc6\u84b8\u998f<\/strong>\u7684\u76f8\u5173\u5de5\u4f5c\uff0c\u5728\u5176\u4e2d\u4e5f\u5bf9<strong>\u81ea\u76d1\u7763\u5b66\u4e60\u84b8\u998f<\/strong>\u548c<strong><a data-paste-text=\"true\" data-za-not-track-link=\"true\" href=\"https:\/\/zhida.zhihu.com\/search?content_id=235760946&amp;content_type=Article&amp;match_order=1&amp;q=%E8%A7%86%E8%A7%89Transformer&amp;zhida_source=entity\" target=\"_blank\">\u89c6\u89c9Transformer<\/a>\uff08ViT\uff09\u84b8\u998f<\/strong>\u4e5f\u8fdb\u884c\u4e86\u4ecb\u7ecd\u3002\u6700\u540e\u8bb2\u89e3\u4e86\u6269\u5c55\u7684\u84b8\u998f\u6280\u672f\u9886\u57df\uff0c\u5305\u62ec<strong>\u591a\u6559\u5e08\u77e5\u8bc6\u84b8\u998f\u3001\u8de8\u6a21\u6001\u77e5\u8bc6\u84b8\u998f\u3001<a data-paste-text=\"true\" data-za-not-track-link=\"true\" href=\"https:\/\/zhida.zhihu.com\/search?content_id=235760946&amp;content_type=Article&amp;match_order=1&amp;q=%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%9C%BA%E5%88%B6%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F&amp;zhida_source=entity\" target=\"_blank\">\u6ce8\u610f\u529b\u673a\u5236\u77e5\u8bc6\u84b8\u998f<\/a>\u3001<a data-paste-text=\"true\" data-za-not-track-link=\"true\" href=\"https:\/\/zhida.zhihu.com\/search?content_id=235760946&amp;content_type=Article&amp;match_order=1&amp;q=%E6%97%A0%E6%95%B0%E6%8D%AE%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F&amp;zhida_source=entity\" target=\"_blank\">\u65e0\u6570\u636e\u77e5\u8bc6\u84b8\u998f<\/a>\u548c\u5bf9\u6297\u77e5\u8bc6\u84b8\u998f<\/strong>\u3002\u672c\u89e3\u8bfb\u4e3b\u8981\u8bb2\u89e3\u6838\u5fc3\u5185\u5bb9\u548c\u89c1\u89e3\uff0c\u5177\u4f53\u7684\u53c2\u8003\u6587\u732e\u8bf7\u89c1\u539f\u8457\u4f5c\u3002<\/p>\n<p data-pid=\"EGo6EpeZ\">\u8bba\u6587\u5b98\u65b9\u5730\u5740\uff1a<a data-za-detail-view-id=\"1043\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/link.springer.com\/chapter\/10.1007\/978-3-031-32095-8_1\" rel=\"nofollow noreferrer\" target=\"_blank\">https:\/\/link.springer.com\/chapter\/10.1007\/978-3-031-32095-8_1<\/a><\/p>\n<p data-pid=\"X_sZoGaP\">arXiv\u5730\u5740\uff1a<a data-za-detail-view-id=\"1043\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/arxiv.org\/pdf\/2306.10687.pdf\" rel=\"nofollow noreferrer\" target=\"_blank\">https:\/\/arxiv.org\/pdf\/2306.10687.pdf<\/a><\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c807a03a5f3d24b73ad14f2a22d7e2d3.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c807a03a5f3d24b73ad14f2a22d7e2d3.jpg\"><\/a><\/figure>\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_78 ez-toc-wrap-left-text counter-hierarchy ez-toc-counter ez-toc-light-blue ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">\u6587\u7ae0\u76ee\u5f55<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"Toggle Table of Content\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Toggle<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewBox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewBox=\"0 0 24 24\" version=\"1.2\" baseProfile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1 ' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/17aitech.com\/?p=38631\/#1_Categories_of_response-based_feature-based_and_relation-based_knowledge_distillation_%E5%9F%BA%E4%BA%8E%E5%93%8D%E5%BA%94%E3%80%81%E7%89%B9%E5%BE%81%E3%80%81%E5%85%B3%E7%B3%BB%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%E5%88%92%E5%88%86\" >1. Categories of response-based, feature-based, and relation-based knowledge distillation. (\u57fa\u4e8e\u54cd\u5e94\u3001\u7279\u5f81\u3001\u5173\u7cfb\u7684\u77e5\u8bc6\u84b8\u998f\u5212\u5206)<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/17aitech.com\/?p=38631\/#11_Response-based_Knowledge_Distillation%EF%BC%88%E5%9F%BA%E4%BA%8E%E5%93%8D%E5%BA%94%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >1.1 Response-based Knowledge Distillation\uff08\u57fa\u4e8e\u54cd\u5e94\u7684\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/17aitech.com\/?p=38631\/#12_Feature-based_Knowledge_Distillation%EF%BC%88%E5%9F%BA%E4%BA%8E%E7%89%B9%E5%BE%81%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >1.2 Feature-based Knowledge Distillation\uff08\u57fa\u4e8e\u7279\u5f81\u7684\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/17aitech.com\/?p=38631\/#13_Relation-based_Knowledge_Distillation%EF%BC%88%E5%9F%BA%E4%BA%8E%E5%85%B3%E7%B3%BB%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >1.3 Relation-based Knowledge Distillation\uff08\u57fa\u4e8e\u5173\u7cfb\u7684\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/17aitech.com\/?p=38631\/#2_Distillation_Schemes_%EF%BC%88%E8%92%B8%E9%A6%8F%E6%9C%BA%E5%88%B6%EF%BC%89\" >2. Distillation Schemes \uff08\u84b8\u998f\u673a\u5236\uff09<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/17aitech.com\/?p=38631\/#21_Offline_Knowledge_Distillation%EF%BC%88%E7%A6%BB%E7%BA%BF%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >2.1 Offline Knowledge Distillation\uff08\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/17aitech.com\/?p=38631\/#22_Online_Knowledge_Distillation%EF%BC%88%E5%9C%A8%E7%BA%BF%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >2.2 Online Knowledge Distillation\uff08\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/17aitech.com\/?p=38631\/#23_Self-Knowledge_Distillation%EF%BC%88%E8%87%AA%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >2.3 Self-Knowledge Distillation\uff08\u81ea\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/17aitech.com\/?p=38631\/#24_%E4%B8%89%E7%A7%8D%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%E7%BB%BC%E5%90%88%E6%AF%94%E8%BE%83\" >2.4 \u4e09\u79cd\u77e5\u8bc6\u84b8\u998f\u7efc\u5408\u6bd4\u8f83<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/17aitech.com\/?p=38631\/#3_Distillation_Algorithms_%EF%BC%88%E8%92%B8%E9%A6%8F%E7%AE%97%E6%B3%95%EF%BC%89\" >3. Distillation Algorithms \uff08\u84b8\u998f\u7b97\u6cd5\uff09<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-11\" href=\"https:\/\/17aitech.com\/?p=38631\/#31_Multi-Teacher_Distillation%EF%BC%88%E5%A4%9A%E6%95%99%E5%B8%88%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >3.1 Multi-Teacher Distillation\uff08\u591a\u6559\u5e08\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-12\" href=\"https:\/\/17aitech.com\/?p=38631\/#32_Cross-Modal_Distillation%EF%BC%88%E8%B7%A8%E6%A8%A1%E6%80%81%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >3.2 Cross-Modal Distillation\uff08\u8de8\u6a21\u6001\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-13\" href=\"https:\/\/17aitech.com\/?p=38631\/#33_Attention-based_Distillation%EF%BC%88%E5%9F%BA%E4%BA%8E%E6%B3%A8%E6%84%8F%E5%8A%9B%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >3.3 Attention-based Distillation\uff08\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-14\" href=\"https:\/\/17aitech.com\/?p=38631\/#34_Data-free_Distillation%EF%BC%88%E6%97%A0%E9%9C%80%E6%95%B0%E6%8D%AE%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\" >3.4 Data-free Distillation\uff08\u65e0\u9700\u6570\u636e\u7684\u77e5\u8bc6\u84b8\u998f\uff09<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-15\" href=\"https:\/\/17aitech.com\/?p=38631\/#35_Adversarial_Distillation_%EF%BC%88%E5%AF%B9%E6%8A%97%E8%92%B8%E9%A6%8F%EF%BC%89\" >3.5 Adversarial Distillation \uff08\u5bf9\u6297\u84b8\u998f\uff09<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-16\" href=\"https:\/\/17aitech.com\/?p=38631\/#4%E6%80%BB%E7%BB%93\" >4.\u603b\u7ed3<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-17\" href=\"https:\/\/17aitech.com\/?p=38631\/#%E5%8F%82%E8%80%83%E6%96%87%E7%8C%AE\" >\u53c2\u8003\u6587\u732e<\/a><\/li><\/ul><\/nav><\/div>\n<h2><span class=\"ez-toc-section\" id=\"1_Categories_of_response-based_feature-based_and_relation-based_knowledge_distillation_%E5%9F%BA%E4%BA%8E%E5%93%8D%E5%BA%94%E3%80%81%E7%89%B9%E5%BE%81%E3%80%81%E5%85%B3%E7%B3%BB%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%E5%88%92%E5%88%86\"><\/span>1. Categories of response-based, feature-based, and relation-based knowledge distillation. (\u57fa\u4e8e\u54cd\u5e94\u3001\u7279\u5f81\u3001\u5173\u7cfb\u7684\u77e5\u8bc6\u84b8\u998f\u5212\u5206)<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p data-pid=\"kYwjmj1_\">\u76ee\u524d\u7684\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u901a\u5e38\u6d89\u53ca\u77e5\u8bc6\u7c7b\u578b\u548c\u84b8\u998f\u7b56\u7565\u3002\u524d\u8005\u4fa7\u91cd\u4e8e\u63a2\u7d22\u5b66\u751f\u6a21\u4eff\u7684\u5404\u79cd\u4fe1\u606f\u7c7b\u578b\u3002\u540e\u8005\u65e8\u5728\u5e2e\u52a9\u5b66\u751f\u6709\u6548\u5730\u5b66\u4e60\u8001\u5e08\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u7814\u7a76\u57fa\u4e8e\u54cd\u5e94\u3001\u57fa\u4e8e\u7279\u5f81\u548c\u57fa\u4e8e\u5173\u7cfb\u7684\u77e5\u8bc6\uff0c\u4ee5\u53ca\u57fa\u4e8e\u9884\u5b9a\u4e49\u4fe1\u606f\u7684\u5e38\u7528\u7c7b\u578b\u3002\u57fa\u4e8e\u54cd\u5e94\u7684KD\u5f15\u5bfc\u6559\u5e08\u7684\u6700\u7ec8\u8f93\u51fa\u4ee5\u6307\u5bfc\u5b66\u751f\u7684\u8f93\u51fa\u3002\u76f4\u89c2\u4e0a\uff0c\u5b66\u751f\u53ef\u4ee5\u76f4\u63a5\u5b66\u4e60\u5230\u8001\u5e08\u4ea7\u751f\u7684\u9884\u6d4b\u3002\u9664\u4e86\u6700\u7ec8\u8f93\u51fa\u5916\uff0c\u4e2d\u95f4\u7279\u5f81\u8fd8\u5bf9\u795e\u7ecf\u7f51\u7edc\u7684\u77e5\u8bc6\u63d0\u53d6\u8fc7\u7a0b\u8fdb\u884c\u7f16\u7801\u3002\u57fa\u4e8e\u7279\u5f81\u7684KD\u53ef\u4ee5\u6559\u5b66\u751f\u5728\u6574\u4e2a\u9690\u85cf\u5c42\u4e2d\u83b7\u5f97\u66f4\u6709\u610f\u4e49\u7684\u8bed\u4e49\u4fe1\u606f\u3002\u57fa\u4e8e\u54cd\u5e94\u548c\u57fa\u4e8e\u7279\u5f81\u7684KD\u901a\u5e38\u8003\u8651\u4ece\u5355\u4e2a\u6570\u636e\u6837\u672c\u4e2d\u63d0\u53d6\u77e5\u8bc6\u3002\u76f8\u53cd\uff0c\u57fa\u4e8e\u5173\u7cfb\u7684KD\u8bd5\u56fe\u6316\u6398\u6574\u4e2a\u6570\u636e\u96c6\u7684\u8de8\u6837\u672c\u5173\u7cfb\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u56de\u987e\u4e86\u6bcf\u79cd\u77e5\u8bc6\u7c7b\u578b\u7684\u4e00\u4e9b\u4ee3\u8868\u6027\u65b9\u6cd5\uff0c\u5e76\u603b\u7ed3\u4e86\u5b83\u4eec\u7684\u5dee\u5f02\u3002\u793a\u610f\u56fe\u6982\u8ff0\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-365ec9057310f1fca63d73d27d615791.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-365ec9057310f1fca63d73d27d615791.jpg\"><\/a><figcaption>\u56fe1 \u57fa\u4e8e\u54cd\u5e94\u3001\u7279\u5f81\u3001\u5173\u7cfb\u7684\u6559\u5e08-\u5b66\u751f\u77e5\u8bc6\u84b8\u998f\u6574\u4f53\u793a\u610f\u56fe<\/figcaption><\/figure>\n<h3><span class=\"ez-toc-section\" id=\"11_Response-based_Knowledge_Distillation%EF%BC%88%E5%9F%BA%E4%BA%8E%E5%93%8D%E5%BA%94%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>1.1 Response-based Knowledge Distillation\uff08\u57fa\u4e8e\u54cd\u5e94\u7684\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"HL2pzisv\">\u57fa\u4e8e\u54cd\u5e94\u7684KD\u4fa7\u91cd\u4e8e\u4ece\u6700\u540e\u4e00\u5c42\u5b66\u4e60\u77e5\u8bc6\u4f5c\u4e3a\u54cd\u5e94\u3002\u5b83\u65e8\u5728\u4f7f\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u7684\u6700\u7ec8\u9884\u6d4b\u4fdd\u6301\u4e00\u81f4\u3002\u57fa\u4e8e\u54cd\u5e94\u7684KD\u7684\u7279\u6027\u662f\u7ed3\u679c\u9a71\u52a8\u5b66\u4e60\uff0c\u4f7f\u5176\u6613\u4e8e\u6269\u5c55\u5230\u5404\u79cd\u4efb\u52a1\u3002\u5f00\u521b\u6027\u7684KD\u53ef\u4ee5\u8ffd\u6eaf\u5230Hinton\u7b49\u4eba\u5de5\u4f5c[1]\u3002\u6838\u5fc3\u601d\u60f3\u662f\u901a\u8fc7\u8f6f\u5316\u7684softmax\uff08\u5373\u201c\u8f6f\u6807\u7b7e\u201d\uff09\u63d0\u53d6\u7c7b\u6982\u7387\u5206\u5e03\u3002\u5bf9\u4e8e\u5206\u7c7b\u4efb\u52a1\uff0c\u8f6f\u6982\u7387\u5206\u5e03\u88ab\u516c\u5f0f\u5316\u4e3a\uff1a<\/p>\n<p data-pid=\"MJphdWXL\"><span data-eeimg=\"1\" data-tex=\"p(z_{i};T)=frac{exp{(z_{i}\/T)}}{sum_{j=1}^{N}exp{(z_{j}\/T)}},\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mi&gt;p&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mo&gt;;&lt;\/mo&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;=&lt;\/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;exp&lt;\/mi&gt;&lt;mo&gt;\u2061&lt;\/mo&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mo&gt;\/&lt;\/mo&gt;&lt;\/mrow&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/mrow&gt;&lt;\/mrow&gt;&lt;mrow&gt;&lt;munderover&gt;&lt;mo&gt;\u2211&lt;\/mo&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;mo&gt;=&lt;\/mo&gt;&lt;mn&gt;1&lt;\/mn&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;N&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/munderover&gt;&lt;mi&gt;exp&lt;\/mi&gt;&lt;mo&gt;\u2061&lt;\/mo&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mo&gt;\/&lt;\/mo&gt;&lt;\/mrow&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/mrow&gt;&lt;\/mrow&gt;&lt;\/mfrac&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;\/math&gt;'>p(zi;T)=exp\u2061(zi\/T)\u2211j=1Nexp\u2061(zj\/T),<\/span><\/span><\/p>\n<p data-pid=\"nfs33TKO\">\u7ed9\u5b9a\u6765\u81ea\u6559\u5e08\u548c\u5b66\u751f\u7684\u7c7b\u522b\u6982\u7387\u5206\u5e03\u00a0<span