{"id":1069932,"date":"2022-11-09T22:57:00","date_gmt":"2022-11-10T06:57:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&#038;p=1069932"},"modified":"2024-09-25T01:42:00","modified_gmt":"2024-09-25T08:42:00","slug":"neurips-2022-the-big-convergence-of-ai-theory","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/neurips-2022-the-big-convergence-of-ai-theory\/","title":{"rendered":"NeurIPS 2022 | \u4e00\u7aa5\u4eba\u5de5\u667a\u80fd\u5927\u4e00\u7edf\u4e0e\u7406\u8bba\u7814\u7a76\u7684\u6700\u65b0\u8fdb\u5c55"},"content":{"rendered":"\n<p>\u7f16\u8005\u6309\uff1a\u4f5c\u4e3a\u76ee\u524d\u5168\u7403\u6700\u8d1f\u76db\u540d\u7684\u4eba\u5de5\u667a\u80fd\u76db\u4f1a\u4e4b\u4e00\uff0cNeurIPS (Conference on Neural Information Processing Systems) \u5728\u6bcf\u5e74\u5e74\u672b\u90fd\u662f\u8ba1\u7b97\u673a\u79d1\u5b66\u9886\u57df\u77a9\u76ee\u7684\u7126\u70b9\u3002\u88ab NeurIPS \u63a5\u6536\u7684\u8bba\u6587\uff0c\u4ee3\u8868\u7740\u5f53\u4eca\u795e\u7ecf\u79d1\u5b66\u548c\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u7684\u6700\u9ad8\u6c34\u5e73\u3002\u4eca\u5e74\u7684 NeurIPS \u5927\u4f1a\u5c06\u4e8e11\u670828\u65e5\u81f312\u67089\u65e5\u4e3e\u884c\uff0c\u672c\u5c4a\u5927\u4f1a\u5171\u6536\u523010411\u7bc7\u6709\u6548\u6295\u7a3f\uff0c\u5176\u4e2d2672\u7bc7\u83b7\u63a5\u6536\uff0c\u6700\u7ec8\u63a5\u6536\u7387\u4e3a25.6%\u3002\u76f8\u6bd4\u53bb\u5e74\uff0c\u6295\u7a3f\u6570\u91cf\u7ee7\u7eed\u589e\u52a0\u3002<\/p>\n\n\n\n<p>\u5728\u672c\u5c4a\u5927\u4f1a\u4e2d\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e5f\u6709\u8bf8\u591a\u8bba\u6587\u5165\u9009\uff0c\u5185\u5bb9\u4e3b\u8981\u6db5\u76d6\u4eba\u5de5\u667a\u80fd\u4e94\u5927\u70ed\u70b9\u8bdd\u9898\uff1a\u4eba\u5de5\u667a\u80fd\u8d70\u5411\u5927\u4e00\u7edf\u3001\u8ba1\u7b97\u673a\u7406\u8bba\u3001\u8d4b\u80fd\u4ea7\u4e1a\u754c\u7684\u4eba\u5de5\u667a\u80fd\u3001\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u3001\u4eba\u5de5\u667a\u80fd\u8d4b\u80fd\u5185\u5bb9\u4e0e\u8bbe\u8ba1\u751f\u6210\u3002\u5728\u63a5\u4e0b\u6765\u7684\u51e0\u5468\u91cc\uff0c\u6211\u4eec\u5c06\u6309\u8bdd\u9898\u4e0e\u5927\u5bb6\u5206\u4eab\u76f8\u5173\u9886\u57df\u7684\u5b66\u672f\u6700\u524d\u6cbf\uff01\u4eca\u5929\uff0c\u8ba9\u6211\u4eec\u5148\u4ece\u201c\u4eba\u5de5\u667a\u80fd\u8d70\u5411\u5927\u4e00\u7edf\u201d\u548c\u201c\u8ba1\u7b97\u673a\u7406\u8bba\u201d\u8bdd\u9898\u4e0b\u76848\u7bc7\u8bba\u6587\u7cbe\u534e\u5f00\u59cb\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u4eba\u5de5\u667a\u80fd\u8d70\u5411\u5927\u4e00\u7edf\">\u4eba\u5de5\u667a\u80fd\u8d70\u5411\u5927\u4e00\u7edf<\/h2>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"the-big-convergence-of-ai\">The Big Convergence of AI<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u9488\u5bf9\u5f3a\u5316\u5b66\u4e60\u7684\u63a9\u7801\u9690\u7a7a\u95f4\u91cd\u5efa\">\u9488\u5bf9\u5f3a\u5316\u5b66\u4e60\u7684\u63a9\u7801\u9690\u7a7a\u95f4\u91cd\u5efa<\/h3>\n\n\n\n<h3 id=\"\" class=\"wp-block-heading\"><\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"690\" height=\"163\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-1.png\" alt=\"paper screenshot\" class=\"wp-image-1070313\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-1.png 690w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-1-300x71.