data-eeimg=\"1\" data-tex=\"p(z^{T};T)\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mi&gt;p&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;;&lt;\/mo&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/math&gt;'>p(zT;T)<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"p(z^{S};T)\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mi&gt;p&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;;&lt;\/mo&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/math&gt;'>p(zS;T)<\/span><\/span> \uff0c\u57fa\u4e8e\u54cd\u5e94\u7684KD\u5c1d\u8bd5\u4f7f\u7528\u8ddd\u79bb\u51fd\u6570\u00a0<span data-eeimg=\"1\" data-tex=\"mathcal{L}_{dis}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>Ldis<\/span><\/span> \u5bf9\u9f50\u4e8c\u8005\u3002<\/p>\n<p data-pid=\"MAYYVpTe\"><span data-eeimg=\"1\" data-tex=\"mathcal{L}_{response_kd}(p(z^{S};T),p(z^{T};T))=mathcal{L}_{dis}(p(z^{S};T),p(z^{T};T)),\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;r&lt;\/mi&gt;&lt;mi&gt;e&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;mi&gt;p&lt;\/mi&gt;&lt;mi&gt;o&lt;\/mi&gt;&lt;mi&gt;n&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;mi&gt;e&lt;\/mi&gt;&lt;mi mathvariant=\"normal\"&gt;_&lt;\/mi&gt;&lt;mi&gt;k&lt;\/mi&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;mi&gt;p&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;;&lt;\/mo&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;mi&gt;p&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;;&lt;\/mo&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;=&lt;\/mo&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;mi&gt;p&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;;&lt;\/mo&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;mi&gt;p&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;z&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;;&lt;\/mo&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;\/math&gt;'>Lresponse_kd(p(zS;T),p(zT;T))=Ldis(p(zS;T),p(zT;T)),<\/span><\/span><\/p>\n<p data-pid=\"N9DClt42\">\u5176\u4e2d\u00a0<span data-eeimg=\"1\" data-tex=\"mathcal{L}_{dis}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>Ldis<\/span><\/span> \u53ef\u4ee5\u662fKullback-Leibler\uff08KL\uff09\u6563\u5ea6\u3001\u5747\u5e73\u65b9\u8bef\u5dee\u7b49\u3002<\/p>\n<p data-pid=\"iRxPxKFK\">\u4e3b\u8981\u8bb2\u89e3\u4e09\u4e2a\u65b9\u9762\u7684\u5de5\u4f5c\u5185\u5bb9\uff1a<\/p>\n<p data-pid=\"Yut7b8N1\">1.1.1 Interpret response-based KD\uff08\u89e3\u91ca\u57fa\u4e8e\u54cd\u5e94\u7684\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"bMs3G8ML\">\u57fa\u4e8e\u54cd\u5e94\u7684KD\u7684\u6838\u5fc3\u601d\u60f3\u5f88\u5bb9\u6613\u7406\u89e3\u3002\u5b83\u5f15\u5bfc\u5b66\u751f\u5b66\u4e60\u6559\u5e08\u7f51\u7edc\u751f\u6210\u7684\u6700\u7ec8\u7ed3\u679c\u3002\u57fa\u4e8e\u54cd\u5e94\u7684KD\u7684\u6709\u6548\u6027\u4e5f\u53ef\u4ee5\u4e0e\u6807\u7b7e\u5e73\u6ed1\u6b63\u5219\u5316\u8054\u7cfb\u8d77\u6765\u3002Yuan\u7b49\u4eba\u5c06\u57fa\u4e8eKL\u6563\u5ea6\u7684KD\u65b9\u7a0b\u6c42\u89e3\u4e3a\u6807\u7b7e\u5e73\u6ed1\u3002Muller\u7b49\u4eba\u89c2\u5bdf\u5230\u6807\u7b7e\u5e73\u6ed1\u6291\u5236\u4e86KD\u7684\u6709\u6548\u6027\u3002\u4ed6\u4eec\u8ba4\u4e3a\uff0c\u6807\u7b7e\u5e73\u6ed1\u4f1a\u5bfc\u81f4\u4e0d\u540c\u7c7b\u522b\u6837\u672c\u4e4b\u95f4\u76f8\u4f3c\u6027\u7684\u903b\u8f91\u4fe1\u606f\u4e22\u5931\uff0c\u8fd9\u5bf9\u57fa\u4e8e\u54cd\u5e94\u7684KD\u5f88\u6709\u4ef7\u503c\u3002\u7136\u800c\uff0cShen\u7b49\u4eba\u63d0\u4f9b\u4e86\u4e00\u9879\u5b9e\u8bc1\u7814\u7a76\uff0c\u8bf4\u660e\u6807\u7b7e\u5e73\u6ed1\u901a\u5e38\u4e0d\u4f1a\u6291\u5236KD\u7684\u6709\u6548\u6027\u3002\u4ed6\u4eec\u53d1\u73b0\uff0c\u5728\u4e24\u79cd\u60c5\u51b5\u4e0b\uff0c\u6807\u7b7e\u5e73\u6ed1\u53ef\u80fd\u4f1a\u4ea7\u751f\u8d1f\u9762\u5f71\u54cd\uff1a\u957f\u5c3e\u7c7b\u5206\u5e03\u548c\u7c7b\u6570\u91cf\u589e\u52a0\u3002Mobahi\u7b49\u4eba\u5c55\u793a\u4e86\u4e00\u79cd\u7406\u8bba\u89c2\u70b9\uff0c\u5373\u4e24\u4e2a\u76f8\u540c\u7f51\u7edc\u67b6\u6784\u4e4b\u95f4\u7684\u84b8\u998f\u653e\u5927\u4e86Hilbert\u7a7a\u95f4\u4e2d\u7684\u6b63\u5219\u5316\u3002\u6700\u65b0\u7684DKD\u5c06\u539f\u59cbKD\u635f\u5931\u89e3\u8026\u4e3a\u76ee\u6807\u7c7bKD\u548c\u975e\u76ee\u6807\u7c7bKD\u3002\u901a\u8fc7\u4ec5\u5f15\u5165\u4e24\u4e2a\u8d85\u53c2\u6570\u6765\u7075\u6d3b\u5e73\u8861\u4e24\u4e2a\u9879\uff0cDKD\u63d0\u9ad8\u4e86\u539f\u59cbKD\u7684\u6709\u6548\u6027\u3002<\/p>\n<p data-pid=\"3RfdFMX6\">1.1.2 Reduce performance gap with auxiliary architecture \uff08\u901a\u8fc7\u8f85\u52a9\u7ed3\u6784\u51cf\u5c11\u6559\u5e08-\u5b66\u751f\u6027\u80fd\u5dee\u8ddd\uff09<\/p>\n<p data-pid=\"H-CUoeXd\">\u7531\u4e8e\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u7684\u80fd\u529b\u5dee\u8ddd\uff0cKD\u53ef\u80fd\u5b58\u5728\u6027\u80fd\u5dee\u8ddd\uff0c\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\u95ee\u9898\u3002\u4e3a\u4e86\u7f13\u89e3\u8fd9\u4e2a\u95ee\u9898\uff0cTAKD\u5f15\u5165\u4e86\u4e00\u4e2a\u4e2d\u7b49\u89c4\u6a21\u7684\u7f51\u7edc\u4f5c\u4e3a\u6559\u5e08\u52a9\u7406\uff0c\u5e76\u6267\u884c\u4e86\u4e00\u4e2a\u8fde\u7eed\u7684KD\u8fc7\u7a0b\u3002\u6cbf\u7740\u8fd9\u6761\u8109\u7edc\uff0cHKD\u5e94\u7528\u4e86\u4e00\u4e2a\u8f85\u52a9\u6559\u5e08\u7f51\u7edc\u6765\u4f20\u8f93\u5206\u5c42\u4fe1\u606f\u6d41\u3002DGKD\u63d0\u51fa\u4e86\u4e00\u79cd\u5bc6\u96c6\u5f15\u5bfc\u6a21\u5f0f\uff0c\u5e76\u7531\u6240\u6709\u524d\u4efb\u6559\u5e08\u6267\u884c\u591a\u6b65KD\u3002SFTN\u9996\u5148\u8bad\u7ec3\u4e00\u540d\u6559\u5e08\u548c\u5b66\u751f\u5206\u652f\uff0c\u7136\u540e\u5c06\u66f4\u5bb9\u6613\u4f20\u6388\u7684\u77e5\u8bc6\u4f20\u6388\u7ed9\u5b66\u751f\u3002TOFD\u548cHSAKD\u5728\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u9644\u52a0\u4e86\u51e0\u4e2a\u8f85\u52a9\u5206\u652f\uff0c\u4ee5\u4fc3\u8fdb\u77e5\u8bc6\u4e92\u52a8\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u4e3a\u8bad\u7ec3\u56fe\u5f15\u5165\u4e86\u989d\u5916\u7684\u67b6\u6784\uff0c\u5e76\u589e\u52a0\u4e86\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n<p data-pid=\"pOp5x0WV\">1.1.3 Reduce performance gap with adaptive distillation \uff08\u901a\u8fc7\u81ea\u9002\u5e94\u84b8\u998f\u51cf\u5c11\u6559\u5e08-\u5b66\u751f\u6027\u80fd\u5dee\u8ddd\uff09<\/p>\n<p data-pid=\"ZvBraJiF\">\u4e00\u4e9b\u5de5\u4f5c\u8bd5\u56fe\u7814\u7a76\u6837\u672c\u81ea\u9002\u5e94\u84b8\u998f\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002WSL\u4ece\u504f\u5dee-\u65b9\u5dee\u6743\u8861\u7684\u89d2\u5ea6\u63d0\u51fa\u4e86\u6837\u672c\u52a0\u6743\u8f6f\u6807\u7b7e\u3002ATKD\u4f7f\u7528\u57fa\u4e8e\u6807\u51c6\u504f\u5dee\u7684\u81ea\u9002\u5e94\u6e29\u5ea6\u6765\u7f29\u5c0f\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u7684\u9510\u5ea6\u5dee\u8ddd\u3002MKD\u5229\u7528\u5143\u5b66\u4e60\u6765\u641c\u7d22\u53ef\u5b66\u4e60\u7684\u6e29\u5ea6\u53c2\u6570\u3002SCKD\u4ece\u68af\u5ea6\u76f8\u4f3c\u6027\u7684\u89d2\u5ea6\u7814\u7a76\u4e86\u4e0d\u5339\u914d\u95ee\u9898\uff0c\u4f7f\u5b66\u751f\u81ea\u9002\u5e94\u5730\u5b66\u4e60\u5176\u6709\u76ca\u7684\u77e5\u8bc6\u3002PAD\u63d0\u51fa\u4e86\u4e00\u79cd\u5177\u6709\u6570\u636e\u4e0d\u786e\u5b9a\u6027\u7684\u81ea\u9002\u5e94\u6837\u672c\u52a0\u6743\u673a\u5236\uff0c\u8be5\u673a\u5236\u57fa\u4e8e\u56f0\u96be\u5b9e\u4f8b\u53ef\u80fd\u5bf9KD\u96be\u4ee5\u5904\u7406\u7684\u8fd9\u4e00\u73b0\u8c61\u3002\u9664\u4e86\u63a2\u7d22\u4f9d\u8d56\u4e8e\u6837\u672c\u7684\u84b8\u998f\u5916\uff0cSong\u7b49\u4eba\u548cLi\u7b49\u4eba\u8fd8\u63d0\u51fa\u4e86\u4e00\u79cd\u6df7\u5408\u524d\u5411\u4f20\u64ad\u65b9\u6848\uff0c\u4f7f\u5b66\u751f\u901a\u8fc7\u8054\u5408\u8bad\u7ec3\u9690\u5f0f\u5b66\u4e60\u6559\u5e08\u7684\u77e5\u8bc6\u3002\u5728\u8bfe\u7a0b\u5b66\u4e60\u7684\u9a71\u4f7f\u4e0b\uff0cRCO\u5f15\u5bfc\u5b66\u751f\u4ece\u5934\u5f00\u59cb\u6a21\u4eff\u8001\u5e08\u7684\u8bad\u7ec3\u8f68\u8ff9\uff0c\u76f4\u81f3\u6536\u655b\u3002ESKD\u63d0\u524d\u505c\u6b62\u6559\u5e08\u8bad\u7ec3\uff0c\u4ee5\u4ea7\u751f\u66f4\u67d4\u548c\u7684\u903b\u8f91\u5206\u5e03\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c0827ec76ce5fb5cdc76e852973cf23b.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c0827ec76ce5fb5cdc76e852973cf23b.jpg\"><\/a><figcaption>\u88681 \u57fa\u4e8e\u54cd\u5e94\u7684KD\u65b9\u6cd5\u5b9e\u9a8c\u7ed3\u679c\u5bf9\u6bd4<\/figcaption><\/figure>\n<p data-pid=\"QQda496F\"><strong>\u603b\u7ed3\u3002<\/strong>\u6211\u4eec\u5728\u88681\u4e2d\u5bf9\u5404\u79cd\u57fa\u4e8e\u54cd\u5e94\u7684KD\u65b9\u6cd5\u8fdb\u884c\u4e86\u5b9e\u9a8c\u7ed3\u679c\u5bf9\u6bd4\u3002\u5927\u591a\u6570\u65b9\u6cd5\u90fd\u5e94\u7528\u4e8e\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff0c\u800c\u6700\u8fd1\u7684MKD\u8fdb\u4e00\u6b65\u65e8\u5728\u6539\u8fdb\u89c6\u89c9Transformer\u3002\u57fa\u4e8e\u54cd\u5e94\u7684KD\u7684\u6838\u5fc3\u601d\u60f3\u662f\u7ed3\u679c\u9a71\u52a8\u5b66\u4e60\uff0c\u5e76\u4e14\u5f88\u5bb9\u6613\u5e94\u7528\u4e8e\u73b0\u6709\u7684\u6df1\u5ea6\u5b66\u4e60\u4efb\u52a1\u3002\u5bf9\u4e8e\u76ee\u6807\u68c0\u6d4b\uff0cChen\u7b49\u4eba\u5efa\u8bae\u5f15\u5bfc\u5b66\u751f\u6a21\u4eff\u8001\u5e08\u7684\u76ee\u6807\u6982\u7387\u548c\u56de\u5f52\u8fb9\u754c\u6846\u3002\u5bf9\u4e8e\u8bed\u4e49\u5206\u5272\uff0c\u7ed3\u679c\u9a71\u52a8\u7684\u77e5\u8bc6\u662f\u9010\u50cf\u7d20\u7684\u7c7b\u6982\u7387\u5206\u5e03\u3002\u7c7b\u4f3c\u5730\uff0c\u5bf9\u4e8e\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\u7684Bert\u538b\u7f29\uff0cDistilBERT\u4f20\u9012\u63a9\u7801\u8bcd\u6c47\u7684\u7c7b\u9884\u6d4b\u3002\u5c3d\u7ba1\u57fa\u4e8e\u54cd\u5e94\u7684KD\u5df2\u6210\u529f\u5e94\u7528\u4e8e\u8bb8\u591a\u4efb\u52a1\uff0c\u4f46\u6027\u80fd\u5dee\u8ddd\u4ecd\u7136\u662f\u4e00\u4e2a\u503c\u5f97\u63a2\u7d22\u7684\u95ee\u9898\u3002\u5f53\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u8ddd\u5f88\u5927\u65f6\uff0c\u5b66\u751f\u53ef\u80fd\u65e0\u6cd5\u5438\u6536\u6709\u610f\u4e49\u7684\u77e5\u8bc6\uff0c\u8fd9\u53ef\u80fd\u4f1a\u5bfc\u81f4\u4e0d\u5229\u7684\u76d1\u7ba1\u6548\u679c\u3002\u6b64\u5916\uff0c\u57fa\u4e8e\u54cd\u5e94\u7684KD\u5ffd\u7565\u4e86\u795e\u7ecf\u7f51\u7edc\u9690\u85cf\u5c42\u4e2d\u7f16\u7801\u7684\u4e2d\u95f4\u4fe1\u606f\uff0c\u5bfc\u81f4\u6027\u80fd\u63d0\u5347\u6709\u9650\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"12_Feature-based_Knowledge_Distillation%EF%BC%88%E5%9F%BA%E4%BA%8E%E7%89%B9%E5%BE%81%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>1.2 Feature-based Knowledge Distillation\uff08\u57fa\u4e8e\u7279\u5f81\u7684\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"xNjGnpsv\">\u6b63\u5982\u6211\u4eec\u4e0a\u9762\u8ba8\u8bba\u7684\uff0c\u57fa\u4e8e\u54cd\u5e94\u7684KD\u5ffd\u7565\u4e86\u4e2d\u5c42\u76d1\u7763\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u7f3a\u9677\uff0c\u57fa\u4e8e\u7279\u5f81\u7684KD\u4fa7\u91cd\u4e8e\u63a2\u7d22\u4e2d\u95f4\u7279\u5f81\u4fe1\u606f\uff0c\u4ee5\u63d0\u4f9b\u5168\u9762\u7684\u76d1\u7763\uff0c\u5982\u7279\u5f81\u56fe\u53ca\u5176\u63d0\u53d6\u7684\u4fe1\u606f\u3002\u57fa\u4e8e\u7279\u5f81\u7684\u84b8\u998f\u635f\u5931\u53ef\u4ee5\u516c\u5f0f\u5316\u4e3a\uff1a<\/p>\n<p data-pid=\"2qCpCU-E\"><span data-eeimg=\"1\" data-tex=\"mathcal{L}_{feature_kd}(F^{S},F^{T})=mathcal{L}_{dis}{(phi^{S}(F^{S}),phi^{T}(F^{T}))},\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mi&gt;e&lt;\/mi&gt;&lt;mi&gt;a&lt;\/mi&gt;&lt;mi&gt;t&lt;\/mi&gt;&lt;mi&gt;u&lt;\/mi&gt;&lt;mi&gt;r&lt;\/mi&gt;&lt;mi&gt;e&lt;\/mi&gt;&lt;mi mathvariant=\"normal\"&gt;_&lt;\/mi&gt;&lt;mi&gt;k&lt;\/mi&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;=&lt;\/mo&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;\u03d5&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;\u03d5&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/mrow&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;\/math&gt;'>Lfeature_kd(FS,FT)=Ldis(\u03d5S(FS),\u03d5T(FT)),<\/span><\/span><\/p>\n<p data-pid=\"trkys0w0\">\u5176\u4e2d\u00a0<span data-eeimg=\"1\" data-tex=\"F^{S}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>FS<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"F^{T}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>FT<\/span><\/span> \u8868\u793a\u6765\u81ea\u5b66\u751f\u548c\u6559\u5e08\u7684\u4e2d\u95f4\u7279\u5f81\u56fe\uff0c<span data-eeimg=\"1\" data-tex=\"phi^{S}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;\u03d5&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>\u03d5S<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"phi^{T}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;\u03d5&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>\u03d5T<\/span><\/span> \u662f\u4ea7\u751f\u84b8\u998f\u4fe1\u606f\u7684\u8f6c\u6362\u51fd\u6570\uff0c\u5982\u6ce8\u610f\u529b\u673a\u5236\u3001\u6fc0\u6d3b\u8fb9\u754c\u3001\u795e\u7ecf\u5143\u9009\u62e9\u6027\u548c\u6982\u7387\u5206\u5e03\u7b49\uff0c<span data-eeimg=\"1\" data-tex=\"mathcal{L}_{dis}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>Ldis<\/span><\/span> \u662f\u4e00\u4e2a\u8ddd\u79bb\u51fd\u6570\uff0c\u7528\u4e8e\u8861\u91cf\u5339\u914d\u7279\u5f81\u4fe1\u606f\u7684\u76f8\u4f3c\u6027\uff0c\u4f8b\u5982\uff0c\u5747\u65b9\u8bef\u5dee\u635f\u5931\u548cKL\u6563\u5ea6\u635f\u5931\u3002<\/p>\n<p data-pid=\"QI1x2NWr\">\u4e3b\u8981\u8bb2\u89e3\u4e09\u4e2a\u65b9\u9762\u7684\u5de5\u4f5c\u5185\u5bb9\uff1a<\/p>\n<p data-pid=\"Fi92EJ-c\">1.2.1 Knowledge exploration: transform intermediate feature maps to meaningful knowledge \uff08\u77e5\u8bc6\u63a2\u7d22\uff1a\u8f6c\u6362\u4e2d\u95f4\u5c42\u7279\u5f81\u56fe\u5f97\u5230\u6709\u610f\u4e49\u7684\u77e5\u8bc6\uff09<\/p>\n<p data-pid=\"uRUvOior\">\u5f00\u521b\u6027\u7684FitNet\u662f\u7b2c\u4e00\u79cd\u57fa\u4e8e\u7279\u5f81\u7684KD\u65b9\u6cd5\uff0c\u5176\u6838\u5fc3\u601d\u60f3\u662f\u5728\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u4ee5\u9010\u5c42\u7684\u65b9\u5f0f\u5bf9\u9f50\u4ece\u9690\u85cf\u5c42\u751f\u6210\u7684\u4e2d\u5c42\u7279\u5f81\u56fe\u3002\u8fd9\u9879\u7b80\u5355\u76f4\u89c2\u7684\u5de5\u4f5c\u53ef\u80fd\u4e0d\u9700\u8981\u9ad8\u7ea7\u77e5\u8bc6\u3002\u540e\u7eed\u7684\u65b9\u6cd5\u8bd5\u56fe\u63a2\u7d22\u4ece\u539f\u59cb\u7279\u5f81\u56fe\u4e2d\u7f16\u7801\u66f4\u6709\u610f\u4e49\u7684\u4fe1\u606f\u3002AT\u5c06\u7279\u5f81\u56fe\u8f6c\u6362\u4e3a\u7a7a\u95f4\u6ce8\u610f\u529b\u56fe\uff0c\u4f5c\u4e3a\u6709\u4ef7\u503c\u7684\u4fe1\u606f\u3002NST\u63d0\u53d6\u6fc0\u6d3b\u70ed\u56fe\u4f5c\u4e3a\u795e\u7ecf\u5143\u8f6c\u79fb\u7684\u9009\u62e9\u6027\u3002Srinivas\u7b49\u4eba\u5728\u7279\u5f81\u56fe\u4e4b\u95f4\u5e94\u7528\u4e86\u96c5\u53ef\u6bd4\u5339\u914d\u3002PKT\u5c06\u7279\u5f81\u56fe\u8868\u793a\u4e3a\u6982\u7387\u5206\u5e03\uff0c\u5e76\u7531KL\u6563\u5ea6\u6a21\u62df\u3002FSP\u5f15\u5165\u4e86Gramian\u77e9\u9635\u6765\u6d4b\u91cf\u4e0d\u540c\u5c42\u7279\u5f81\u56fe\u4e4b\u95f4\u7684\u6c42\u89e3\u8fc7\u7a0b\u6d41\u3002Seung\u7b49\u4eba\u4f7f\u7528\u5947\u5f02\u503c\u5206\u89e3\u6765\u89e3\u6790\u7279\u5f81\u77e5\u8bc6\u3002FT\u5f15\u5165\u4e86\u81ea\u52a8\u7f16\u7801\u5668\uff0c\u4ee5\u65e0\u76d1\u7763\u7684\u65b9\u5f0f\u5c06\u6559\u5e08\u7684\u7279\u5f81\u56fe\u89e3\u6790\u4e3a\u201c\u56e0\u7d20\u201d\uff0c\u5e76\u5f15\u5165\u4e86\u7ffb\u8bd1\u5668\uff0c\u5c06\u201c\u56e0\u7d20\u201d\u8f6c\u6362\u4e3a\u6613\u4e8e\u7406\u89e3\u7684\u77e5\u8bc6\u3002AB\u8003\u8651\u9690\u85cf\u7279\u5f81\u7a7a\u95f4\u4e2d\u7684\u6fc0\u6d3b\u8fb9\u754c\uff0c\u5e76\u8feb\u4f7f\u5b66\u751f\u5b66\u4e60\u4e0e\u8001\u5e08\u4e00\u81f4\u7684\u8fb9\u754c\u3002Overhaul\u91cd\u65b0\u601d\u8003\u4e86\u84b8\u998f\u7279\u5f81\u4f4d\u7f6e\uff0c\u4f7f\u7528\u65b0\u8bbe\u8ba1\u7684 ReLU\u548cL2\u8ddd\u79bb\u51fd\u6570\u6765\u8fc7\u6ee4\u5197\u4f59\u4fe1\u606f\u3002<\/p>\n<p data-pid=\"pN6YPHFp\">\u6700\u8fd1\uff0cTOFD\u548cHSAKD\u5c06\u8f85\u52a9\u5206\u7c7b\u5668\u9644\u52a0\u5230\u7531\u989d\u5916\u4efb\u52a1\u76d1\u7763\u7684\u4e2d\u5c42\u7279\u5f81\u56fe\u4e0a\uff0c\u4ee5\u4ea7\u751f\u4fe1\u606f\u4e30\u5bcc\u7684\u6982\u7387\u5206\u5e03\u3002\u524d\u8005\u5229\u7528\u4e86\u539f\u59cb\u7684\u76d1\u7763\u4efb\u52a1\uff0c\u800c\u540e\u8005\u5f15\u5165\u4e86\u6709\u610f\u4e49\u7684\u81ea\u76d1\u7763\u589e\u5f3a\u4efb\u52a1\u3002MGD\u4f7f\u7528\u81ea\u9002\u5e94\u901a\u9053\u5206\u914d\u7b97\u6cd5\u6267\u884c\u7ec6\u7c92\u5ea6\u7279\u5f81\u63d0\u53d6\u3002ICKD\u4ece\u5305\u542b\u7279\u5f81\u7a7a\u95f4\u591a\u6837\u6027\u548c\u540c\u6e90\u6027\u7684\u7279\u5f81\u56fe\u4e2d\u6316\u6398\u901a\u9053\u95f4\u76f8\u5173\u6027\u3002\u5728\u901a\u9053\u7ef4\u5ea6\u4e4b\u5916\uff0cTTKD\u4f7f\u7528\u81ea\u6ce8\u610f\u529b\u673a\u5236\u8fdb\u884c\u7a7a\u95f4\u7ea7\u7279\u5f81\u5339\u914d\u3002\u603b\u4e4b\uff0c\u4ee5\u524d\u7684\u65b9\u6cd5\u901a\u5e38\u4f1a\u4e3aKD\u63d0\u53d6\u66f4\u4e30\u5bcc\u7684\u7279\u5f81\u4fe1\u606f\uff0c\u4ece\u800c\u83b7\u5f97\u6bd4\u7ecf\u5178FitNet\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n<p data-pid=\"OGwa6DIa\">1.2.2 Knowledge transfer: good mimicry algorithm to let the student learn better \uff08\u77e5\u8bc6\u8fc1\u79fb\uff1a\u597d\u7684\u6a21\u4eff\u7b97\u6cd5\u8ba9\u5b66\u751f\u5b66\u5f97\u66f4\u597d\uff09<\/p>\n<p data-pid=\"1nuNWnTh\">\u9664\u4e86\u77e5\u8bc6\u63a2\u7d22\uff0c\u53e6\u4e00\u4e2a\u6709\u4ef7\u503c\u7684\u95ee\u9898\u662f\u5982\u4f55\u6709\u6548\u5730\u8fc1\u79fb\u77e5\u8bc6\u3002\u5927\u591a\u6570\u57fa\u4e8e\u7279\u5f81\u7684KD\u65b9\u6cd5\u4f7f\u7528\u7b80\u5355\u7684\u5747\u65b9\u8bef\u5dee\u635f\u5931\u8fdb\u884c\u77e5\u8bc6\u5bf9\u9f50\u3002\u9664\u6b64\u4e4b\u5916\uff0cVID\u8fd8\u5f15\u7528\u4e86\u4fe1\u606f\u8bba\u6846\u67b6\uff0c\u8ba4\u4e3aKD\u662f\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u4e92\u4fe1\u606f\u7684\u6700\u5927\u5316\u3002Wang\u7b49\u4eba\u5c06\u5b66\u751f\u89c6\u4e3a\u751f\u6210\u5668\uff0c\u5e76\u5e94\u7528\u4e86\u4e00\u4e2a\u989d\u5916\u7684\u9274\u522b\u5668\u6765\u533a\u5206\u5b66\u751f\u6216\u8001\u5e08\u4ea7\u751f\u7684\u7279\u5f81\uff0c\u8fd9\u79cd\u5bf9\u6297\u8fc7\u7a0b\u5f15\u5bfc\u5b66\u751f\u5b66\u4e60\u4e0e\u6559\u5e08\u76f8\u4f3c\u7684\u7279\u5f81\u5206\u5e03\u3002Xu\u7b49\u4eba\u63d0\u51fa\u5bf9\u5012\u6570\u7b2c\u4e8c\u5c42\u7684\u7279\u5f81\u5d4c\u5165\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u4ee5\u6291\u5236\u566a\u58f0\u7684\u8d1f\u9762\u5f71\u54cd\u3002\u9664\u4e86\u7814\u7a76\u6a21\u4eff\u5ea6\u91cf\u635f\u5931\u5916\uff0c\u5728\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u4f7f\u7528\u5171\u4eab\u5206\u7c7b\u5668\u8fd8\u53ef\u4ee5\u5e2e\u52a9\u5b66\u751f\u9690\u5f0f\u5730\u5bf9\u9f50\u6559\u5e08\u7684\u7279\u5f81\u3002<\/p>\n<p data-pid=\"fui43jAy\">1.2.3 Distillation for vision transformer \uff08\u89c6\u89c9Transformer\u65b9\u9762\u7684\u84b8\u998f\u5de5\u4f5c\uff09<\/p>\n<p data-pid=\"00EHCzpD\">Vision Transformer\uff08ViT\uff09\u5728\u56fe\u50cf\u8bc6\u522b\u65b9\u9762\u8868\u73b0\u51fa\u4e86\u5353\u8d8a\u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u57fa\u4e8eViT\u7684\u7f51\u7edc\u9700\u8981\u5f88\u9ad8\u7684\u8ba1\u7b97\u6210\u672c\u3002KD\u4e3a\u8bad\u7ec3\u5177\u6709\u7406\u60f3\u6027\u80fd\u7684\u5c0f\u578bViT\u63d0\u4f9b\u4e86\u51fa\u8272\u7684\u89e3\u51b3\u65b9\u6848\u3002\u5728\u5173\u7cfb\u5c42\u6b21\u4e0a\uff0c\u6d41\u5f62\u84b8\u998f\u5c06\u9010\u5757\u5173\u7cfb\u4f5c\u4e3aViT\u84b8\u998f\u7684\u77e5\u8bc6\u7c7b\u578b\u8fdb\u884c\u63a2\u7d22\u3002\u5728\u7279\u5f81\u7ea7\u522b\u4e0a\uff0cAttnDistill\u5c06\u6ce8\u610f\u529b\u56fe\u4ece\u6559\u5e08\u8f6c\u79fb\u5230\u5b66\u751f\u3002ViTKD\u4e3aViT\u7279\u5f81\u84b8\u998f\u63d0\u4f9b\u4e86\u5b9e\u7528\u6307\u5357\u3002\u9664\u4e86\u540c\u8d28ViT\u4e4b\u95f4\u7684\u7279\u5f81\u6a21\u4eff\u5916\uff0c\u4e00\u4e9b\u5de5\u4f5c\u8fd8\u8bd5\u56fe\u4eceCNN\u4e2d\u63d0\u53d6\u5230\u7684\u5f52\u7eb3\u504f\u7f6e\u84b8\u998f\u5230ViT\u3002\u4e00\u4e9b\u6709\u524d\u666f\u7684\u77e5\u8bc6\u7c7b\u578b\u4ecd\u503c\u5f97\u8fdb\u4e00\u6b65\u6316\u6398\uff0c\u5982\u4e2d\u95f4\u7279\u5f81\u3001\u6ce8\u610f\u529b\u5173\u7cfb\u548c\u84b8\u998f\u4f4d\u7f6e\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-0a3c5025e473d08c317ff32bc90e6a89.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-0a3c5025e473d08c317ff32bc90e6a89.jpg\"><\/a><figcaption>\u88682 \u57fa\u4e8e\u7279\u5f81\u7684KD\u65b9\u6cd5\u5b9e\u9a8c\u7ed3\u679c\u5bf9\u6bd4<\/figcaption><\/figure>\n<p data-pid=\"QacnxMOJ\"><strong>\u603b\u7ed3\u3002<\/strong>\u6211\u4eec\u5728\u88682\u4e2d\u5bf9\u5404\u79cd\u57fa\u4e8e\u7279\u5f81\u7684KD\u65b9\u6cd5\u8fdb\u884c\u4e86\u5b9e\u9a8c\u7ed3\u679c\u5bf9\u6bd4\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u57fa\u4e8e\u7279\u5f81\u7684KD\u662f\u5bf9\u57fa\u4e8e\u54cd\u5e94\u7684KD\u7684\u5168\u9762\u8865\u5145\uff0c\u5b83\u63d0\u4f9b\u4e86\u5c01\u88c5\u5b66\u4e60\u8fc7\u7a0b\u7684\u4e2d\u95f4\u7279\u5f81\u3002\u7136\u800c\uff0c\u7b80\u5355\u5730\u5728\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u5bf9\u9f50\u76f8\u540c\u7684\u9636\u6bb5\u6027\u7279\u5f81\u4fe1\u606f\u53ef\u80fd\u4f1a\u5bfc\u81f4\u8d1f\u9762\u76d1\u7763\uff0c\u7279\u522b\u662f\u5728\u6027\u80fd\u5dee\u8ddd\u6216\u67b6\u6784\u5dee\u5f02\u8f83\u5927\u7684\u60c5\u51b5\u4e0b\u3002\u4e00\u4e2a\u66f4\u6709\u4ef7\u503c\u7684\u65b9\u5411\u53ef\u80fd\u662f\u57fa\u4e8e\u5b66\u751f\u53cb\u597d\u7684\u7279\u5f81KD\uff0c\u5b83\u63d0\u4f9b\u4e86\u8bed\u4e49\u4e00\u81f4\u7684\u76d1\u7763\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"13_Relation-based_Knowledge_Distillation%EF%BC%88%E5%9F%BA%E4%BA%8E%E5%85%B3%E7%B3%BB%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>1.3 Relation-based Knowledge Distillation\uff08\u57fa\u4e8e\u5173\u7cfb\u7684\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"9bgJcfNG\">\u57fa\u4e8e\u54cd\u5e94\u548c\u57fa\u4e8e\u7279\u5f81\u7684KD\u901a\u5e38\u8003\u8651\u4ece\u5355\u4e2a\u6837\u672c\u4e2d\u63d0\u53d6\u77e5\u8bc6\u3002\u76f8\u53cd\uff0c\u57fa\u4e8e\u5173\u7cfb\u7684KD\u5c06\u8de8\u6837\u672c\u6216\u8de8\u5c42\u5173\u7cfb\u89c6\u4e3a\u6709\u610f\u4e49\u7684\u77e5\u8bc6\u3002<\/p>\n<p data-pid=\"IlKCmKSU\">\u4e3b\u8981\u8bb2\u89e3\u4e24\u4e2a\u65b9\u9762\u7684\u5de5\u4f5c\u5185\u5bb9\uff1a<\/p>\n<p data-pid=\"pPGteC5X\">1.3.1 Relation-based Cross-Sample Knowledge Distillation \uff08\u8de8\u6837\u672c\u7684\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"nyZX5kNH\">\u57fa\u4e8e\u5173\u7cfb\u7684\u8de8\u6837\u672c\u84b8\u998f\u635f\u5931\u516c\u5f0f\u4e00\u822c\u4e3a\uff1a<\/p>\n<p data-pid=\"4x1k7EyH\"><span data-eeimg=\"1\" data-tex=\"mathcal{L}_{relation_kd}(F^{S},F^{T})=sum_{i,j}mathcal{L}_{dis}{(psi^{S}(v^{S}_{i},v^{S}_{j}),psi^{T}(v^{T}_{i},v^{T}_{j}))},\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;r&lt;\/mi&gt;&lt;mi&gt;e&lt;\/mi&gt;&lt;mi&gt;l&lt;\/mi&gt;&lt;mi&gt;a&lt;\/mi&gt;&lt;mi&gt;t&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;o&lt;\/mi&gt;&lt;mi&gt;n&lt;\/mi&gt;&lt;mi mathvariant=\"normal\"&gt;_&lt;\/mi&gt;&lt;mi&gt;k&lt;\/mi&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;=&lt;\/mo&gt;&lt;munder&gt;&lt;mo&gt;\u2211&lt;\/mo&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/munder&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;\u03c8&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;\u03c8&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/mrow&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;\/math&gt;'>Lrelation_kd(FS,FT)=\u2211i,jLdis(\u03c8S(viS,vjS),\u03c8T(viT,vjT)),<\/span><\/span><\/p>\n<p data-pid=\"OI4h9GXg\">\u5176\u4e2d\u00a0<span data-eeimg=\"1\" data-tex=\"F^{S}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>FS<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"F^{T}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>FT<\/span><\/span> \u5206\u522b\u8868\u793a\u6559\u5e08\u548c\u5b66\u751f\u7684\u7279\u5f81\u96c6\uff0c\u00a0<span data-eeimg=\"1\" data-tex=\"v_{i}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>vi<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"v_{j}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>vj<\/span><\/span> \u662f\u7b2c\u00a0<span data-eeimg=\"1\" data-tex=\"i\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/math&gt;'>i<\/span><\/span> \u4e2a\u548c\u7b2c\u00a0<span data-eeimg=\"1\" data-tex=\"j\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/math&gt;'>j<\/span><\/span> \u4e2a\u6837\u672c\u7684\u7279\u5f81\u5d4c\u5165\uff0c\u00a0<span data-eeimg=\"1\" data-tex=\"(v^{S}_{i},v^{S}_{j})in F^{S}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;\u2208&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>(viS,vjS)\u2208FS<\/span><\/span> \uff0c\u00a0<span data-eeimg=\"1\" data-tex=\"(v^{T}_{i},v^{T}_{j})in F^{T}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;\u2208&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;F&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>(viT,vjT)\u2208FT<\/span><\/span> \u3002\u00a0<span data-eeimg=\"1\" data-tex=\"psi^{S}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;\u03c8&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>\u03c8S<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"psi^{T}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;\u03c8&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>\u03c8T<\/span><\/span> \u662f\u00a0<span data-eeimg=\"1\" data-tex=\"(v^{S}_{i},v^{S}_{j})\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/math&gt;'>(viS,vjS)<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"(v^{T}_{i},v^{T}_{j})\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/math&gt;'>(viT,vjT)<\/span><\/span> \u7684\u76f8\u4f3c\u5ea6\u5ea6\u91cf\u51fd\u6570\u3002\u00a0<span data-eeimg=\"1\" data-tex=\"mathcal{L}_{dis}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>Ldis<\/span><\/span> \u662f\u4e00\u4e2a\u8ddd\u79bb\u51fd\u6570\uff0c\u7528\u4e8e\u8861\u91cf\u5b9e\u4f8b\u56fe\u7684\u76f8\u4f3c\u6027\uff0c\u4f8b\u5982\uff0c\u5747\u65b9\u8bef\u5dee\u635f\u5931\u548cKullback-Leibler\u6563\u5ea6\u635f\u5931\u3002<\/p>\n<p data-pid=\"Z1lYs5Rm\"><em>1.3.1.1 Constructing relational graph with various edge weights<\/em> \uff08\u5229\u7528\u4e0d\u540c\u7684\u8fb9\u6743\u91cd\u6784\u5efa\u5173\u7cfb\u56fe\uff09<\/p>\n<p