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-1-240x57.png 240w\" sizes=\"auto, (max-width: 690px) 100vw, 690px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/mask-based-latent-reconstruction-for-reinforcement-learning\/<\/p>\n\n\n\n<p>\u89c6\u89c9\u72b6\u6001\u8868\u5f81\u7684\u8d28\u91cf\u5bf9\u57fa\u4e8e\u89c6\u89c9\u7684\u5f3a\u5316\u5b66\u4e60\uff08vision-based reinforcement learning\uff09\u81f3\u5173\u91cd\u8981\u3002\u4e3a\u4e86\u5b66\u4e60\u9ad8\u6548\u7684\u72b6\u6001\u8868\u5f81\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u521b\u65b0\u6027\u5730\u5c06\u57fa\u4e8e\u63a9\u7801\u7684\u5efa\u6a21\u6280\u672f\uff08mask-based modeling\uff09\u5e94\u7528\u5230\u5f3a\u5316\u5b66\u4e60\u4e2d\uff0c\u4ee5\u4fc3\u8fdb\u5176\u72b6\u6001\u8868\u5f81\u5b66\u4e60\u3002\u6b64\u524d\u57fa\u4e8e\u63a9\u7801\u7684\u5efa\u6a21\u6280\u672f\u5df2\u7ecf\u5728 CV \u548c NLP \u9886\u57df\u4e2d\u5927\u653e\u5f02\u5f69\uff0c\u800c\u8fd9\u9879\u5de5\u4f5c\u662f\u5c06\u5176\u5e94\u7528\u5230\u5f3a\u5316\u5b66\u4e60\u9886\u57df\u5e2e\u52a9\u7b56\u7565\u5b66\u4e60\u7684\u9996\u6b21\u63a2\u7d22\u3002<\/p>\n\n\n\n<p>\u5177\u4f53\u5730\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u800c\u6709\u6548\u7684\u81ea\u76d1\u7763\u65b9\u6cd5\uff0c\u5373\u57fa\u4e8e\u63a9\u7801\u7684\u9690\u7a7a\u95f4\u91cd\u5efa (mask-based latent reconstruction\uff0c\u7b80\u79f0\u4e3a MLR)\u3002MLR \u901a\u8fc7\u4ece\u5177\u6709\u65f6\u7a7a\u63a9\u7801\u7684\u89c6\u89c9\u72b6\u6001\u4e2d\u9884\u6d4b\u5176\u5728\u9690\u7a7a\u95f4\u4e2d\u7684\u5b8c\u6574\u8868\u5f81\uff0c\u4ece\u800c\u4f7f\u795e\u7ecf\u7f51\u7edc\u5728\u5b66\u4e60\u72b6\u6001\u8868\u5f81\u65f6\u80fd\u591f\u66f4\u597d\u5730\u5229\u7528\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u7f16\u7801\u66f4\u591a\u7b56\u7565\u5b66\u4e60\u6240\u9700\u8981\u7684\u8bed\u4e49\u4fe1\u606f\u3002\u5927\u91cf\u57fa\u51c6\u5b9e\u9a8c\u8868\u660e\uff0cMLR \u663e\u8457\u63d0\u9ad8\u4e86\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u7684\u6837\u672c\u6548\u7387\uff08sample efficiency\uff09\uff0c\u5728\u591a\u4e2a\u8fde\u7eed\u548c\u79bb\u6563\u7684\u5f3a\u5316\u5b66\u4e60\u73af\u5883\u4e2d\u53d6\u5f97\u4e86 SOTA \u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"882\" height=\"1024\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-2-882x1024.png\" alt=\"\u56fe1\uff1a\u57fa\u4e8e\u63a9\u7801\u7684\u9690\u7a7a\u95f4\u91cd\u5efa\uff08MLR\uff09\u7684\u6846\u67b6\u793a\u610f\u56fe\" class=\"wp-image-1070316\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-2-882x1024.png 882w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-2-258x300.png 258w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-2-768x891.