data-pid=\"hyFvDkgA\">\u57fa\u4e8e\u5173\u7cfb\u7684KD\u77e5\u8bc6\u53ef\u4ee5\u770b\u4f5c\u662f\u4e00\u4e2a\u5b9e\u4f8b\u56fe\uff0c\u5176\u4e2d\u8282\u70b9\u8868\u793a\u6837\u672c\u7684\u7279\u5f81\u5d4c\u5165\u3002\u5927\u591a\u6570\u57fa\u4e8e\u5173\u7cfb\u7684KD\u7814\u7a76\u5404\u79cd\u76f8\u4f3c\u6027\u5ea6\u91cf\u51fd\u6570\u6765\u8ba1\u7b97\u8fb9\u6743\u91cd\u3002DarkRank\u662f\u7b2c\u4e00\u79cd\u57fa\u4e8e\u6df1\u5ea6\u5ea6\u91cf\u5b66\u4e60\u5d4c\u5165\u7684\u6b27\u51e0\u91cc\u5fb7\u8ddd\u79bb\u6765\u5efa\u6a21\u4ea4\u53c9\u6837\u672c\u76f8\u4f3c\u6027\u7684\u65b9\u6cd5\u3002MHGD\u4f7f\u7528\u591a\u5934\u6ce8\u610f\u529b\u7f51\u7edc\u5904\u7406\u57fa\u4e8e\u56fe\u7684\u8868\u793a\u3002RKD\u4f7f\u7528\u76f8\u4e92\u5173\u7cfb\u7684\u8ddd\u79bb\u548c\u89d2\u5ea6\u76f8\u4f3c\u6027\u4f5c\u4e3a\u7ed3\u6784\u5316\u77e5\u8bc6\u3002CCKD\u4f7f\u7528\u57fa\u4e8e\u6838\u7684\u9ad8\u65afRBF\u6355\u83b7\u5b9e\u4f8b\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u3002SP\u5728\u7ed9\u5b9a\u4e00\u4e2a\u6279\u6570\u636e\u7684\u60c5\u51b5\u4e0b\u6784\u5efa\u6210\u5bf9\u76f8\u4f3c\u6027\u77e9\u9635\u3002IRG\u901a\u8fc7\u9876\u70b9\u548c\u8fb9\u53d8\u6362\u5bf9\u5b9e\u4f8b\u5173\u7cfb\u56fe\u8fdb\u884c\u5efa\u6a21\u3002REFILLED\u5f15\u5bfc\u8001\u5e08\u91cd\u65b0\u52a0\u6743\u5bf9\u4e8e\u5b66\u751f\u56f0\u96be\u7684\u4e09\u5143\u7ec4\u6837\u672c\u6743\u91cd\uff0c\u4ee5\u8fdb\u884c\u5173\u7cfb\u5339\u914d\u3002\u6240\u6709\u65b9\u6cd5\u90fd\u4fa7\u91cd\u4e8e\u5728\u6837\u672c\u7ea7\u7279\u5f81\u5d4c\u5165\u4e0a\u5bf9\u5173\u7cfb\u56fe\u8fdb\u884c\u5efa\u6a21\uff0c\u4f46\u5728\u5404\u79cd\u8fb9\u6743\u91cd\u751f\u6210\u7b56\u7565\u4e0a\u6709\u6240\u4e0d\u540c\u3002<\/p>\n<p data-pid=\"bs9dKMux\"><em>1.3.1.2\u00a0<\/em>Constructing relational graph with meaningful transformation\uff08\u5229\u7528\u6709\u610f\u4e49\u7684\u8f6c\u6362\u6784\u5efa\u5173\u7cfb\u56fe\uff09<\/p>\n<p data-pid=\"U7T1X9qC\">\u4f7f\u7528\u7b80\u5355\u7684\u5ea6\u91cf\u51fd\u6570\u76f4\u63a5\u5bf9\u8fb9\u6743\u91cd\u8fdb\u884c\u5efa\u6a21\u53ef\u80fd\u65e0\u6cd5\u6709\u610f\u4e49\u5730\u6355\u6349\u76f8\u5173\u6027\u6216\u9ad8\u9636\u4f9d\u8d56\u6027\u3002CRD\u5f15\u5165\u4e86\u6709\u76d1\u7763\u5bf9\u6bd4\u5b66\u4e60\uff0c\u57fa\u4e8eInfoNCE\u7684\u8bef\u5dee\u5f62\u5f0f\uff0c\u4f7f\u5f97\u8001\u5e08\u548c\u5b66\u751f\u7684\u8868\u5f81\u5bf9\u9f50\u3002\u5728CRD\u4e0a\uff0cCRCD\u6839\u636e\u7279\u5f81\u53ca\u5176\u68af\u5ea6\u63d0\u51fa\u4e86\u4e92\u8865\u5173\u7cfb\u5bf9\u6bd4\u84b8\u998f\u3002\u4e3a\u4e86\u5728\u539f\u59cb\u76d1\u7763\u5b66\u4e60\u7684\u57fa\u7840\u4e0a\u63d0\u53d6\u66f4\u4e30\u5bcc\u7684\u77e5\u8bc6\uff0cSSKD\u9075\u5faaSimCLR\u6846\u67b6\uff0c\u5e76\u5229\u7528\u56fe\u50cf\u65cb\u8f6c\u7684\u81ea\u76d1\u7763\u5bf9\u6bd4\u84b8\u998f\u3002\u4e3a\u4e86\u5229\u7528\u6807\u7b7e\u4e2d\u7684\u7c7b\u522b\u4fe1\u606f\uff0cCSKD\u6784\u5efa\u4e86\u7c7b\u5185\u548c\u7c7b\u95f4\u7684\u7ed3\u6784\u5316\u5173\u7cfb\u3002\u4ee5\u524d\u7684\u65b9\u6cd5\u901a\u5e38\u5173\u6ce8\u5b9e\u4f8b\u7ea7\u7279\u5f81\u53ca\u5176\u5173\u7cfb\uff0c\u4f46\u5ffd\u7565\u4e86\u5c40\u90e8\u7279\u5f81\u548c\u7ec6\u8282\u3002\u56e0\u6b64\uff0cLKD\u5229\u7528\u7c7b\u611f\u77e5\u6ce8\u610f\u529b\u6a21\u5757\u6765\u6355\u83b7\u91cd\u8981\u533a\u57df\uff0c\u7136\u540e\u4f7f\u7528\u5c40\u90e8\u8865\u4e01\u5bf9\u5c40\u90e8\u5173\u7cfb\u77e9\u9635\u8fdb\u884c\u5efa\u6a21\u3002GLD\u6784\u9020\u4e86\u4e00\u4e2a\u5173\u7cfb\u56fe\uff0c\u5176\u4e2d\u5305\u542b\u7531\u5c40\u90e8\u7a7a\u95f4\u6c60\u5316\u5c42\u63d0\u53d6\u7684\u5c40\u90e8\u7279\u5f81\u3002<\/p>\n<p data-pid=\"8xboZ4OW\">1.3.2 Relation-based Cross-Layer Knowledge Distillation \uff08\u8de8\u5c42\u7684\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"y5nsLWh4\">\u9664\u4e86\u5728\u6570\u636e\u6837\u672c\u4e0a\u5efa\u7acb\u5173\u7cfb\u5916\uff0c\u6a21\u578b\u5185\u7684\u8de8\u5c42\u4ea4\u4e92\u4fe1\u606f\u7f16\u7801\u4e5f\u662f\u4e00\u79cd\u6709\u4ef7\u503c\u7684\u77e5\u8bc6\u5f62\u5f0f\u3002\u57fa\u4e8e\u5173\u7cfb\u7684\u8de8\u5c42\u84b8\u998f\u635f\u5931\u516c\u5f0f\u4e3a\uff1a<\/p>\n<p data-pid=\"vNp0FoDm\"><span data-eeimg=\"1\" data-tex=\"mathcal{L}_{relation_kd}(f^{S},f^{T})=mathcal{L}_{dis}{(g^{S}(f^{S}_{i},f^{S}_{j}),g^{T}(f^{T}_{i},f^{T}_{j}))},\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;r&lt;\/mi&gt;&lt;mi&gt;e&lt;\/mi&gt;&lt;mi&gt;l&lt;\/mi&gt;&lt;mi&gt;a&lt;\/mi&gt;&lt;mi&gt;t&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;o&lt;\/mi&gt;&lt;mi&gt;n&lt;\/mi&gt;&lt;mi mathvariant=\"normal\"&gt;_&lt;\/mi&gt;&lt;mi&gt;k&lt;\/mi&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;=&lt;\/mo&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;g&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;g&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/mrow&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;\/math&gt;'>Lrelation_kd(fS,fT)=Ldis(gS(fiS,fjS),gT(fiT,fjT)),<\/span><\/span><\/p>\n<p data-pid=\"VM86eIW3\">\u5176\u4e2d\u00a0<span data-eeimg=\"1\" data-tex=\"f^{S}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>fS<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"f^{T}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>fT<\/span><\/span> \u5206\u522b\u4ee3\u8868\u4ece\u6559\u5e08\u548c\u5b66\u751f\u4e0d\u540c\u5c42\u4e2d\u63d0\u53d6\u7684\u7279\u5f81\u96c6\u5408\uff0c\u00a0<span data-eeimg=\"1\" data-tex=\"f_{i}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>fi<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"f_{j}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>fj<\/span><\/span> \u662f\u6765\u81ea\u7b2c\u00a0<span data-eeimg=\"1\" data-tex=\"i\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/math&gt;'>i<\/span><\/span> \u5c42\u548c\u7b2c\u00a0<span data-eeimg=\"1\" data-tex=\"j\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/math&gt;'>j<\/span><\/span> \u5c42\u7684\u7279\u5f81\u5d4c\u5165\uff0c\u00a0<span data-eeimg=\"1\" data-tex=\"(f^{S}_{i},f^{S}_{j})in f^{S}, (f^{T}_{i},f^{T}_{j})in f^{T}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;\u2208&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;mo&gt;\u2208&lt;\/mo&gt;&lt;msup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>(fiS,fjS)\u2208fS,(fiT,fjT)\u2208fT<\/span><\/span> \u3002\u00a0<span data-eeimg=\"1\" data-tex=\"g^{S}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;g&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>gS<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"g^{T}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msup&gt;&lt;mi&gt;g&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msup&gt;&lt;\/math&gt;'>gT<\/span><\/span> \u662f\u00a0<span data-eeimg=\"1\" data-tex=\"(f^{S}_{i},f^{S}_{j})\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;S&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/math&gt;'>(fiS,fjS)<\/span><\/span> \u548c\u00a0<span data-eeimg=\"1\" data-tex=\"(f^{T}_{i},f^{T}_{j})\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;mo stretchy=\"false\"&gt;(&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo&gt;,&lt;\/mo&gt;&lt;msubsup&gt;&lt;mi&gt;f&lt;\/mi&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;j&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;T&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msubsup&gt;&lt;mo stretchy=\"false\"&gt;)&lt;\/mo&gt;&lt;\/math&gt;'>(fiT,fjT)<\/span><\/span> \u7684\u5c42\u805a\u5408\u51fd\u6570\u3002\u00a0<span data-eeimg=\"1\" data-tex=\"mathcal{L}_{dis}\"><span data-mathml='&lt;math xmlns=\"http:\/\/www.w3.org\/1998\/Math\/MathML\"&gt;&lt;msub&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi class=\"MJX-tex-caligraphic\" mathvariant=\"script\"&gt;L&lt;\/mi&gt;&lt;\/mrow&gt;&lt;mrow class=\"MJX-TeXAtom-ORD\"&gt;&lt;mi&gt;d&lt;\/mi&gt;&lt;mi&gt;i&lt;\/mi&gt;&lt;mi&gt;s&lt;\/mi&gt;&lt;\/mrow&gt;&lt;\/msub&gt;&lt;\/math&gt;'>Ldis<\/span><\/span> \u662f\u8ddd\u79bb\u51fd\u6570\u6765\u8861\u91cf\u8de8\u5c42\u7279\u5f81\u7684\u76f8\u4f3c\u5ea6\u3002<\/p>\n<p data-pid=\"v8ZX5Zbf\">FSP\u662f\u6355\u83b7KD\u8de8\u5c42\u7279\u5f81\u56fe\u4e4b\u95f4\u5173\u7cfb\u7684\u5f00\u521b\u6027\u65b9\u6cd5\uff0c\u5b83\u5f15\u5165\u4e86Gramian\u77e9\u9635\uff0c\u5c06\u6c42\u89e3\u8fc7\u7a0b\u7684\u6d41\u7a0b\u8868\u793a\u4e3a\u77e5\u8bc6\u3002Passalis\u7b49\u4eba\u6307\u51fa\uff0c\u5177\u6709\u4e0d\u540c\u6027\u80fd\u7684\u6559\u5e08\u548c\u5b66\u751f\u7f51\u7edc\u4e4b\u95f4\u7684\u540c\u4e00\u5c42\u53ef\u80fd\u4f1a\u51fa\u73b0\u8bed\u4e49\u62bd\u8c61\u5dee\u8ddd\u3002\u4ee5\u524d\u7684\u65b9\u6cd5\u901a\u5e38\u4f9d\u8d56\u4e8e\u624b\u5de5\u5236\u4f5c\u7684\u4e00\u5bf9\u4e00\u7684\u5c42\u5206\u914d\u7b56\u7565\uff0c\u7136\u800c\uff0c\u8fd9\u79cd\u7b80\u5355\u7684\u5bf9\u9f50\u53ef\u80fd\u4f1a\u5bfc\u81f4\u6210\u5bf9\u5e08\u751f\u5c42\u4e4b\u95f4\u7684\u8bed\u4e49\u4e0d\u5339\u914d\u95ee\u9898\u3002\u8bb8\u591a\u540e\u7eed\u5de5\u4f5c\u8003\u8651\u5728\u591a\u4e2a\u7279\u5f81\u5c42\u4e0a\u5efa\u6a21\u6709\u610f\u4e49\u7684\u77e5\u8bc6\u3002Jang\u7b49\u4eba\u5f15\u5165\u4e86\u7528\u4e8e\u52a0\u6743\u5c42\u7ea7\u7279\u5f81\u5339\u914d\u7684\u5143\u7f51\u7edc\u3002\u5728\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u542f\u53d1\u4e0b\uff0c\u4e00\u4e9b\u5de5\u4f5c\u5229\u7528\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u6743\u91cd\u8fdb\u884c\u81ea\u9002\u5e94\u5c42\u5206\u914d\u3002\u9664\u4e86\u7814\u7a76\u5c42\u5339\u914d\u95ee\u9898\u5916\uff0c\u4e00\u4e9b\u5de5\u4f5c\u8fd8\u8bd5\u56fe\u805a\u5408\u6240\u6709\u9636\u6bb5\u7279\u5f81\u56fe\uff0c\u4ee5\u6784\u5efa\u66f4\u5177\u4fe1\u606f\u91cf\u7684\u7279\u5f81\u4f5c\u4e3a\u76d1\u7763\u4fe1\u53f7\u3002ReviewKD\u5229\u7528\u6559\u5e08\u7684\u591a\u5c42\u6b21\u7279\u5f81\uff0c\u6839\u636e\u5404\u79cd\u7279\u5f81\u878d\u5408\u6a21\u5757\u6307\u5bfc\u5b66\u751f\u7684\u6bcf\u4e00\u5c42\u3002LONDON\u603b\u7ed3\u4e86\u591a\u5c42\u6b21\u7279\u5f81\u56fe\uff0c\u4ee5\u6a21\u62dfLipschitz\u8fde\u7eed\u6027\u3002\u9664\u4e86\u624b\u5de5\u7b56\u7565\u5916\uff0cDFA\u8fd8\u5e94\u7528\u4e86\u641c\u7d22\u65b9\u6cd5\u6765\u81ea\u52a8\u627e\u5230\u5408\u9002\u7684\u7279\u5f81\u805a\u5408\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-f1218616551ef1d37b6d8acded04fcad.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-f1218616551ef1d37b6d8acded04fcad.jpg\"><\/a><figcaption>\u88683 \u57fa\u4e8e\u5173\u7cfb\u7684KD\u65b9\u6cd5\u5b9e\u9a8c\u7ed3\u679c\u5bf9\u6bd4<\/figcaption><\/figure>\n<p data-pid=\"mLfnA-gS\"><strong>\u603b\u7ed3\u3002<\/strong>\u6211\u4eec\u5728\u88683\u4e2d\u5bf9\u5404\u79cd\u57fa\u4e8e\u5173\u7cfb\u7684KD\u65b9\u6cd5\u8fdb\u884c\u4e86\u5b9e\u9a8c\u7ed3\u679c\u5bf9\u6bd4\u3002\u4e0d\u540c\u4e8e\u57fa\u4e8e\u7279\u5f81\u548c\u57fa\u4e8e\u54cd\u5e94\u7684KD\uff0c\u57fa\u4e8e\u5173\u7cfb\u7684\u65b9\u6cd5\u65e8\u5728\u6355\u6349\u5404\u79cd\u6837\u672c\u6216\u4e0d\u540c\u5c42\u4e4b\u95f4\u7684\u9ad8\u9636\u5173\u7cfb\u3002\u5173\u7cfb\u56fe\u6355\u83b7\u4e86\u6574\u4e2a\u6570\u636e\u96c6\u7684\u7ed3\u6784\u5316\u4f9d\u8d56\u5173\u7cfb\u3002\u8de8\u5c42\u7279\u5f81\u5173\u7cfb\u5bf9\u8bed\u4e49\u8fc7\u7a0b\u7684\u4fe1\u606f\u8fdb\u884c\u7f16\u7801\u3002\u5982\u4f55\u4f7f\u7528\u66f4\u6709\u610f\u4e49\u7684\u8282\u70b9\u8f6c\u6362\u548c\u5ea6\u91cf\u51fd\u6570\u6216\u805a\u5408\u9002\u5f53\u7684\u5c42\u4fe1\u606f\u6765\u5efa\u6a21\u66f4\u597d\u7684\u5173\u7cfb\u4ecd\u7136\u662f\u9700\u8981\u8fdb\u4e00\u6b65\u7814\u7a76\u7684\u6838\u5fc3\u95ee\u9898\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"2_Distillation_Schemes_%EF%BC%88%E8%92%B8%E9%A6%8F%E6%9C%BA%E5%88%B6%EF%BC%89\"><\/span>2. Distillation Schemes \uff08\u84b8\u998f\u673a\u5236\uff09<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p data-pid=\"vfuV56Vp\">\u84b8\u998f\u673a\u5236\u901a\u5e38\u5206\u4e3a\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\u3001\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\u548c\u81ea\u77e5\u8bc6\u84b8\u998f\uff0c\u5982\u56fe2\u6240\u793a\uff1a<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-0fb8aa0d296b4281286b67fe0d8a8964.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-0fb8aa0d296b4281286b67fe0d8a8964.jpg\"><\/a><figcaption>\u56fe2 \u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\u3001\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\u548c\u81ea\u77e5\u8bc6\u84b8\u998f\u673a\u5236\u793a\u610f\u56fe<\/figcaption><\/figure>\n<h3><span class=\"ez-toc-section\" id=\"21_Offline_Knowledge_Distillation%EF%BC%88%E7%A6%BB%E7%BA%BF%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>2.1 Offline Knowledge Distillation\uff08\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"iDBMZBgj\">\u79bb\u7ebfKD\u662f\u6240\u8c13\u7684\u57fa\u4e8e\u6559\u5e08\u548c\u5b66\u751f\u7684\u5b66\u4e60\uff0c\u4e4b\u524d\u7684\u7814\u7a76\u5df2\u7ecf\u5bf9\u5176\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u7814\u7a76\u3002\u79bb\u7ebfKD\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u77e5\u8bc6\u4ece\u4e00\u4e2a\u9ad8\u6027\u80fd\u7684\u5927\u578b\u9884\u8bad\u7ec3\u6559\u5e08\u7f51\u7edc\u8f6c\u79fb\u5230\u4e00\u4e2a\u5c0f\u578b\u5feb\u901f\u7684\u5b66\u751f\u7f51\u7edc\u3002\u5728\u5b9e\u8df5\u4e2d\uff0c\u79bb\u7ebfKD\u901a\u5e38\u4f1a\u8fdb\u884c\u4e24\u9636\u6bb5\u7684\u8bad\u7ec3\uff1a\uff081\uff09\u6559\u5e08\u7f51\u7edc\u5728\u4efb\u52a1\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u4ee5\u5b9e\u73b0\u5353\u8d8a\u7684\u6027\u80fd\uff1b\u4ee5\u53ca\uff082\uff09\u5728\u8bad\u7ec3\u9636\u6bb5\u5f15\u5bfc\u5b66\u751f\u6a21\u4eff\u8001\u5e08\u7684\u4fe1\u606f\u3002\u5f53\u79bb\u7ebfKD\u4f7f\u7528\u516c\u5f00\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u6765\u8bad\u7ec3\u5b66\u751f\u7f51\u7edc\u65f6\uff0c\u79bb\u7ebfKD\u4e5f\u53ef\u4ee5\u88ab\u89c6\u4e3a\u4e00\u4e2a\u5355\u9636\u6bb5\u7684\u8bad\u7ec3\u6d41\u7a0b\u3002\u7531\u4e8e\u6559\u5e08\u7f51\u7edc\u662f\u9884\u5148\u8bad\u7ec3\u548c\u51bb\u7ed3\u7684\uff0c\u6211\u4eec\u5c06\u57fa\u4e8e\u6559\u5e08-\u5b66\u751f\u7684\u5b66\u4e60\u79f0\u4e3a\u79bb\u7ebfKD\u3002<\/p>\n<p data-pid=\"xhWCG2Ps\">\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\u7684\u5de5\u4f5c\u4e3b\u8981\u5728\u7b2c\u4e00\u7ae0\u8282\u201dCategories of response-based, feature-based, and relation-based knowledge distillation(\u57fa\u4e8e\u54cd\u5e94\u3001\u7279\u5f81\u3001\u5173\u7cfb\u7684\u77e5\u8bc6\u84b8\u998f\u5212\u5206)\u201c\u8fdb\u884c\u8bb2\u89e3\u3002\u56e0\u6b64\u5728\u8fd9\u91cc\u4ec5\u5c55\u793a\u4e86\u4e0d\u540c\u84b8\u998f\u65b9\u6cd5\u5bf9\u4e8e\u84b8\u998f\u6027\u80fd\u548c\u84b8\u998f\u65f6\u95f4\u7684\u6743\u8861,\u5982\u88684\u6240\u793a\u3002\u6211\u4eec\u5168\u9762\u6bd4\u8f83\u4e86\u4ee3\u8868\u6027\u7684\u79bb\u7ebfKD\u65b9\u6cd5\u5728\u51c6\u786e\u6027\u548c\u84b8\u998f\u65f6\u95f4\u65b9\u9762\u7684\u8868\u73b0\u3002\u6211\u4eec\u53ef\u4ee5\u89c2\u5bdf\u5230\uff0c\u5404\u79cdKD\u65b9\u6cd5\u5177\u6709\u4e0d\u540c\u7684\u6027\u8d28\u3002\u4f20\u7edf\u7684KD\u84b8\u998f\u65f6\u95f4\u6700\u77ed\uff0c\u4f46\u4ec5\u80fd\u83b7\u5f97\u9002\u5ea6\u7684\u589e\u76ca\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cHSAKD\u5b9e\u73b0\u4e86\u6700\u4f73\u7684\u84b8\u998f\u6027\u80fd\uff0c\u751a\u81f3\u4e0e\u6559\u5e08\u7684\u51c6\u786e\u7387\u76f8\u5339\u914d\uff0c\u4f46\u65f6\u95f4\u6bd4\u4f20\u7edfKD\u9ad8\u51fa3\u500d\u3002DKD\u662f\u4f20\u7edfKD\u7684\u6539\u8fdb\u7248\u672c\uff0c\u5728\u7cbe\u5ea6\u548c\u84b8\u998f\u65f6\u95f4\u4e4b\u95f4\u8fbe\u5230\u4e86\u7406\u60f3\u7684\u5e73\u8861\u3002\u4ece\u6a21\u578b\u538b\u7f29\u7684\u89d2\u5ea6\u6765\u770b\uff0c\u6700\u4f73\u84b8\u998f\u7684ResNet-20\u7684\u53c2\u6570\u548cFLOP\u51cf\u5c11\u4e863\u500d\uff0c\u4f46\u4e0e\u6559\u5e08ResNet-56\u76f8\u6bd4\uff0c\u6027\u80fd\u4ec5\u4e0b\u964d\u4e860.1%\u3002\u5728\u5b9e\u8df5\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u6839\u636e\u5b9e\u9645\u9700\u6c42\u548c\u8ba1\u7b97\u8d44\u6e90\u9009\u62e9\u5408\u9002\u7684KD\u7b97\u6cd5\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15a920c16747f60b73c1e3764503aab7.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15a920c16747f60b73c1e3764503aab7.jpg\"><\/a><figcaption>\u88684 \u4e0d\u540c\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u5728CIFAR-100\u6570\u636e\u96c6\u4e0a\u6027\u80fd\u548c\u8bad\u7ec3\u65f6\u95f4\u7684\u6743\u8861<\/figcaption><\/figure>\n<h3><span class=\"ez-toc-section\" id=\"22_Online_Knowledge_Distillation%EF%BC%88%E5%9C%A8%E7%BA%BF%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>2.2 Online Knowledge Distillation\uff08\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-4a28f175bc2a4acf6349874926525a1c.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-4a28f175bc2a4acf6349874926525a1c.jpg\"><\/a><figcaption>\u56fe3 \u4e24\u4e2a\u5b66\u751f\u7f51\u7edc\u57fa\u4e8e\u54cd\u5e94\u3001\u7279\u5f81\u3001\u5173\u7cfb\u7684\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\u6574\u4f53\u793a\u610f\u56fe<\/figcaption><\/figure>\n<p data-pid=\"1Mw9yCW_\">\u5728\u7ebfKD\u65e8\u5728\u4ece\u96f6\u5f00\u59cb\u540c\u65f6\u8bad\u7ec3\u4e00\u7ec4\u5b66\u751f\u7f51\u7edc\uff0c\u5e76\u5728\u8bad\u7ec3\u9636\u6bb5\u76f8\u4e92\u4f20\u9012\u77e5\u8bc6\u3002\u4e0e\u79bb\u7ebfKD\u4e0d\u540c\uff0c\u5728\u7ebfKD\u662f\u4e00\u4e2a\u7aef\u5230\u7aef\u7684\u4f18\u5316\u8fc7\u7a0b\uff0c\u4e0d\u9700\u8981\u4e8b\u5148\u660e\u786e\u7684\u9884\u8bad\u7ec3\u6559\u5e08\u7f51\u7edc\u3002\u6839\u636e\u77e5\u8bc6\u7c7b\u578b\uff0c\u5f53\u524d\u7684\u5728\u7ebfKD\u4e3b\u8981\u5206\u4e3a\u57fa\u4e8e\u54cd\u5e94\u3001\u57fa\u4e8e\u7279\u5f81\u548c\u57fa\u4e8e\u5173\u7cfb\u7684\u65b9\u6cd5\uff0c\u5982\u56fe3\u6240\u793a\u3002\u6211\u4eec\u5728\u88685\u4e2d\u63d0\u4f9b\u4e86\u5404\u79cd\u5728\u7ebfKD\u65b9\u6cd5\u7684\u5b9e\u9a8c\u7ed3\u679c\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-94a54a4d004af9e601c42151a59d4976.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-94a54a4d004af9e601c42151a59d4976.jpg\"><\/a><figcaption>\u88685 \u4e0d\u540c\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u5728CIFAR-100\u4e0a\u7684\u51c6\u786e\u7387\u5bf9\u6bd4<\/figcaption><\/figure>\n<p data-pid=\"_ICOX_gl\">\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\u6309\u7167\u8fc1\u79fb\u7684\u77e5\u8bc6\u7c7b\u578b\u5206\u4e3a\u57fa\u4e8e\u54cd\u5e94\u3001\u7279\u5f81\u548c\u5173\u7cfb\u7684\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\uff1a<\/p>\n<p data-pid=\"6czibaPm\">2.2.1 Response-based Online KD\uff08\u57fa\u4e8e\u54cd\u5e94\u7684\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"jHvUl0pm\">\u5728\u7ebfKD\u53ef\u4ee5\u8ffd\u6eaf\u5230\u6df1\u5ea6\u76f8\u4e92\u5b66\u4e60\uff08Deep Mutual Learning, DML\uff09\u3002DML\u8868\u660e\uff0c\u5728\u4f20\u7edf\u5b66\u4e60\u65b9\u6848\u4e2d\uff0c\u5c06\u6bcf\u4e2a\u5b66\u751f\u7684\u7c7b\u522b\u540e\u9a8c\u4e0e\u5176\u4ed6\u5b66\u751f\u7684\u7c7b\u522b\u540e\u9a8c\u5bf9\u9f50\u6bd4\u5355\u72ec\u8bad\u7ec3\u66f4\u597d\u3002Song\u7b49\u4eba\u5c06\u8fd9\u4e00\u60f3\u6cd5\u8fdb\u4e00\u6b65\u6269\u5c55\u5230\u5177\u6709\u5171\u4eab\u6d45\u5c42\u548c\u5206\u79bb\u7684\u9ad8\u5c42\u5206\u652f\u7684\u5206\u5c42\u67b6\u6784\u3002Anil\u7b49\u4eba\u5c06\u76f8\u4e92\u84b8\u998f\u5e94\u7528\u4e8e\u5927\u89c4\u6a21\u5206\u5e03\u5f0f\u795e\u7ecf\u7f51\u7edc\u3002DCML\u901a\u8fc7\u5728\u9690\u85cf\u5c42\u4e2d\u6dfb\u52a0\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u8f85\u52a9\u5206\u7c7b\u5668\u6765\u589e\u5f3a\u76f8\u4e92\u5b66\u4e60\u3002MutualNet\u4f7f\u7528\u4e0d\u540c\u8f93\u5165\u5206\u8fa8\u7387\u5bf9\u5e94\u7684\u4e0d\u540c\u5bbd\u5ea6\u7684\u5b50\u7f51\u7edc\u8fdb\u884c\u76f8\u4e92\u5b66\u4e60\uff0c\u4ee5\u63a2\u7d22\u591a\u5c3a\u5ea6\u7279\u5f81\u3002MMT\u548cPCL\u4e3a\u6bcf\u4e2a\u5bf9\u7b49\u4f53\u5f15\u5165\u4e86\u4e00\u4e2a\u65f6\u95f4\u5747\u503c\u6559\u5e08\uff0c\u4e3a\u76f8\u4e92\u5b66\u4e60\u751f\u6210\u66f4\u597d\u7684\u4f2a\u6807\u7b7e\u3002\u9664\u4e86\u540c\u4f34\u6559\u5b66\u65b9\u5f0f\uff0cONE\u8fd8\u7ec4\u88c5\u4e86\u96c6\u6210\u6982\u7387\u5206\u5e03\uff0c\u6784\u5efa\u4e86\u4e00\u4e2a\u865a\u62df\u6559\u5e08\u89d2\u8272\uff0c\u4ee5\u63d0\u4f9b\u8f6f\u6807\u7b7e\u3002OKDDip\u5229\u7528\u81ea\u6ce8\u610f\u529b\u673a\u5236\u6765\u63d0\u9ad8\u540c\u4f34\u591a\u6837\u6027\uff0c\u7136\u540e\u5c06\u8f85\u52a9\u540c\u4f34\u7684\u96c6\u5408\u77e5\u8bc6\u4f20\u9012\u7ed9\u7ec4\u957f\u3002KDCL\u7814\u7a76\u4e86\u4ece\u4e24\u4e2a\u6570\u636e\u589e\u5f3a\u89c6\u56fe\u4e2d\u4f7f\u7528\u5404\u79cd\u805a\u5408\u7b56\u7565\u751f\u6210\u8f6f\u96c6\u6210\u76ee\u6807\u3002\u4e00\u4e9b\u5de5\u4f5c\u8003\u8651\u4f7f\u7528\u7279\u5f81\u878d\u5408\u548c\u989d\u5916\u7684\u5206\u7c7b\u5668\u6765\u8f93\u51fa\u6709\u610f\u4e49\u7684\u6807\u7b7e\u3002<\/p>\n<p data-pid=\"IxCutwDD\">2.2.2 Feature-based Online KD\uff08\u57fa\u4e8e\u7279\u5f81\u7684\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"gp8YAUcC\">\u4ee5\u524d\u7684\u5728\u7ebfKD\u65b9\u6cd5\u901a\u5e38\u4fa7\u91cd\u4e8e\u5b66\u4e60\u7c7b\u6982\u7387\uff0c\u4e3b\u8981\u5728\u5404\u79cd\u7b56\u7565\u6216\u67b6\u6784\u4e0a\u6709\u6240\u4e0d\u540c\uff0c\u4f46\u5ffd\u7565\u4e86\u5728\u7ebf\u5b66\u4e60\u7684\u7279\u5f81\u7ea7\u4fe1\u606f\u3002Walawalkar\u7b49\u4eba\u5bf9\u6a21\u578b\u538b\u7f29\u7684\u4e2d\u95f4\u7279\u5f81\u56fe\u8fdb\u884c\u4e86\u5728\u7ebf\u84b8\u998f\u3002Zhang\u7b49\u4eba\u8868\u660e\uff0c\u76f4\u63a5\u5bf9\u9f50\u7279\u5f81\u56fe\u53ef\u80fd\u4f1a\u51cf\u5c0f\u7fa4\u4f53\u591a\u6837\u6027\u5e76\u635f\u5bb3\u5728\u7ebfKD\u3002\u8bb8\u591a\u65b9\u6cd5\u90fd\u63d0\u51fa\u4e86\u5728\u7ebf\u5bf9\u6297\u7279\u5f81\u63d0\u53d6\u6765\u76f8\u4e92\u5b66\u4e60\u7279\u5f81\u5206\u5e03\u3002\u5bf9\u6297\u6027\u5728\u7ebfKD\u7684\u60f3\u6cd5\u662f\u4e3a\u6bcf\u4e2a\u7f51\u7edc\u6dfb\u52a0\u4e00\u4e2a\u9274\u522b\u5668\uff0c\u8be5\u9274\u522b\u5668\u53ef\u4ee5\u5c06\u81ea\u8eab\u7684\u7279\u5f81\u56fe\u5206\u7c7b\u4e3a\u4f2a\u9020\uff0c\u6216\u5c06\u53e6\u4e00\u4e2a\u7f51\u7edc\u7684\u7279\u5f81\u56fe\u5206\u7c7b\u4e3a\u771f\u5b9e\u3002<\/p>\n<p data-pid=\"nUy1kUCH\">2.2.3 Relation-based Online KD\uff08\u57fa\u4e8e\u5173\u7cfb\u7684\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"N0pXHOLK\">\u76f8\u4e92\u5bf9\u6bd4\u5b66\u4e60\uff08Mutual Contrastive Learning, MCL)\u5c06\u6bcf\u4e2a\u7f51\u7edc\u89c6\u4e3a\u4e00\u4e2a\u5355\u72ec\u7684\u89c6\u56fe\uff0c\u5e76\u4ece\u5bf9\u6bd4\u8868\u5f81\u5b66\u4e60\u7684\u89d2\u5ea6\u5f15\u5165\u4e86\u57fa\u4e8e\u76f8\u4e92\u5173\u7cfb\u7684\u84b8\u998f\u3002\u4e0e\u4e4b\u524d\u7684\u5de5\u4f5c\u76f8\u6bd4\uff0cMCL\u5e2e\u52a9\u6bcf\u4e2a\u7f51\u7edc\u5b66\u4e60\u66f4\u597d\u7684\u89c6\u89c9\u7279\u5f81\u8868\u793a\u3002\u5148\u524d\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u7684\u4e00\u4e2a\u5171\u540c\u70b9\u662f\u4ece\u5355\u4e2a\u539f\u59cb\u4efb\u52a1\u4e2d\u63d0\u53d6\u77e5\u8bc6\u3002HSSAKD\u5728\u7279\u5f81\u6620\u5c04\u540e\u9644\u52a0\u5206\u7c7b\u5668\uff0c\u4ee5\u5b66\u4e60\u989d\u5916\u7684\u81ea\u76d1\u7763\u589e\u5f3a\u4efb\u52a1\uff0c\u5e76\u5f15\u5bfc\u7f51\u7edc\u76f8\u4e92\u84b8\u998f\u81ea\u76d1\u7763\u5206\u5e03\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"23_Self-Knowledge_Distillation%EF%BC%88%E8%87%AA%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>2.3 Self-Knowledge Distillation\uff08\u81ea\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"sLrWn-cv\">Self-KD\u65e8\u5728\u4ece\u7f51\u7edc\u4e2d\u63d0\u53d6\u77e5\u8bc6\u5e76\u81ea\u5b66\u3002\u4e0e\u79bb\u7ebf\u548c\u5728\u7ebfKD\u4e0d\u540c\uff0cSelf-KD\u6ca1\u6709\u989d\u5916\u7684\u6559\u5e08\u6216\u540c\u4f34\u7f51\u7edc\u8fdb\u884c\u77e5\u8bc6\u4ea4\u6d41\u3002\u56e0\u6b64\uff0c\u73b0\u6709\u7684Self-KD\u65b9\u6cd5\u901a\u5e38\u5229\u7528\u8f85\u52a9\u67b6\u6784\uff0c\u6570\u636e\u589e\u5f3a\u6216\u5e8f\u5217\u5feb\u7167\u84b8\u998f\u6765\u63a2\u7d22\u5916\u90e8\u77e5\u8bc6\u4ee5\u5b9e\u73b0\u81ea\u6211\u63d0\u5347\u3002\u6b64\u5916\uff0c\u901a\u8fc7\u624b\u52a8\u8bbe\u8ba1\u6b63\u5219\u5316\u5206\u5e03\u6765\u4ee3\u66ff\u6559\u5e08\uff0cSelf-KD\u4e5f\u53ef\u4ee5\u4e0e\u6807\u7b7e\u5e73\u6ed1\u8054\u7cfb\u8d77\u6765\u3002\u6211\u4eec\u5728\u88686\u4e2d\u63d0\u4f9b\u4e86\u5404\u79cdSelf-KD\u65b9\u6cd5\u7684\u5b9e\u9a8c\u7ed3\u679c\u5bf9\u6bd4\u3002\u9664\u4e86\u5c06Self-KD\u5e94\u7528\u4e8e\u4f20\u7edf\u7684\u76d1\u7763\u5b66\u4e60\u4e4b\u5916\uff0c\u6700\u8fd1\u7684\u7814\u7a76\u8fd8\u8bd5\u56fe\u501f\u9274Self-KD\u7684\u601d\u60f3\u8fdb\u884c\u81ea\u76d1\u7763\u5b66\u4e60\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-02c608a07a43fe4d660e45f4dbc4704d.