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-2-155x180.png 155w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-2.png 1058w\" sizes=\"auto, (max-width: 882px) 100vw, 882px\" \/><figcaption class=\"wp-element-caption\"><em>\u56fe1\uff1a\u57fa\u4e8e\u63a9\u7801\u7684\u9690\u7a7a\u95f4\u91cd\u5efa\uff08MLR\uff09\u7684\u6846\u67b6\u793a\u610f\u56fe<\/em><\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u57fa\u4e8e\u6ed1\u52a8\u8bed\u8a00\u6a21\u578b\u7684\u53e5\u5b50\u8bc4\u5206\u8f6c\u6362\u5668\">\u57fa\u4e8e\u6ed1\u52a8\u8bed\u8a00\u6a21\u578b\u7684\u53e5\u5b50\u8bc4\u5206\u8f6c\u6362\u5668<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"685\" height=\"159\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-3.png\" alt=\"paper screenshot\" class=\"wp-image-1070319\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-3.png 685w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-3-300x70.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-3-240x56.png 240w\" sizes=\"auto, (max-width: 685px) 100vw, 685px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/transcormer-transformer-for-sentence-scoring-with-sliding-language-modeling\/<\/p>\n\n\n\n<p>\u53e5\u5b50\u8bc4\u5206\u65e8\u5728\u8bc4\u4f30\u4e00\u4e2a\u53e5\u5b50\u7684\u6700\u5927\u4f3c\u7136\u4f30\u8ba1\uff0c\u88ab\u5e7f\u6cdb\u5e94\u7528\u4e8e\u8bb8\u591a\u81ea\u7136\u8bed\u8a00\u4efb\u52a1\u7684\u573a\u666f\u4e2d\uff0c\u5305\u62ec\u91cd\u6392\u5e8f\u3001\u8bed\u8a00\u53ef\u63a5\u53d7\u6027\u7b49\u3002\u8fc7\u53bb\u7528\u4e8e\u89e3\u51b3\u53e5\u5b50\u8bc4\u5206\u7684\u5de5\u4f5c\u4e3b\u8981\u4ee5\u4e24\u79cd\u7ecf\u5178\u8bed\u8a00\u6a21\u578b\u4e3a\u4e3b\uff1a\u56e0\u679c\u8bed\u8a00\u6a21\u578b\uff08causal language modeling, CLM\uff09\u548c\u63a9\u7801\u8bed\u8a00\u6a21\u578b\uff08masked language modeling, MLM\uff09\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u5de5\u4f5c\u90fd\u5b58\u5728\u4e00\u5b9a\u7684\u74f6\u9888\uff1aCLM \u867d\u7136\u53ea\u9700\u8981\u8ba1\u7b97\u4e00\u6b21\u4f46\u5374\u53ea\u5229\u7528\u4e86\u5355\u5411\u4fe1\u606f\uff1bMLM \u80fd\u591f\u5229\u7528\u53cc\u5411\u8bed\u4e49\uff0c\u4f46\u6bcf\u6b21\u53ea\u80fd\u9884\u6d4b\u90e8\u5206\u5355\u8bcd\u800c\u4e0d\u5f97\u4e0d\u9700\u8981\u591a\u6b21\u63a8\u7406\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u6ed1\u52a8\u8bed\u8a00\u6a21\u578b\u7684 Transformer \u6a21\u578b Transcormer\uff0c\u5e76\u5728\u5176\u4e2d\u8bbe\u8ba1\u4e86\u4e00\u79cd\u4e09\u6d41\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7528\u4e8e\u7ef4\u62a4\u6ed1\u52a8\u8bed\u8a00\u6a21\u578b\u3002\u5229\u7528\u8fd9\u6837\u7684\u8bbe\u8ba1\uff0cTranscormer \u53ef\u4ee5\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u5229\u7528\u53cc\u5411\u4fe1\u606f\u8fdb\u884c\u9884\u6d4b\u7684\u540c\u65f6\uff0c\u53ea\u9700\u4e00\u6b21\u8ba1\u7b97\u5373\u53ef\u5f97\u5230\u6240\u6709\u5355\u8bcd\u7684\u6982\u7387\u3002\u6ed1\u52a8\u8bed\u8a00\u6a21\u578b\u5728\u8ba1\u7b97\u53e5\u5b50\u8bc4\u5206\u65f6\uff0cTranscormer \u8fd8\u53ef\u4ee5\u907f\u514d CLM \u53ea\u80fd\u5229\u7528\u5355\u5411\u4fe1\u606f\u7684\u7f3a\u70b9\u4ee5\u53ca MLM \u9700\u8981\u591a\u6b21\u8ba1\u7b97\u7684\u4e0d\u8db3\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cTranscormer \u5728\u53e5\u5b50\u8bc4\u5206\u4efb\u52a1\u4e0a\u80fd\u591f\u53d6\u5f97\u6bd4\u5176\u4ed6\u65b9\u6cd5\u66f4\u597d\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"800\" height=\"394\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-4.png\" alt=\"\u56fe2\uff1aTranscormer \u7ed3\u6784\u793a\u610f\u56fe\uff1a\u5de6\u4fa7\u4e3a\u524d\u5411\u6d41\uff0c\u53f3\u4fa7\u4e3a\u540e\u5411\u6d41\uff0c\u4e2d\u95f4\u4e3a\u8be2\u95ee\u6d41\u3002\u5176\u4e2d\uff0c\u524d\u5411\u6d41\u7528\u4e8e\u6536\u96c6\u524d\u5411\u8bed\u4e49\uff0c\u540e\u5411\u6d41\u7528\u4e8e\u6536\u96c6\u540e\u5411\u8bed\u4e49\uff0c\u800c\u8be2\u95ee\u6d41\u7528\u4e8e\u6355\u83b7\u5f53\u524d\u4f4d\u7f6e\u5728\u5176\u4e4b\u524d\u7684\u524d\u5411\u6d41\u8bed\u4e49\u548c\u5176\u4e4b\u540e\u7684\u540e\u5411\u6d41\u8bed\u4e49\" class=\"wp-image-1070322\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-4.png 800w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-4-300x148.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-4-768x378.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-4-240x118.png 240w\" sizes=\"auto, (max-width: 800px) 100vw, 800px\" \/><figcaption class=\"wp-element-caption\"><em>\u56fe2\uff1aTranscormer \u7ed3\u6784\u793a\u610f\u56fe\uff1a\u5de6\u4fa7\u4e3a\u524d\u5411\u6d41\uff0c\u53f3\u4fa7\u4e3a\u540e\u5411\u6d41\uff0c\u4e2d\u95f4\u4e3a\u8be2\u95ee\u6d41\u3002\u5176\u4e2d\uff0c\u524d\u5411\u6d41\u7528\u4e8e\u6536\u96c6\u524d\u5411\u8bed\u4e49\uff0c\u540e\u5411\u6d41\u7528\u4e8e\u6536\u96c6\u540e\u5411\u8bed\u4e49\uff0c\u800c\u8be2\u95ee\u6d41\u7528\u4e8e\u6355\u83b7\u5f53\u524d\u4f4d\u7f6e\u5728\u5176\u4e4b\u524d\u7684\u524d\u5411\u6d41\u8bed\u4e49\u548c\u5176\u4e4b\u540e\u7684\u540e\u5411\u6d41\u8bed\u4e49<\/em><\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u5468\u8fb9\u89c6\u89c9\u6ce8\u610f\u529b\u7f51\u7edc\">\u5468\u8fb9\u89c6\u89c9\u6ce8\u610f\u529b\u7f51\u7edc<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"547\" height=\"138\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-5.png\" alt=\"paper screenshot\" class=\"wp-image-1070325\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-5.png 547w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-5-300x76.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-5-240x61.png 240w\" sizes=\"auto, (max-width: 547px) 100vw, 547px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/peripheral-vision-transformer\/<\/p>\n\n\n\n<p>\u4eba\u7c7b\u62e5\u6709\u5468\u8fb9\u89c6\u89c9\u8fd9\u79cd\u7279\u6b8a\u7684\u89c6\u89c9\u5904\u7406\u7cfb\u7edf\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u7684\u6574\u4e2a\u89c6\u91ce\u53ef\u4ee5\u6839\u636e\u5230\u51dd\u89c6\u4e2d\u5fc3\u7684\u8ddd\u79bb\u88ab\u5212\u5206\u4e3a\u591a\u4e2a\u8f6e\u5ed3\u533a\u57df\uff0c\u800c\u5468\u8fb9\u89c6\u89c9\u4f7f\u6211\u4eec\u80fd\u591f\u611f\u77e5\u4e0d\u540c\u533a\u57df\u7684\u5404\u79cd\u89c6\u89c9\u7279\u5f81\u3002\u53d7\u8be5\u751f\u7269\u5b66\u542f\u53d1\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5f00\u59cb\u63a2\u7d22\u5728\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u4e2d\u6a21\u62df\u5468\u8fb9\u89c6\u89c9\u8fdb\u884c\u89c6\u89c9\u8bc6\u522b\u7684\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u7814\u7a76\u5458\u4eec\u6240\u8bbe\u8ba1\u7684 PerViT \u7f51\u7edc\uff0c\u53ef\u4ee5\u5c06\u8f6e\u5ed3\u533a\u57df\u901a\u8fc7\u4f4d\u7f6e\u7f16\u7801\u7ed3\u5408\u5230\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e2d\uff0c\u4f7f\u7f51\u7edc\u638c\u63e1\u5982\u4f55\u5c06\u89c6\u91ce\u5212\u5206\u4e3a\u4e0d\u540c\u8f6e\u5ed3\u533a\u57df\u7684\u65b9\u6cd5\uff0c\u5e76\u80fd\u591f\u4ece\u4e0d\u540c\u533a\u57df\u4e2d\u63d0\u53d6\u76f8\u5e94\u7684\u7279\u5f81\u3002\u7814\u7a76\u5458\u4eec\u7cfb\u7edf\u5730\u7814\u7a76\u4e86\u673a\u5668\u611f\u77e5\u6a21\u578b\u7684\u5185\u90e8\u5de5\u4f5c\u539f\u7406\uff0c\u53d1\u73b0\u7f51\u7edc\u5b66\u4e60\u611f\u77e5\u89c6\u89c9\u6570\u636e\u7684\u65b9\u5f0f\u4e0e\u4eba\u7c7b\u89c6\u89c9\u76f8\u4f3c\u3002\u5728 ImageNet-1K \u4e0a\u5bf9 PerViT \u7f51\u7edc\u8bc4\u4f30\u7684\u7ed3\u679c\u663e\u793a\uff0cPerViT \u5728\u4e0d\u540c\u6a21\u578b\u5927\u5c0f\u4e0a\u7684\u56fe\u50cf\u5206\u7c7b\u6027\u80fd\u5747\u4f18\u4e8e\u57fa\u7ebf\uff0c\u8bc1\u660e\u4e86\u8be5\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"577\" height=\"723\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-6.png\" alt=\"\u56fe3\uff1a\u4eba\u7c7b\u5468\u8fb9\u89c6\u89c9\uff08\u4e0a\uff09\u4e0e\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u795e\u7ecf\u7f51\u7edc\uff08\u4e0b\uff09\u76f8\u7ed3\u5408\u4ee5\u8fdb\u884c\u89c6\u89c9\u8bc6\u522b\u7684\u793a\u610f\u56fe\" class=\"wp-image-1070328\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-6.png 577w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-6-239x300.png 239w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-6-144x180.png 144w\" sizes=\"auto, (max-width: 577px) 100vw, 577px\" \/><figcaption class=\"wp-element-caption\"><em>\u56fe3\uff1a\u4eba\u7c7b\u5468\u8fb9\u89c6\u89c9\uff08\u4e0a\uff09\u4e0e\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u795e\u7ecf\u7f51\u7edc\uff08\u4e0b\uff09\u76f8\u7ed3\u5408\u4ee5\u8fdb\u884c\u89c6\u89c9\u8bc6\u522b\u7684\u793a\u610f\u56fe<\/em><\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"vrl3-\u7531\u6570\u636e\u9a71\u52a8\u7684\u89c6\u89c9\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\">VRL3\uff1a\u7531\u6570\u636e\u9a71\u52a8\u7684\u89c6\u89c9\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u6846\u67b6<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"658\" height=\"191\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-7.png\" alt=\"paper screenshot\" class=\"wp-image-1070331\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-7.png 658w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-7-300x87.