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-02c608a07a43fe4d660e45f4dbc4704d.jpg\"><\/a><figcaption>\u88686 \u4e0d\u540c\u81ea\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u5728CIFAR-100\u4e0a\u7684\u51c6\u786e\u7387\u5bf9\u6bd4<\/figcaption><\/figure>\n<p data-pid=\"70t4CqaW\">2.3.1 Self-KD with auxiliary architecture\uff08\u57fa\u4e8e\u8f85\u52a9\u7ed3\u6784\u7684\u81ea\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"Ni-VQa9o\">\u8fd9\u79cd\u65b9\u6cd5\u7684\u60f3\u6cd5\u662f\u9644\u52a0\u8f85\u52a9\u67b6\u6784\u6765\u6355\u83b7\u989d\u5916\u7684\u77e5\u8bc6\uff0c\u4ee5\u8865\u5145\u4e3b\u7f51\u7edc\u3002DKS\u63d2\u5165\u51e0\u4e2a\u8f85\u52a9\u5206\u652f\uff0c\u5e76\u5728\u8fd9\u4e9b\u5206\u652f\u548c\u4e3b\u9aa8\u5e72\u4e4b\u95f4\u8fdb\u884c\u6210\u5bf9\u7684\u77e5\u8bc6\u8f6c\u79fb\u3002BYOT\u5c06\u6982\u7387\u548c\u7279\u5f81\u4fe1\u606f\u4ece\u7f51\u7edc\u7684\u8f83\u6df1\u90e8\u5206\u8f6c\u79fb\u5230\u8f83\u6d45\u90e8\u5206\u3002SAD\u4f7f\u7528\u6765\u81ea\u6df1\u5c42\u7684\u6ce8\u610f\u529b\u56fe\uff0c\u4ee5\u9010\u5c42\u7684\u65b9\u5f0f\u76d1\u7763\u6d45\u5c42\u7684\u6ce8\u610f\u529b\u56fe\u3002\u9664\u4e86\u70b9\u5bf9\u70b9\u4f20\u8f93\uff0cMetaDistiller\u8fd8\u901a\u8fc7\u81ea\u4e0a\u800c\u4e0b\u7684\u65b9\u5f0f\u878d\u5408\u7279\u5f81\u56fe\u6765\u6784\u5efa\u6807\u7b7e\u751f\u6210\u5668\uff0c\u5e76\u4f7f\u7528\u5143\u5b66\u4e60\u5bf9\u5176\u8fdb\u884c\u4f18\u5316\u3002FRSKD\u4ee5\u7c7b\u4f3cBiFPN\u7684\u65b9\u5f0f\u805a\u5408\u7279\u5f81\u56fe\uff0c\u4ee5\u6784\u5efa\u4e00\u4e2a\u81ea\u5b66\u7f51\u7edc\uff0c\u63d0\u4f9b\u9ad8\u8d28\u91cf\u7684\u7279\u5f81\u56fe\u548c\u8f6f\u6807\u7b7e\u3002\u4e00\u4e2a\u95ee\u9898\u662f\uff0c\u57fa\u4e8e\u8f85\u52a9\u67b6\u6784\u7684\u65b9\u6cd5\u9ad8\u5ea6\u4f9d\u8d56\u4e8e\u4eba\u5de5\u8bbe\u8ba1\u7684\u7f51\u7edc\uff0c\u5176\u53ef\u6269\u5c55\u6027\u8f83\u5dee\u3002<\/p>\n<p data-pid=\"w8T-KYX9\">2.3.2 Self-KD with data augmentation\uff08\u57fa\u4e8e\u6570\u636e\u589e\u5f3a\u7684\u81ea\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"THhAuCJp\">\u57fa\u4e8e\u6570\u636e\u589e\u5f3a\u7684\u65b9\u6cd5\u901a\u5e38\u4f1a\u5f3a\u5236\u4ece\u4e24\u4e2a\u4e0d\u540c\u7684\u589e\u5f3a\u89c6\u56fe\u751f\u6210\u7c7b\u4f3c\u7684\u9884\u6d4b\u3002\u6cbf\u7740\u8fd9\u4e00\u601d\u8def\uff0cDDGSD\u5bf9\u540c\u4e00\u56fe\u50cf\u5e94\u7528\u4e86\u4e24\u79cd\u4e0d\u540c\u7684\u589e\u5f3a\u7b97\u5b50\u3002CS-KD\u4ece\u540c\u4e00\u7c7b\u522b\u4e2d\u968f\u673a\u62bd\u53d6\u4e24\u4e2a\u4e0d\u540c\u7684\u5b9e\u4f8b\u3002MixSKD\u5c06Mixup\u56fe\u50cf\u89c6\u4e3a\u7279\u5f81\u548c\u6982\u7387\u7a7a\u95f4\u4e2d\u7684\u4e00\u4e2a\u89c6\u56fe\uff0c\u5c06\u7ebf\u6027\u63d2\u503c\u56fe\u50cf\u89c6\u4e3a\u53e6\u4e00\u4e2a\u89c6\u56fe\u3002\u4e3a\u4e86\u6316\u6398\u8de8\u56fe\u50cf\u77e5\u8bc6\uff0cBAKE\u8bd5\u56fe\u901a\u8fc7\u52a0\u6743\u805a\u5408\u5176\u4ed6\u6837\u672c\u7684\u77e5\u8bc6\uff0c\u5f62\u6210\u4e00\u4e2a\u8f6f\u76ee\u6807\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u4e0e\u57fa\u7ebf\u76f8\u6bd4\uff0c\u57fa\u4e8e\u6570\u636e\u589e\u5f3a\u7684Self-KD\u9700\u8981\u591a\u4e2a\u6b63\u5411\u8fc7\u7a0b\uff0c\u56e0\u6b64\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n<p data-pid=\"9KOtv6C9\">2.3.3 Self-KD with sequential snapshot distillation\uff08\u57fa\u4e8e\u5e8f\u5217\u5feb\u7167\u7684\u81ea\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"eHLxgm4-\">\u8fd9\u4e00\u601d\u8def\u8003\u8651\u5229\u7528\u8bad\u7ec3\u8f68\u8ff9\u4e0a\u7684\u7f51\u7edc\u5feb\u7167\u6765\u63d0\u4f9b\u76d1\u7763\u4fe1\u53f7\u3002BAN\u5728\u5148\u524d\u53d7\u8fc7\u8bad\u7ec3\u7684\u7f51\u7edc\u5feb\u7167\u76d1\u7763\u4e0b\uff0c\u4ee5\u8fde\u7eed\u7684\u65b9\u5f0f\u9010\u6b65\u6539\u8fdb\u7f51\u7edc\u3002SD\u4ece\u65e9\u671f\u65f6\u4ee3\u83b7\u53d6\u7f51\u7edc\u5feb\u7167\uff0c\u4ee5\u6559\u5bfc\u5176\u540e\u671f\u65f6\u4ee3\u3002PS-KD\u5efa\u8bae\u901a\u8fc7\u603b\u7ed3\u771f\u5b9e\u6807\u7b7e\u548c\u8fc7\u53bb\u7684\u9884\u6d4b\u6765\u9010\u6b65\u5b8c\u5584\u8f6f\u76ee\u6807\u3002DLB\u5728\u4e0a\u4e00\u4e2a\u548c\u5f53\u524d\u7684\u5c0f\u6279\u91cf\u6570\u636e\u4e4b\u95f4\u6267\u884c\u4e00\u81f4\u6027\u6b63\u5219\u5316\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u57fa\u4e8e\u5feb\u7167\u7684Self-KD\u9700\u8981\u4fdd\u5b58\u8bad\u7ec3\u6a21\u578b\u7684\u591a\u4e2a\u526f\u672c\uff0c\u56e0\u6b64\u589e\u52a0\u4e86\u5185\u5b58\u6210\u672c\u3002<\/p>\n<p data-pid=\"edkmBGaW\">2.3.4 Self-supervised learning with Self-KD\uff08\u81ea\u76d1\u7763\u5b66\u4e60\u548c\u81ea\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"I2yzVjl0\">\u81ea\u76d1\u7763\u5b66\u4e60\u4fa7\u91cd\u4e8e\u7ed9\u5b9a\u672a\u6807\u6ce8\u6570\u636e\u5b66\u4e60\u5230\u597d\u7684\u7279\u5f81\u8868\u793a\u3002Self-KD\u548c\u81ea\u76d1\u7763\u5b66\u4e60\u4e4b\u95f4\u6709\u4e00\u4e9b\u6709\u8da3\u7684\u8054\u7cfb\u3002\u5728\u81ea\u76d1\u7763\u7684\u573a\u666f\u4e2d\uff0c\u8be5\u6846\u67b6\u901a\u5e38\u6784\u5efa\u4e24\u4e2a\u89d2\u8272\uff1a\u5728\u7ebf\u548c\u76ee\u6807\u7f51\u7edc\u3002\u524d\u8005\u662f\u8bad\u7ec3\u7f51\u7edc\uff0c\u540e\u8005\u662f\u5e73\u5747\u6559\u5e08\uff0c\u5177\u6709\u6765\u81ea\u5728\u7ebf\u7f51\u7edc\u7684\u79fb\u52a8\u5e73\u5747\u6743\u91cd\u3002\u76ee\u6807\u7f51\u7edc\u4e0e\u5728\u7ebf\u7f51\u7edc\u5177\u6709\u76f8\u540c\u7684\u67b6\u6784\uff0c\u4f46\u6743\u91cd\u4e0d\u540c\u3002\u76ee\u6807\u7f51\u7edc\u901a\u5e38\u7528\u4e8e\u63d0\u4f9b\u76d1\u7763\u4fe1\u53f7\u6765\u8bad\u7ec3\u5728\u7ebf\u7f51\u7edc\u3002MoCo\u5229\u7528\u76ee\u6807\u7f51\u7edc\u751f\u6210\u4e00\u81f4\u7684\u6b63\u8d1f\u5bf9\u6bd4\u6837\u672c\u3002\u4e00\u4e9b\u81ea\u76d1\u7763\u65b9\u6cd5\u5c06\u76ee\u6807\u7f51\u7edc\u89c6\u4e3a\u4e00\u4e2a\u81ea\u6559\u5e08\u6765\u63d0\u4f9b\u56de\u5f52\u76ee\u6807\uff0c\u5982BYOL\u3001DINO\u548cSimSiam\u3002\u53d7BYOT\u7684\u542f\u53d1\uff0cSDSSL\u5f15\u5bfc\u4e2d\u95f4\u7279\u5f81\u5d4c\u5165\uff0c\u4ee5\u5bf9\u6bd4\u6700\u7ec8\u5c42\u7684\u7279\u5f81\u3002\u5c3d\u7ba1\u4e4b\u524d\u7684\u65b9\u6cd5\u5728\u81ea\u76d1\u7763\u8868\u793a\u5b66\u4e60\u65b9\u9762\u53d6\u5f97\u4e86\u7406\u60f3\u7684\u6027\u80fd\uff0c\u4f46\u53ef\u80fd\u4ecd\u5b58\u5728\u4e24\u4e2a\u503c\u5f97\u63a2\u7d22\u7684\u65b9\u5411\u3002\u9996\u5148\uff0c\u76ee\u6807\u7f51\u7edc\u662f\u4ee5\u79fb\u52a8\u5e73\u5747\u7684\u65b9\u5f0f\u4ece\u5728\u7ebf\u7f51\u7edc\u4e2d\u6784\u5efa\u7684\u3002\u6211\u4eec\u662f\u5426\u6709\u66f4\u6709\u610f\u4e49\u7684\u65b9\u6cd5\u6765\u6784\u5efa\u76ee\u6807\u7f51\u7edc\uff1f\u5176\u6b21\uff0c\u635f\u5931\u901a\u5e38\u662f\u4e3a\u4e86\u5bf9\u9f50\u6700\u7ec8\u7684\u7279\u5f81\u5d4c\u5165\uff0c\u672a\u6765\u53ef\u4ee5\u8fdb\u4e00\u6b65\u6316\u6398\u5728\u7ebf\u7f51\u7edc\u548c\u76ee\u6807\u7f51\u7edc\u4e4b\u95f4\u7684\u4e00\u4e9b\u4e2d\u95f4\u7279\u5f81\u6216\u5bf9\u6bd4\u5173\u7cfb\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"24_%E4%B8%89%E7%A7%8D%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%E7%BB%BC%E5%90%88%E6%AF%94%E8%BE%83\"><\/span>2.4 \u4e09\u79cd\u77e5\u8bc6\u84b8\u998f\u7efc\u5408\u6bd4\u8f83<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-192200b53a5b2224d694f4947f5f41c0.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-192200b53a5b2224d694f4947f5f41c0.jpg\"><\/a><figcaption>\u88687 \u4e09\u79cd\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u5728\u4e0d\u540c\u65b9\u9762\u7684\u6bd4\u8f83\uff0c\u5176\u4e2d*\u4ee3\u8868\u9884\u8bad\u7ec3\u7684\u6559\u5e08\u662f\u73b0\u6210\u7684<\/figcaption><\/figure>\n<p data-pid=\"rKaFOjGW\">\u5982\u88687\u6240\u793a\uff0c\u672c\u6587\u5c55\u793a\u4e09\u79cd\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u5728\u4e0d\u540c\u65b9\u9762\u7684\u6bd4\u8f83\u3002\u79bb\u7ebf\u84b8\u998f\u9700\u8981\u4e00\u4e2a\u989d\u5916\u7684\u6559\u5e08\u6a21\u578b\u6765\u8bad\u7ec3\u5b66\u751f\uff0c\u800c\u5728\u7ebf\u84b8\u998f\u6216\u81ea\u84b8\u998f\u901a\u8fc7\u7aef\u5230\u7aef\u4f18\u5316\u8bad\u7ec3\u4e00\u7ec4\u6a21\u578b\u6216\u5355\u4e2a\u6a21\u578b\u3002\u5f53\u516c\u5f00\u53ef\u7528\u7684\u9884\u8bad\u7ec3\u6559\u5e08\u6a21\u578b\u4e0d\u53ef\u7528\u65f6\uff0c\u79bb\u7ebf\u84b8\u998f\u9700\u8981\u9884\u5148\u8bad\u7ec3\u4e00\u4e2a\u9ad8\u6027\u80fd\u7684\u6559\u5e08\u7f51\u7edc\uff0c\u56e0\u6b64\u8bad\u7ec3\u4ee3\u4ef7\u8f83\u9ad8\u3002\u81ea\u84b8\u998f\u5229\u7528\u5355\u4e00\u6a21\u578b\u5b9e\u73b0\u81ea\u6211\u63d0\u5347\uff0c\u901a\u5e38\u5177\u6709\u8f83\u4f4e\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u79bb\u7ebf\u84b8\u998f\u5728\u9884\u8bad\u7ec3\u6559\u5e08\u6a21\u578b\u4e3a\u516c\u5f00\u6a21\u578b\u65f6\u4e5f\u5177\u6709\u8f83\u4f4e\u7684\u8bad\u7ec3\u590d\u6742\u6027\uff0c\u56e0\u4e3a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6559\u5e08\u7f51\u7edc\u662f\u4e0d\u9700\u8981\u68af\u5ea6\u4f20\u64ad\u8bad\u7ec3\u7684\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"3_Distillation_Algorithms_%EF%BC%88%E8%92%B8%E9%A6%8F%E7%AE%97%E6%B3%95%EF%BC%89\"><\/span>3. Distillation Algorithms \uff08\u84b8\u998f\u7b97\u6cd5\uff09<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<h3><span class=\"ez-toc-section\" id=\"31_Multi-Teacher_Distillation%EF%BC%88%E5%A4%9A%E6%95%99%E5%B8%88%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>3.1 Multi-Teacher Distillation\uff08\u591a\u6559\u5e08\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"IWTjXjT4\">\u5728\u4f20\u7edf\u7684KD\u4e2d\uff0c\u77e5\u8bc6\u4ece\u4e00\u4e2a\u9ad8\u6027\u80fd\u7684\u8001\u5e08\u8f6c\u79fb\u5230\u4e00\u4e2a\u8f7b\u91cf\u7684\u5b66\u751f\u8eab\u4e0a\uff0c\u4f46\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u77e5\u8bc6\u7684\u591a\u6837\u6027\u548c\u6709\u6548\u6027\u662f\u6709\u9650\u7684\u3002\u4e0d\u540c\u7684\u8001\u5e08\u53ef\u4ee5\u4e3a\u5b66\u751f\u63d0\u4f9b\u4ed6\u4eec\u72ec\u7279\u800c\u6709\u4ef7\u503c\u7684\u77e5\u8bc6\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u5b66\u751f\u53ef\u4ee5\u4ece\u591a\u4e2a\u6559\u5e08\u7f51\u7edc\u4e2d\u5b66\u4e60\u5404\u79cd\u77e5\u8bc6\u8868\u793a\u3002\u9075\u5faa\u4f20\u7edf\u7684KD\uff0c\u903b\u8f91\u5206\u5e03\u6216\u4e2d\u5c42\u7279\u5f81\u5f62\u5f0f\u7684\u77e5\u8bc6\u53ef\u4ee5\u7528\u4f5c\u76d1\u7763\u4fe1\u53f7\u3002\u591a\u6559\u5e08KD\u7684\u793a\u610f\u56fe\u5982\u4e0b\u56fe4\u6240\u793a\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8b3dc239e96efcdf0692144e19bd459f.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8b3dc239e96efcdf0692144e19bd459f.jpg\"><\/a><figcaption>\u56fe4 \u591a\u6559\u5e08\u77e5\u8bc6\u84b8\u998f\u57fa\u672c\u601d\u60f3\u793a\u610f\u56fe<\/figcaption><\/figure>\n<p data-pid=\"R5kW3Nr7\">3.1.1 KD from ensemble logits\uff08\u4ece\u96c6\u6210\u7684\u903b\u8f91\u5206\u5e03\u4e2d\u84b8\u998f\uff09<\/p>\n<p data-pid=\"A_9TlWv0\">\u6a21\u578b\u96c6\u6210\u7684\u903b\u8f91\u5206\u5e03\u662f\u591a\u6559\u5e08\u77e5\u8bc6\u84b8\u998f\u7684\u76f4\u63a5\u65b9\u6cd5\u4e4b\u4e00\u3002\u57fa\u4e8e\u8fd9\u4e00\u7406\u5ff5\uff0c\u5f15\u5bfc\u5b66\u751f\u5b66\u4e60\u6559\u5e08\u96c6\u6210\u903b\u8f91\u7684\u8f6f\u8f93\u51fa\u3002\u7136\u800c\uff0c\u5bf9\u591a\u6559\u5e08\u7684\u9884\u6d4b\u8fdb\u884c\u7b80\u5355\u5e73\u5747\u53ef\u80fd\u4f1a\u5ffd\u89c6\u6559\u5e08\u7fa4\u4f53\u591a\u6837\u6027\u3002\u56e0\u6b64\uff0c\u4e00\u4e9b\u5de5\u4f5c\u63d0\u51fa\u4e86\u901a\u8fc7\u81ea\u9002\u5e94\u5730\u6a21\u4eff\u6559\u5e08\u7684\u8f93\u51fa\u5e76\u4f7f\u7528\u5404\u79cd\u805a\u5408\u6743\u91cd\u6765\u5b66\u4e60\u5b66\u751f\u6a21\u578b\u3002<\/p>\n<p data-pid=\"ZGhGck4_\">3.1.2 KD from ensemble feature representations\uff08\u4ece\u96c6\u6210\u7684\u7279\u5f81\u8868\u8fbe\u4e2d\u84b8\u998f\uff09<\/p>\n<p data-pid=\"zX7I9EAI\">\u9664\u4e86\u4ece\u903b\u8f91\u5206\u5e03\u4e2d\u63d0\u53d6\u5916\uff0c\u4e2d\u95f4\u5c42\u7279\u5f81\u7684\u96c6\u6210\u53ef\u4ee5\u4e3a\u5b66\u751f\u63d0\u4f9b\u66f4\u591a\u7684\u8bed\u4e49\u4fe1\u606f\u3002\u7136\u800c\uff0c\u4ece\u7279\u5f81\u8868\u793a\u4e2d\u63d0\u53d6\u66f4\u5177\u6311\u6218\u6027\uff0c\u56e0\u4e3a\u6559\u5e08\u6c60\u4e2d\u7684\u6bcf\u4e2a\u6210\u5458\u5728\u7279\u5b9a\u5c42\u4e2d\u90fd\u6709\u4e0d\u540c\u7684\u7279\u5f81\u8868\u793a\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cPark\u7b49\u4eba\u5728\u7279\u5f81\u56fe\u7ea7\u522b\u5bf9\u591a\u4e2a\u6559\u5e08\u7f51\u7edc\u5e94\u7528\u4e86\u975e\u7ebf\u6027\u53d8\u6362\u3002Wu\u7b49\u4eba\u63d0\u51fa\u901a\u8fc7\u6700\u5c0f\u5316\u6559\u5e08\u548c\u5b66\u751f\u7684\u76f8\u4f3c\u6027\u77e9\u9635\u4e4b\u95f4\u7684\u8ddd\u79bb\u6765\u63d0\u53d6\u77e5\u8bc6\u3002\u5218\u7b49\u4eba\u63d0\u51fa\u8ba9\u5b66\u751f\u7f51\u7edc\u5b66\u4e60\u6559\u5e08\u6a21\u578b\u7684\u53ef\u5b66\u4e60\u8f6c\u6362\u77e9\u9635\u3002\u4e3a\u4e86\u5229\u7528\u903b\u8f91\u548c\u4e2d\u95f4\u7279\u5f81\uff0cChen\u7b49\u4eba\u5f15\u5165\u4e86\u53cc\u6559\u5e08\u7f51\u7edc\uff0c\u5206\u522b\u63d0\u4f9b\u54cd\u5e94\u7ea7\u548c\u7279\u5f81\u7ea7\u77e5\u8bc6\u3002<\/p>\n<p data-pid=\"e2ZszCzP\">3.1.3 Computation-efficient multi-teacher KD from sub-networks\uff08\u4ece\u5b50\u7f51\u7edc\u5f97\u5230\u8ba1\u7b97\u9ad8\u6548\u7684\u84b8\u998f\uff09<\/p>\n<p data-pid=\"OEWMmW8U\">\u4f7f\u7528\u591a\u6559\u5e08\u5f15\u5165\u4e86\u989d\u5916\u7684\u8bad\u7ec3\u8ba1\u7b97\u6210\u672c\uff0c\u56e0\u6b64\u4e00\u4e9b\u65b9\u6cd5\u4ece\u5355\u4e2a\u6559\u5e08\u7f51\u7edc\u4e2d\u521b\u5efa\u4e86\u4e00\u4e9b\u5b50\u6559\u5e08\u3002Nguyen\u7b49\u4eba\u5229\u7528\u968f\u673a\u5757\u548c\u8df3\u8fc7\u6559\u5e08\u7f51\u7edc\u4e0a\u7684\u8fde\u63a5\u6765\u751f\u6210\u591a\u4e2a\u6559\u5e08\u89d2\u8272\u3002\u4e00\u4e9b\u65b9\u6cd5\u8bbe\u8ba1\u4e86\u591a\u5934\u67b6\u6784\uff0c\u4ee5\u4ea7\u751f\u591a\u4e2a\u6559\u5e08\u89d2\u8272\u3002<\/p>\n<p data-pid=\"nwOLl_y0\">3.1.4 Multi-task multi-teacher KD\uff08\u591a\u4efb\u52a1\u591a\u6559\u5e08\u84b8\u998f\uff09<\/p>\n<p data-pid=\"S0n-ZMbK\">\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\uff0c\u591a\u6559\u5e08KD\u662f\u57fa\u4e8e\u540c\u4e00\u4efb\u52a1\u7684\u3002\u77e5\u8bc6\u878d\u5408\u65e8\u5728\u901a\u8fc7\u4ece\u6240\u6709\u63a5\u53d7\u8fc7\u4e0d\u540c\u4efb\u52a1\u8bad\u7ec3\u7684\u6559\u5e08\u90a3\u91cc\u5b66\u4e60\u77e5\u8bc6\uff0c\u6765\u57f9\u517b\u4e00\u4e2a\u591a\u624d\u591a\u827a\u7684\u5b66\u751f\u3002Luoemph\u7b49\u4eba\u65e8\u5728\u5b66\u4e60\u4e00\u4e2a\u80fd\u591f\u4ece\u5f02\u8d28\u6559\u5e08\u90a3\u91cc\u5438\u6536\u5168\u9762\u77e5\u8bc6\u7684\u591a\u4eba\u624d\u5b66\u751f\u7f51\u7edc\u3002Ye \u7b49\u4eba\u5c06\u76ee\u6807\u7f51\u7edc\u96c6\u4e2d\u7528\u4e8e\u5b9a\u5236\u4efb\u52a1\uff0c\u7531\u4ece\u4e0d\u540c\u4efb\u52a1\u4e2d\u9884\u5148\u8bad\u7ec3\u7684\u591a\u540d\u6559\u5e08\u6307\u5bfc\u3002\u5b66\u751f\u4ece\u5f02\u8d28\u6559\u5e08\u90a3\u91cc\u7ee7\u627f\u4e86\u7406\u60f3\u7684\u80fd\u529b\uff0c\u56e0\u6b64\u53ef\u4ee5\u540c\u65f6\u6267\u884c\u591a\u9879\u4efb\u52a1\u3002Rusu\u7b49\u4eba\u5f15\u5165\u4e86\u4e00\u79cd\u591a\u6559\u5e08\u7b56\u7565\u84b8\u998f\u65b9\u6cd5\uff0c\u5c06\u4ee3\u7406\u7684\u591a\u79cd\u7b56\u7565\u8f6c\u79fb\u5230\u5355\u4e2a\u5b66\u751f\u7f51\u7edc\u3002<\/p>\n<p data-pid=\"pHCrnmU0\"><strong>\u603b\u7ed3\u3002<\/strong>\u603b\u4e4b\uff0c\u7531\u4e8e\u4e0d\u540c\u7684\u6559\u5e08\u63d0\u4f9b\u4e0d\u540c\u7684\u77e5\u8bc6\uff0c\u56e0\u6b64\u53ef\u4ee5\u901a\u8fc7\u591a\u6559\u5e08\u84b8\u998f\u6765\u57f9\u517b\u591a\u624d\u591a\u827a\u7684\u5b66\u751f\u3002\u7136\u800c\uff0c\u4ecd\u6709\u51e0\u4e2a\u95ee\u9898\u9700\u8981\u89e3\u51b3\u3002\u4e00\u65b9\u9762\uff0c\u6559\u5e08\u6570\u91cf\u662f\u8bad\u7ec3\u6210\u672c\u548c\u6027\u80fd\u63d0\u9ad8\u4e4b\u95f4\u7684\u6743\u8861\u95ee\u9898\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u6709\u6548\u6574\u5408\u591a\u6559\u5e08\u7684\u5404\u79cd\u77e5\u8bc6\u4ecd\u7136\u662f\u4e00\u4e2a\u503c\u5f97\u63a2\u7d22\u7684\u95ee\u9898\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"32_Cross-Modal_Distillation%EF%BC%88%E8%B7%A8%E6%A8%A1%E6%80%81%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>3.2 Cross-Modal Distillation\uff08\u8de8\u6a21\u6001\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"eGrLBddE\">\u5728\u5e38\u89c1\u7684KD\u65b9\u6cd5\u4e2d\uff0c\u6559\u5e08\u548c\u5b66\u751f\u5f80\u5f80\u5177\u6709\u76f8\u540c\u7684\u6a21\u6001\u3002\u7136\u800c\uff0c\u53e6\u4e00\u79cd\u6a21\u6001\u7684\u8bad\u7ec3\u6570\u636e\u6216\u6807\u7b7e\u53ef\u80fd\u662f\u4e0d\u53ef\u7528\u7684\u3002\u5728\u4e0d\u540c\u6a21\u6001\u4e4b\u95f4\u8f6c\u79fb\u77e5\u8bc6\u662f\u4e00\u4e2a\u6709\u4ef7\u503c\u7684\u5b9e\u8df5\u9886\u57df\u3002\u8de8\u6a21\u6001KD\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u77e5\u8bc6\u4ece\u53d7\u8fc7\u6570\u636e\u6a21\u6001\u8bad\u7ec3\u7684\u6559\u5e08\u8f6c\u79fb\u5230\u53e6\u4e00\u79cd\u6570\u636e\u6a21\u6001\u7684\u5b66\u751f\u7f51\u7edc\u3002\u8de8\u6a21\u6001KD\u7684\u793a\u610f\u56fe\u5982\u56fe5\u6240\u793a\u3002<\/p>\n<p><\/p>\n<p><\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-efc5b2dac71a53ae325164f98d320895.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-efc5b2dac71a53ae325164f98d320895.jpg\"><\/a><figcaption>\u56fe5 \u8de8\u6a21\u6001\u77e5\u8bc6\u84b8\u998f\u57fa\u672c\u601d\u60f3\u793a\u610f\u56fe<\/figcaption><\/figure>\n<p data-pid=\"YDKlGPQY\">\u7ed9\u5b9a\u4e00\u4e2a\u5728\u5177\u6709\u826f\u597d\u6807\u8bb0\u6837\u672c\u7684\u4e00\u79cd\u6a21\u6001\u4e0a\u9884\u5148\u8bad\u7ec3\u7684\u6559\u5e08\u6a21\u578b\uff0cGupta\u7b49\u4eba\u5229\u7528\u65e0\u76d1\u7763\u914d\u5bf9\u6837\u672c\u5728\u5e26\u6ce8\u91ca\u7684RGB\u56fe\u50cf\u548c\u672a\u5e26\u6ce8\u91ca\u7684\u5149\u6d41\u56fe\u50cf\u4e4b\u95f4\u4f20\u9012\u4fe1\u606f\u3002\u901a\u8fc7\u6807\u7b7e\u5f15\u5bfc\u7684\u6210\u5bf9\u6837\u672c\u8303\u5f0f\u5df2\u88ab\u5e7f\u6cdb\u5e94\u7528\u4e8e\u8de8\u6a21\u6001KD\u3002Thoker\u7b49\u4eba\u4f7f\u7528\u914d\u5bf9\u6837\u672c\u5c06RGB\u89c6\u9891\u4e2d\u7684\u77e5\u8bc6\u8f6c\u79fb\u52303D\u4eba\u4f53\u52a8\u4f5c\u8bc6\u522b\u6a21\u578b\u4e2d\u3002Roheda\u7b49\u4eba\u63d0\u51fa\u4e86\u4f7f\u7528GANs\u4ece\u53ef\u7528\u6a21\u6001\u5230\u7f3a\u5931\u6a21\u6001\u7684\u8de8\u6a21\u6001\u84b8\u998f\u3002Do\u7b49\u4eba\u63a2\u7d22\u4e86\u4e00\u79cd\u57fa\u4e8eKD\u7684\u89c6\u89c9\u95ee\u7b54\u65b9\u6cd5\uff0c\u5e76\u4f9d\u9760\u76d1\u7763\u5b66\u4e60\u4f7f\u7528\u771f\u5b9e\u6807\u7b7e\u8fdb\u884c\u8de8\u6a21\u6001\u8fc1\u79fb\u3002Passalis\u7b49\u4eba\u63d0\u51fa\u4e86\u6982\u7387KD\uff0c\u5c06\u77e5\u8bc6\u4ece\u6587\u672c\u6a21\u6001\u8f6c\u79fb\u5230\u89c6\u89c9\u6a21\u6001\u3002<\/p>\n<p data-pid=\"IQO2VhIR\"><strong>\u603b\u7ed3\u3002<\/strong>\u4e00\u822c\u6765\u8bf4\uff0cKD\u5728\u8de8\u6a21\u5f0f\u573a\u666f\u4e2d\u8868\u73b0\u826f\u597d\u3002\u7136\u800c\uff0c\u5f53\u5b58\u5728\u663e\u8457\u7684\u6a21\u6001\u7f3a\u5931\u65f6\uff0c\u8de8\u6a21\u6001KD\u5f88\u96be\u5bf9\u77e5\u8bc6\u4ea4\u4e92\u8fdb\u884c\u5efa\u6a21\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"33_Attention-based_Distillation%EF%BC%88%E5%9F%BA%E4%BA%8E%E6%B3%A8%E6%84%8F%E5%8A%9B%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>3.3 Attention-based Distillation\uff08\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"v3-5Muat\">\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u84b8\u998f\u5229\u7528\u6ce8\u610f\u529b\u4fe1\u606f\u8fdb\u884c\u6709\u6548\u7684\u77e5\u8bc6\u8f6c\u79fb\u3002\u76ee\u524d\u7684\u5de5\u4f5c\u9075\u5faa\u4e24\u6761\u4e3b\u7ebf\uff1a\uff081\uff09\u63d0\u53d6\u4ece\u7279\u5f81\u56fe\u4e2d\u63d0\u70bc\u51fa\u7684\u6ce8\u610f\u529b\u56fe\uff0c\u4ee5\u53ca\uff082\uff09\u57fa\u4e8e\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u52a0\u6743\u63d0\u53d6\uff0c\u5982\u4e0b\u56fe6\u6240\u793a\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-3495624d078a6079a02700ac09525bc6.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-3495624d078a6079a02700ac09525bc6.jpg\"><\/a><figcaption>\u56fe6 \u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236\u7684\u77e5\u8bc6\u84b8\u998f\u57fa\u672c\u601d\u60f3\u793a\u610f\u56fe<\/figcaption><\/figure>\n<p data-pid=\"DxvyR-QP\">3.3.1 Distilling attention maps\uff08\u84b8\u998f\u6ce8\u610f\u529b\u56fe\uff09<\/p>\n<p data-pid=\"e12PMp25\">\u6ce8\u610f\u529b\u56fe\u901a\u5e38\u53cd\u6620\u6709\u4ef7\u503c\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u5e76\u6291\u5236\u4e0d\u91cd\u8981\u7684\u90e8\u5206\u3002\u5f00\u521b\u6027\u7684AT\u901a\u8fc7\u8ba1\u7b97\u901a\u9053\u7ef4\u5ea6\u4e0a\u7279\u5f81\u56fe\u7684\u7edf\u8ba1\u6570\u636e\u6765\u6784\u5efa\u7a7a\u95f4\u6ce8\u610f\u529b\u56fe\uff0c\u5e76\u5728\u6559\u5e08\u548c\u5b66\u751f\u7f51\u7edc\u4e4b\u95f4\u5bf9\u6ce8\u610f\u529b\u56fe\u8fdb\u884c\u5bf9\u9f50\u3002\u7a7a\u95f4\u6ce8\u610f\u529b\u56fe\u5305\u542b\u7c7b\u522b\u611f\u77e5\u8bed\u4e49\u533a\u57df\uff0c\u5e2e\u52a9\u5b66\u751f\u6355\u6349\u8fa8\u522b\u7279\u5f81\u3002CD\u91c7\u7528\u538b\u7f29\u548c\u6fc0\u6d3b\u6a21\u5757\u751f\u6210\u901a\u9053\u6ce8\u610f\u529b\u56fe\uff0c\u8ba9\u5b66\u751f\u5b66\u4e60\u8001\u5e08\u7684\u901a\u9053\u6ce8\u610f\u529b\u6743\u91cd\u3002CWD\u4e3a\u6bcf\u4e2a\u901a\u9053\u63d0\u53d6\u4e00\u4e2a\u7a7a\u95f4\u6ce8\u610f\u529b\u56fe\uff0c\u8868\u793a\u5bc6\u96c6\u9884\u6d4b\u7684\u8bed\u4e49\u63a9\u7801\u3002TinyBert\u8f6c\u79fbTransformer\u5c42\u84b8\u998f\u7684\u81ea\u6ce8\u610f\u529b\u77e9\u9635\u3002LKD\u5f15\u5165\u4e86\u4e00\u4e2a\u7c7b\u522b\u611f\u77e5\u6ce8\u610f\u529b\u6a21\u5757\uff0c\u7528\u4e8e\u6355\u83b7\u7c7b\u76f8\u5173\u533a\u57df\u4ee5\u6784\u5efa\u5c40\u90e8\u76f8\u5173\u77e9\u9635\u3002<\/p>\n<p data-pid=\"IEtG7xji\">3.3.2 Self-attention-based weighted distillation\uff08\u81ea\u6ce8\u610f\u529b\u673a\u5236\u52a0\u6743\u84b8\u998f\uff09<\/p>\n<p data-pid=\"bigihrUL\">\u81ea\u6ce8\u610f\u529b\u6280\u672f\u662f\u4e00\u79cd\u6355\u83b7\u7279\u5f81\u4e4b\u95f4\u76f8\u4f3c\u5173\u7cfb\u7684\u673a\u5236\u3002\u4e00\u4e9b\u65b9\u6cd5\u5e94\u7528\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u6743\u91cd\u8fdb\u884c\u81ea\u9002\u5e94\u5c42\u95f4\u8bed\u4e49\u5339\u914d\u3002SemCKD\u81ea\u52a8\u4e3a\u6bcf\u4e2a\u5b66\u751f\u5c42\u5206\u914d\u4ece\u5408\u9002\u7684\u6559\u5e08\u5c42\u805a\u5408\u7684\u76ee\u6807\uff0c\u8fd9\u4e9b\u76ee\u6807\u5177\u6709\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u76f8\u4f3c\u6027\u3002AFD\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u5143\u7f51\u7edc\u6765\u6a21\u62df\u6559\u5e08\u548c\u5b66\u751f\u7279\u5f81\u4e4b\u95f4\u7684\u76f8\u5bf9\u76f8\u4f3c\u6027\u3002ALP-KD\u5c06\u6559\u5e08\u7aef\u7684\u4fe1\u606f\u4e0e\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u5c42\u6295\u5f71\u878d\u5408\uff0c\u7528\u4e8eBert\u84b8\u998f\u3002\u4e0e\u7279\u5f81\u5c42\u5206\u914d\u6b63\u4ea4\uff0cTTKD\u5e94\u7528\u81ea\u6ce8\u610f\u529b\u673a\u5236\u8fdb\u884c\u7a7a\u95f4\u7ea7\u7279\u5f81\u5339\u914d\u3002<\/p>\n<p data-pid=\"DPDMlSD0\"><strong>\u603b\u7ed3\u3002<\/strong>\u57fa\u4e8e\u6ce8\u610f\u529b\u56fe\u6355\u6349\u663e\u8457\u533a\u57df\u5e76\u8fc7\u6ee4\u5197\u4f59\u4fe1\u606f\uff0c\u5e2e\u52a9\u5b66\u751f\u5b66\u4e60\u6700\u5173\u952e\u7684\u7279\u5f81\u3002\u7136\u800c\uff0c\u6ce8\u610f\u529b\u56fe\u538b\u7f29\u4e86\u7279\u5f81\u56fe\u7684\u7ef4\u5ea6\uff0c\u53ef\u80fd\u4f1a\u4e22\u5931\u6709\u610f\u4e49\u7684\u77e5\u8bc6\u3002\u6b64\u5916\uff0c\u6ce8\u610f\u529b\u56fe\u6709\u65f6\u53ef\u80fd\u4e0d\u4f1a\u805a\u7126\u5728\u6b63\u786e\u7684\u533a\u57df\uff0c\u4ece\u800c\u5bfc\u81f4\u8d1f\u9762\u7684\u76d1\u7763\u5f71\u54cd\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"34_Data-free_Distillation%EF%BC%88%E6%97%A0%E9%9C%80%E6%95%B0%E6%8D%AE%E7%9A%84%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>3.4 Data-free Distillation\uff08\u65e0\u9700\u6570\u636e\u7684\u77e5\u8bc6\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"WDMNOATf\">\u4f20\u7edf\u7684KD\u65b9\u6cd5\u901a\u5e38\u9700\u8981\u5927\u91cf\u7684\u8bad\u7ec3\u6837\u672c\u3002\u7136\u800c\uff0c\u51fa\u4e8e\u9690\u79c1\u6216\u5b89\u5168\u8003\u8651\uff0c\u8bad\u7ec3\u6570\u636e\u96c6\u6709\u65f6\u53ef\u80fd\u662f\u4e0d\u53ef\u5f97\u5230\u7684\u3002\u73b0\u6709\u5de5\u4f5c\u5df2\u7ecf\u63d0\u51fa\u4e86\u4e00\u4e9b\u65b9\u6cd5\u6765\u5904\u7406\u8fd9\u4e2a\u95ee\u9898\uff0c\u4e3b\u8981\u5206\u4e3a\u65e0\u6570\u636eKD\u548c\u6570\u636e\u96c6KD\u3002\u65e0\u6570\u636eKD\u7684\u793a\u610f\u56fe\u5982\u4e0b\u56fe7\u6240\u793a\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8b8bc749b336c14f0bf095d5dc917f8e.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8b8bc749b336c14f0bf095d5dc917f8e.jpg\"><\/a><figcaption>\u56fe7 \u65e0\u9700\u6570\u636e\u7684\u77e5\u8bc6\u84b8\u998f\u57fa\u672c\u601d\u60f3\u793a\u610f\u56fe<\/figcaption><\/figure>\n<p data-pid=\"ZkPeoTuy\">3.4.1 Data-free KD \uff08\u65e0\u9700\u6570\u636e\u7684\u84b8\u998f\uff09<\/p>\n<p data-pid=\"07FezKDE\">\u8bad\u7ec3\u6837\u672c\u901a\u5e38\u662f\u4f7f\u7528\u751f\u6210\u5bf9\u6297\u7f51\u7edc\uff08GAN\uff09\u5408\u6210\u7684\u3002\u6559\u5e08\u7f51\u7edc\u4ee5\u751f\u6210\u7684\u6837\u672c\u4f5c\u4e3a\u8f93\u5165\u6765\u76d1\u7763\u5b66\u751f\u7f51\u7edc\u3002Lopes \u7b49\u4eba\u4f7f\u7528\u4e0d\u540c\u7c7b\u578b\u7684\u6fc0\u6d3b\u8bb0\u5f55\u6765\u91cd\u5efa\u539f\u59cb\u6570\u636e\u3002DeepInversion\u63a2\u7d22\u5b58\u50a8\u5728\u6279\u5f52\u4e00\u5316\u5c42\u4e2d\u7684\u4fe1\u606f\uff0c\u4ee5\u751f\u6210\u65e0\u6570\u636eKD\u7684\u6837\u672c\u3002Nayak\u7b49\u4eba\u901a\u8fc7\u5c06softmax\u7a7a\u95f4\u5efa\u6a21\u4e3a\u6559\u5e08\u53c2\u6570\u7684\u72c4\u5229\u514b\u96f7\u5206\u5e03\uff0c\u4ece\u800c\u5f15\u5165\u4e86\u4e00\u79cd\u6837\u672c\u63d0\u53d6\u673a\u5236\u3002\u9664\u4e86\u6700\u7ec8\u8f93\u51fa\u4e4b\u5916\uff0c\u8fd8\u53ef\u4ee5\u4f7f\u7528\u6559\u5e08\u7279\u5f81\u8868\u793a\u7684\u4fe1\u606f\u751f\u6210\u76ee\u6807\u6570\u636e\u3002Paul\u7b49\u4eba\u4f18\u5316\u4e86\u4e00\u4e2a\u5bf9\u6297\u751f\u6210\u5668\u6765\u641c\u7d22\u56f0\u96be\u7684\u56fe\u50cf\uff0c\u7136\u540e\u4f7f\u7528\u8fd9\u4e9b\u56fe\u50cf\u6765\u8bad\u7ec3\u5b66\u751f\u3002CMI\u5f15\u5165\u4e86\u5bf9\u6bd4\u5b66\u4e60\uff0c\u4f7f\u5408\u6210\u5b9e\u4f8b\u4e0e\u5df2\u7ecf\u5408\u6210\u7684\u5b9e\u4f8b\u533a\u5206\u5f00\u6765\u3002FastDFKD\u4f18\u5316\u4e86\u5143\u5408\u6210\u5668\uff0c\u4ee5\u91cd\u7528\u5171\u4eab\u7684\u901a\u7528\u529f\u80fd\uff0c\u5b9e\u73b0\u66f4\u5feb\u7684\u65e0\u6570\u636eKD\u3002<\/p>\n<p data-pid=\"k9LCPoet\">3.4.2 Dataset KD\uff08\u6570\u636e\u96c6\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"bL1hjm3R\">\u9664\u4e86\u65e0\u6570\u636eKD\u4e4b\u5916\uff0c\u6570\u636e\u96c6\u84b8\u998f\u662f\u5408\u6210\u5c0f\u6570\u636e\u96c6\u4ee5\u8868\u793a\u539f\u59cb\u5b8c\u6574\u6570\u636e\u96c6\u800c\u4e0d\u4f1a\u964d\u4f4e\u7cbe\u5ea6\u7684\u91cd\u8981\u65b9\u5411\u3002\u4e3a\u4e86\u5229\u7528\u5168\u76d1\u7763\u8bbe\u7f6e\uff0cRadosavovic\u7b49\u4eba\u901a\u8fc7\u5355\u4e2a\u6a21\u578b\u4ece\u672a\u6807\u8bb0\u6570\u636e\u7684\u591a\u6b21\u8f6c\u6362\u4e2d\u7ec4\u88c5\u9884\u6d4b\uff0c\u4ee5\u4ea7\u751f\u65b0\u7684\u8bad\u7ec3\u6ce8\u91ca\u3002DDFlow\u63d0\u51fa\u5b66\u4e60\u5149\u6d41\u4f30\u8ba1\uff0c\u5e76\u4ece\u6559\u5e08\u6a21\u578b\u4e2d\u63d0\u53d6\u9884\u6d4b\uff0c\u4ee5\u76d1\u7763\u5b66\u751f\u7f51\u7edc\u8fdb\u884c\u5149\u6d41\u5b66\u4e60\u3002\u672a\u6807\u8bb0\u7684\u6570\u636e\u53ef\u80fd\u4f1a\u963b\u788d\u56fe\u5377\u79ef\u7f51\u7edc\uff08GCN\uff09\u5b66\u4e60\u57fa\u4e8e\u56fe\u7684\u6570\u636e\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u672a\u6807\u8bb0\u6570\u636e\u7684\u4f2a\u6807\u7b7e\u53ef\u4ee5\u4e3a\u8bad\u7ec3GCN\u63d0\u4f9b\u989d\u5916\u7684\u76d1\u7763\u3002RDD\u63d0\u51fa\u4e86\u4e00\u79cd\u53ef\u9760\u7684\u6570\u636e\u9a71\u52a8\u7684\u534a\u76d1\u7763GCN\u8bad\u7ec3\u65b9\u6cd5\uff0c\u5b83\u53ef\u4ee5\u66f4\u597d\u5730\u4f7f\u7528\u9ad8\u8d28\u91cf\u7684\u6570\u636e\uff0c\u5e76\u901a\u8fc7\u5b9a\u4e49\u8282\u70b9\u548c\u8fb9\u7684\u53ef\u9760\u6027\u6765\u6539\u8fdb\u56fe\u8868\u793a\u5b66\u4e60\u3002Cazenavette\u7b49\u4eba\u6cbf\u7740\u63d0\u53d6\u7684\u5408\u6210\u6570\u636e\u548c\u771f\u5b9e\u6570\u636e\u4e4b\u95f4\u7684\u8bad\u7ec3\u8f68\u8ff9\u8fdb\u884c\u4e86\u957f\u8ddd\u79bb\u53c2\u6570\u5339\u914d\u3002<\/p>\n<p