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-7-655x191.png 655w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-7-240x70.png 240w\" sizes=\"auto, (max-width: 658px) 100vw, 658px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/vrl3-a-data-driven-framework-for-visual-deep-reinforcement-learning\/<\/p>\n\n\n\n<p>\u5728\u5f3a\u5316\u5b66\u4e60\uff0c\u5c24\u5176\u662f\u673a\u5668\u4eba\u7cfb\u7edf\u7684\u8bad\u7ec3\u4e2d\uff0c\u65b0\u6570\u636e\u6837\u672c\u7684\u91c7\u96c6\u5f80\u5f80\u5341\u5206\u6602\u8d35\u3002\u4e3a\u4e86\u5b9e\u73b0\u7ecf\u6d4e\uff0c\u9ad8\u6548\uff0c\u670d\u52a1\u4e8e\u5927\u4f17\u7684\u6cdb\u7528\u6027\u5f3a\u5316\u5b66\u4e60\u548c\u673a\u5668\u4eba\u6280\u672f\uff0c\u7814\u7a76\u5458\u4eec\u5c1d\u8bd5\u7ed3\u5408\u5229\u7528\u591a\u79cd\u6570\u636e\u6765\u6e90\uff0c\u5927\u5e45\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u3002\u7814\u7a76\u5458\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u5168\u65b0\u7684\u6570\u636e\u9a71\u52a8\u7684\u5b66\u4e60\u6846\u67b6 VRL3\u3002VRL3 \u4f7f\u7528\u4e86\u4e09\u9636\u6bb5\u7684\u8bad\u7ec3\u65b9\u5f0f\uff0c\u6574\u5408\u4e86\u975e\u5f3a\u5316\u5b66\u4e60\u7684\u5927\u89c4\u6a21\u56fe\u50cf\u6570\u636e\u96c6\uff0c\u6709\u9650\u7684\u4eba\u7c7b\u4e13\u5bb6\u793a\u8303\u4ee5\u53ca\u5728\u7ebf\u5f3a\u5316\u5b66\u4e60\u6570\u636e\uff0c\u5e76\u52a0\u4ee5\u5145\u5206\u5229\u7528\uff0c\u5176\u53ef\u5728\u57fa\u4e8e\u89c6\u89c9\u8f93\u5165\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u4efb\u52a1\u5c24\u5176\u662f\u6a21\u62df\u673a\u5668\u4eba\u4efb\u52a1\u4e0a\uff0c\u4ee5\u60ca\u4eba\u7684\u6837\u672c\u6548\u7387\u8fdb\u884c\u5b66\u4e60\u3002<\/p>\n\n\n\n<p>\u76f8\u6bd4\u4e4b\u524d\u7684\u6700\u5148\u8fdb\u65b9\u6cd5\uff0c\u5728\u6781\u5bcc\u6311\u6218\u6027\u7684 Adroit \u673a\u68b0\u624b\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u6700\u96be\u7684\u4efb\u52a1\u4e0a\uff0cVRL3 \u53ef\u6781\u5176\u663e\u8457\u5730\u5c06\u6837\u672c\u6548\u7387\u63d0\u9ad824\u500d\uff0c\u5e76\u4ee510\u500d\u66f4\u5feb\u8ba1\u7b97\u901f\u5ea6\u548c3\u500d\u66f4\u5c11\u53c2\u6570\u9700\u6c42\u5b8c\u6210\u8bad\u7ec3\u3002\u5728\u8fbe\u5230\u6781\u9ad8\u6027\u80fd\u7684\u540c\u65f6\uff0cVRL3 \u8ffd\u6c42\u5927\u9053\u81f3\u7b80\u7684\u8bbe\u8ba1\u7406\u5ff5\uff0c\u7528\u7b80\u5355\u6613\u61c2\u7684\u8bbe\u8ba1\u601d\u8def\u548c\u4ee3\u7801\u5b9e\u73b0\u3002\u8fd9\u9879\u7814\u7a76\u5411\u5b9e\u73b0\u9ad8\u6548\u3001\u4fbf\u643a\u3001\u4f4e\u6210\u672c\u53ef\u5e7f\u6cdb\u670d\u52a1\u4e8e\u5927\u4f17\u7684\u5f3a\u5316\u5b66\u4e60\u548c\u673a\u5668\u4eba\u7cfb\u7edf\u8fc8\u51fa\u4e86\u5173\u952e\u4e00\u6b65\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"404\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-8-1024x404.png\" alt=\"\u56fe4\uff1aVRL3 \u6a21\u578b\u8bbe\u8ba1\u56fe\" class=\"wp-image-1070334\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-8-1024x404.