data-pid=\"t4mNndcd\"><strong>\u603b\u7ed3\u3002<\/strong>\u5728\u5927\u591a\u6570\u65e0\u6570\u636eKD\u65b9\u6cd5\u4e2d\uff0c\u5408\u6210\u6570\u636e\u901a\u5e38\u662f\u4ece\u9884\u8bad\u7ec3\u7684\u6559\u5e08\u7f51\u7edc\u7684\u7279\u5f81\u8868\u793a\u4e2d\u751f\u6210\u7684\u3002\u5c3d\u7ba1\u5f53\u524d\u7684\u65e0\u6570\u636eKD\u5de5\u4f5c\u5728\u5904\u7406\u6570\u636e\u4e0d\u53ef\u7528\u95ee\u9898\u65b9\u9762\u8868\u73b0\u51fa\u4e86\u663e\u8457\u7684\u6027\u80fd\uff0c\u4f46\u751f\u6210\u66f4\u9ad8\u8d28\u91cf\u548c\u591a\u6837\u5316\u7684\u8bad\u7ec3\u6837\u672c\u4ecd\u7136\u662f\u4e00\u4e2a\u6709\u5f85\u7814\u7a76\u7684\u6311\u6218\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"35_Adversarial_Distillation_%EF%BC%88%E5%AF%B9%E6%8A%97%E8%92%B8%E9%A6%8F%EF%BC%89\"><\/span>3.5 Adversarial Distillation \uff08\u5bf9\u6297\u84b8\u998f\uff09<span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p data-pid=\"DN4mjo3Z\">\u5bf9\u6297\u84b8\u998f\u662f\u5229\u7528\u751f\u6210\u5bf9\u6297\u7f51\u7edc\u7684\u57fa\u672c\u601d\u60f3\u6765\u6539\u8fdbKD\uff0c\u4e3b\u8981\u5206\u4e3a\u4e09\u6761\u4e3b\u7ebf\uff1a\uff081\uff09\u4f7f\u7528GAN\u751f\u6210\u989d\u5916\u7684\u6570\u636e\u6837\u672c\uff0c\uff082\uff09\u5bf9\u6297\u673a\u5236\u6765\u8f85\u52a9KD\u7b97\u6cd5\uff0c\uff083\uff09\u538b\u7f29GAN\u4ee5\u9ad8\u6548\u751f\u6210\u56fe\u50cf\u3002\u5bf9\u6297\u6027KD\u7684\u793a\u610f\u56fe\u5982\u4e0b\u56fe8\u6240\u793a\u3002<\/p>\n<figure data-size=\"normal\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8f6a5bed87b484919ea7162ec7ea06ad.jpg\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8f6a5bed87b484919ea7162ec7ea06ad.jpg\"><\/a><figcaption>\u56fe8 \u5bf9\u6297\u77e5\u8bc6\u84b8\u998f\u57fa\u672c\u601d\u60f3\u793a\u610f\u56fe<\/figcaption><\/figure>\n<p data-pid=\"fSCY9_uX\">3.5.1 Using GAN to generate extra data samples\uff08\u4f7f\u7528\u5bf9\u6297\u751f\u6210\u7f51\u7edc\u4ea7\u751f\u989d\u5916\u7684\u6570\u636e\u6837\u672c\uff09<\/p>\n<p data-pid=\"8Om5YFRa\">\u51e0\u4e4e\u73b0\u6709\u7684KD\u7b97\u6cd5\u90fd\u662f\u6570\u636e\u9a71\u52a8\u7684\uff0c\u5373\u4f9d\u8d56\u4e8e\u539f\u59cb\u6216\u66ff\u4ee3\u6570\u636e\uff0c\u8fd9\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u53ef\u80fd\u662f\u4e0d\u53ef\u7528\u7684\u3002\u751f\u6210\u5bf9\u6297\u7f51\u7edc\u53ef\u4ee5\u5e94\u7528\u4e8e\u5b66\u4e60\u771f\u5b9e\u7684\u6570\u636e\u5206\u5e03\u5e76\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002DFAD\u8ba9\u6559\u5e08\u548c\u5b66\u751f\u7f51\u7edc\u5171\u540c\u53d1\u6325\u9274\u522b\u5668\u7684\u4f5c\u7528\uff0c\u4ee5\u51cf\u5c11\u5dee\u5f02\u3002\u540c\u65f6\uff0c\u5b83\u6dfb\u52a0\u4e86\u4e00\u4e2a\u989d\u5916\u7684\u751f\u6210\u5668\u6765\u4ea7\u751f\u56f0\u96be\u6837\u672c\uff0c\u4ee5\u5bf9\u6297\u6027\u5730\u653e\u5927\u5b83\u3002\u4e00\u4e9b\u65b9\u6cd5\u5f15\u5165\u4e86\u6761\u4ef6GAN\uff08CGAN\uff09\u6765\u751f\u6210\u6570\u636e\u3002Roheda\u7b49\u4eba\u4f7f\u7528CGAN\u5728\u7ed9\u5b9a\u5176\u4ed6\u53ef\u7528\u6a21\u6001\u7684\u524d\u63d0\u4e0b\u4ece\u7f3a\u5931\u6a21\u6001\u4e2d\u84b8\u998f\u77e5\u8bc6\u3002\u7ec8\u8eabGAN\u5c06\u4ee5\u524d\u7f51\u7edc\u4e2d\u5b66\u4e60\u5230\u7684\u77e5\u8bc6\u8f6c\u79fb\u5230\u65b0\u7f51\u7edc\u4e2d\uff0c\u4ee5\u6301\u7eed\u751f\u6210\u6709\u6761\u4ef6\u7684\u56fe\u50cf\u3002<\/p>\n<p data-pid=\"emyw7wMz\">3.5.2 Adversarial mechanism to assist general KD\uff08\u5bf9\u6297\u673a\u5236\u6765\u8f85\u52a9\u77e5\u8bc6\u84b8\u998f\uff09<\/p>\n<p data-pid=\"K_axT6bY\">\u4f20\u7edf\u7684KD\u901a\u5e38\u901a\u8fc7\u8c03\u6574\u77e5\u8bc6\u5206\u5e03\u6765\u7f29\u5c0f\u6559\u5e08\u548c\u5b66\u751f\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002\u5bf9\u6297\u673a\u5236\u53ef\u4ee5\u4f5c\u4e3a\u63d0\u9ad8\u6a21\u4eff\u96be\u5ea6\u7684\u8f85\u52a9\u65b9\u6cd5\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u6838\u5fc3\u601d\u60f3\u662f\u5f15\u5165\u4e00\u4e2a\u989d\u5916\u7684\u9274\u522b\u5668\u6765\u5bf9\u6559\u5e08\u6216\u5b66\u751f\u7f51\u7edc\u4e2d\u7684\u7279\u5f81\u8868\u793a\u8fdb\u884c\u5206\u7c7b\u3002Wang\u7b49\u4eba\u5229\u7528\u9274\u522b\u5668\u4f5c\u4e3a\u6559\u5b66\u52a9\u624b\uff0c\u4f7f\u5b66\u751f\u4e0e\u6559\u5e08\u5b66\u4e60\u76f8\u4f3c\u7684\u7279\u5f81\u5206\u5e03\uff0c\u4ee5\u8fdb\u884c\u56fe\u50cf\u5206\u7c7b\u3002Wang\u7b49\u4eba\u91c7\u7528\u5bf9\u6297\u6027KD\u8fdb\u884c\u5355\u9636\u6bb5\u76ee\u6807\u68c0\u6d4b\u3002Liu\u7b49\u4eba\u5c06\u9010\u50cf\u7d20\u7684\u7c7b\u6982\u7387\u4ee5\u5bf9\u6297\u7684\u65b9\u5f0f\u5e94\u7528\u4e8e\u8bed\u4e49\u5206\u5272\u84b8\u998f\u4e2d\u3002\u9664\u4e86\u57fa\u4e8e\u5e08\u751f\u7684\u5bf9\u6297\u5b66\u4e60\u5916\uff0c\u4e00\u4e9b\u65b9\u6cd5\u8fd8\u5e94\u7528\u4e86\u5728\u7ebf\u5bf9\u6297KD\u5728\u591a\u4e2a\u5b66\u751f\u7f51\u7edc\u4e4b\u95f4\u76f8\u4e92\u63d0\u53d6\u7279\u5f81\u56fe\u3002<\/p>\n<p data-pid=\"xS76A6XR\">3.5.3 Compressing GAN for efficient image generation\uff08\u538b\u7f29\u5bf9\u6297\u751f\u6210\u7f51\u7edc\u6765\u8fdb\u884c\u9ad8\u6548\u7684\u56fe\u50cf\u751f\u6210\uff09<\/p>\n<p data-pid=\"-TEOJKcU\">Aguinaldo\u7b49\u4eba\u5f15\u5bfc\u4e00\u4e2a\u8f83\u5c0f\u7684\u201c\u5b66\u751f\u201dGAN\u5bf9\u9f50\u66f4\u5927\u7684\u201c\u6559\u5e08\u201dGAN\u3002Chen\u7b49\u4eba\u8ba9\u5b66\u751f\u751f\u6210\u5668\u4ece\u76f8\u5e94\u7684\u8001\u5e08\u90a3\u91cc\u5b66\u4e60\u4f4e\u9636\u548c\u9ad8\u9636\u77e5\u8bc6\u3002\u6b64\u5916\uff0c\u5b66\u751f\u9274\u522b\u5668\u7531\u6559\u5e08\u7f51\u7edc\u901a\u8fc7triplet\u635f\u5931\u8fdb\u884c\u76d1\u7763\u3002Li\u7b49\u4eba\u901a\u8fc7\u4ece\u4e2d\u95f4\u8868\u793a\u4e2d\u8f6c\u79fb\u77e5\u8bc6\u5e76\u901a\u8fc7\u795e\u7ecf\u67b6\u6784\u641c\u7d22\u63a2\u7d22\u9ad8\u6548\u67b6\u6784\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u7528\u4e8e\u6761\u4ef6GAN\u7684\u901a\u7528\u538b\u7f29\u6846\u67b6\u3002Zhang\u7b49\u4eba\u6307\u51fa\uff0c\u5c0f\u578bGAN\u901a\u5e38\u96be\u4ee5\u751f\u6210\u6240\u9700\u7684\u9ad8\u9891\u4fe1\u606f\u3002\u5c0f\u6ce2KD\u901a\u8fc7\u79bb\u6563\u5c0f\u6ce2\u53d8\u6362\u5c06\u56fe\u50cf\u5206\u89e3\u4e3a\u4e0d\u540c\u7684\u9891\u5e26\uff0c\u7136\u540e\u53ea\u4f20\u8f93\u6709\u4ef7\u503c\u7684\u9ad8\u9891\u5e26\u3002<\/p>\n<p data-pid=\"9taYG9-6\"><strong>\u603b\u7ed3\u3002<\/strong>\u5c3d\u7ba1\u57fa\u4e8e\u5bf9\u6297\u7684KD\u6709\u52a9\u4e8e\u77e5\u8bc6\u6a21\u4eff\uff0c\u4f46\u5728\u5b9e\u8df5\u4e2d\u53ef\u80fd\u5f88\u96be\u786e\u4fddGAN\u7f51\u7edc\u7684\u6536\u655b\u6027\u3002\u5bf9\u4e8eGAN\u538b\u7f29\uff0c\u4ece\u7279\u5f81\u4e2d\u63d0\u53d6\u54ea\u4e9b\u4fe1\u606f\u9002\u5408\u84b8\u998fGAN\u4ecd\u7136\u662f\u4e00\u4e2a\u503c\u5f97\u63a2\u7d22\u7684\u95ee\u9898\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"4%E6%80%BB%E7%BB%93\"><\/span>4.\u603b\u7ed3<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p data-pid=\"SsYle2r8\">\u672c\u7bc7\u7efc\u8ff0\u9996\u5148\u6309\u7167\u57fa\u4e8e\u54cd\u5e94\u3001\u7279\u5f81\u548c\u5173\u7cfb\u4e09\u79cd\u5c42\u9762\u7684\u77e5\u8bc6\u7c7b\u578b\u603b\u7ed3\u4e86\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\u7684\u5de5\u4f5c\u3002\u901a\u8fc7\u5256\u6790\u79bb\u7ebf\u77e5\u8bc6\u84b8\u998f\u9700\u8981\u9884\u8bad\u7ec3\u6559\u5e08\u7684\u95ee\u9898\uff0c\u5f15\u51fa\u4e86\u5728\u7ebf\u77e5\u8bc6\u84b8\u998f\u548c\u81ea\u77e5\u8bc6\u84b8\u998f\u3002\u9762\u5411\u5b9e\u9645\u573a\u666f\u4e2d\uff0c\u672c\u6587\u53c8\u5f15\u5165\u4e86\u8de8\u6a21\u6001\u77e5\u8bc6\u84b8\u998f\u548c\u65e0\u6570\u636e\u7684\u77e5\u8bc6\u84b8\u998f\u3002\u6b64\u5916\uff0c\u672c\u6587\u5c55\u793a\u4e86\u4e00\u4e9b\u7528\u4e8e\u63d0\u5347\u77e5\u8bc6\u84b8\u998f\u6548\u679c\u7684\u6269\u5c55\u673a\u5236\uff0c\u4f8b\u5982\u591a\u6559\u5e08\u77e5\u8bc6\u84b8\u998f\u3001\u6ce8\u610f\u529b\u673a\u5236\u77e5\u8bc6\u84b8\u998f\u548c\u5bf9\u6297\u77e5\u8bc6\u84b8\u998f\u3002<\/p>\n<p data-pid=\"pfLrkAn3\">\u672c\u6587\u5bf9\u4ee5\u4e0a\u9886\u57df\u7684\u4ee3\u8868\u6027\u76f8\u5173\u5de5\u4f5c\u8fdb\u884c\u8c03\u7814\uff0c\u5e76\u603b\u7ed3\u6838\u5fc3\u601d\u60f3\u4e0e\u8d21\u732e\uff0c\u5bf9\u672a\u6765\u77e5\u8bc6\u84b8\u998f\u7684\u6311\u6218\u8fdb\u884c\u5c55\u671b\u3002\u76f8\u6bd4\u4e8e\u5148\u524d\u53d1\u5e03\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0[2,3]\uff0c\u672c\u6587\u5305\u542b\u4e86\u66f4\u591a\u53d1\u8868\u4e8e2022\u5e74\u4ee5\u540e\u7684\u5de5\u4f5c\uff0c\u4ee5\u53ca\u4ecb\u7ecd\u4e86\u4e00\u4e9b\u5148\u8fdb\u7684\u77e5\u8bc6\u84b8\u998f\u65b9\u5411\uff0c\u4f8b\u5982\u81ea\u76d1\u7763\u84b8\u998f\u548cViT\u84b8\u998f\u3002\u672c\u6587\u5e0c\u671b\u53ef\u4ee5\u901a\u8fc7\u603b\u7ed3\u8fc7\u53bb\u5de5\u4f5c\u8ba9\u8bfb\u8005\u66f4\u5feb\u5730\u5b66\u4e60\u9886\u57df\u7684\u73b0\u72b6\uff0c\u4ece\u800c\u63d0\u51fa\u66f4\u5148\u8fdb\u7684\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\uff0c\u63a8\u52a8\u9886\u57df\u7684\u53d1\u5c55\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"%E5%8F%82%E8%80%83%E6%96%87%E7%8C%AE\"><\/span>\u53c2\u8003\u6587\u732e<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p data-pid=\"Bwif_Vfj\">[1] Hinton G. Distilling the Knowledge in a Neural Network[J]. arXiv preprint arXiv:1503.02531, 2015.<\/p>\n<p data-pid=\"roNk0D5z\">[2] Gou J, Yu B, Maybank S J, et al. Knowledge distillation: A survey[J]. International Journal of Computer Vision, 2021, 129: 1789-1819.<\/p>\n<p data-pid=\"i-PNZaKf\">[3] Wang L, Yoon K J. Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks[J]. IEEE transactions on pattern analysis and machine intelligence, 2021, 44(6): 3048-3068.<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-02-19-12\" target=\"_blank\">Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:Springer\u77e5\u8bc6\u84b8 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","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":"","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":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","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":[27],"tags":[65,67],"class_list":["post-38631","post","type-post","status-publish","format-standard","hentry","category-news","tag-65","tag-67"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=38631\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=38631\",\"url\":\"https:\/\/17aitech.com\/?p=38631\",\"name\":\"Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=38631#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=38631#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c807a03a5f3d24b73ad14f2a22d7e2d3.jpg\",\"datePublished\":\"2025-02-19T15:00:56+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=38631#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=38631\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=38631#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c807a03a5f3d24b73ad14f2a22d7e2d3.jpg\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c807a03a5f3d24b73ad14f2a22d7e2d3.jpg\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=38631#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=38631","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=38631","url":"https:\/\/17aitech.com\/?p=38631","name":"Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=38631#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=38631#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c807a03a5f3d24b73ad14f2a22d7e2d3.jpg","datePublished":"2025-02-19T15:00:56+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=38631#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=38631"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=38631#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c807a03a5f3d24b73ad14f2a22d7e2d3.jpg","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c807a03a5f3d24b73ad14f2a22d7e2d3.jpg"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=38631#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"Springer\u77e5\u8bc6\u84b8\u998f\u4e13\u8457\u89e3\u8bfb | \u9762\u5411\u56fe\u50cf\u8bc6\u522b\u7684\u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/38631","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=38631"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/38631\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=38631"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=38631"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=38631"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}