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-8-300x118.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-8-768x303.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-8-240x95.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-8.png 1267w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\"><em>\u56fe4\uff1aVRL3 \u6a21\u578b\u8bbe\u8ba1\u56fe<\/em><\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u4eba\u5de5\u667a\u80fd\u7406\u8bba\">\u4eba\u5de5\u667a\u80fd\u7406\u8bba<\/h2>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"theory\">Theory<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u7ec4\u5408\u591a\u81c2\u8001\u864e\u673a\u5728\u968f\u673a\u89e6\u53d1\u81c2\u6216\u72ec\u7acb\u81c2\u573a\u666f\u4e0b\u4e0e\u6700\u5927\u89e6\u53d1\u81c2\u6570\u91cf\u65e0\u5173\u7684\u635f\u5931\u5206\u6790\">\u7ec4\u5408\u591a\u81c2\u8001\u864e\u673a\u5728\u968f\u673a\u89e6\u53d1\u81c2\u6216\u72ec\u7acb\u81c2\u573a\u666f\u4e0b\u4e0e\u6700\u5927\u89e6\u53d1\u81c2\u6570\u91cf\u65e0\u5173\u7684\u635f\u5931\u5206\u6790<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"695\" height=\"488\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-9.png\" alt=\"paper screenshot\" class=\"wp-image-1070337\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-9.png 695w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-9-300x211.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-9-240x169.png 240w\" sizes=\"auto, (max-width: 695px) 100vw, 695px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/batch-size-independent-regret-bounds-for-combinatorial-semi-bandits-with-probabilistically-triggered-arms-or-independent-arms\/<\/p>\n\n\n\n<p>\u7ec4\u5408\u591a\u81c2\u8001\u864e\u673a\uff08combinatorial multi-armed bandit\uff09\u5c06\u4f20\u7edf\u7684\u7ec4\u5408\u4f18\u5316\u548c\u5728\u7ebf\u5b66\u4e60\u76f8\u7ed3\u5408\uff0c\u901a\u8fc7\u5728\u7ebf\u53cd\u9988\u673a\u5236\u4e0d\u65ad\u6539\u8fdb\u6a21\u578b\u7684\u4f18\u5316\u6548\u679c\u3002\u5176\u5e94\u7528\u6db5\u76d6\u63a8\u8350\u7cfb\u7edf\u3001\u5728\u7ebf\u5e7f\u544a\u3001\u793e\u4ea4\u7f51\u7edc\u3001\u65e0\u7ebf\u7f51\u7edc\u7b49\u591a\u4e2a\u9886\u57df\u3002\u5728\u672c\u6587\u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u901a\u8fc7\u65b9\u5dee\u5206\u6790\u7684\u65b9\u6cd5\u964d\u4f4e\u4e86\u6bcf\u4e2a\u65f6\u523b\u53ef\u80fd\u88ab\u6fc0\u6d3b\u7684\u81c2\u6570 K \u5bf9\u7b97\u6cd5\u6240\u627f\u53d7\u635f\u5931\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u7814\u7a76\u5458\u4eec\u627e\u5230\u4e86\u4e00\u79cd\u5168\u65b0\u7684\u5149\u6ed1\u6761\u4ef6\uff0c\u79f0\u4e3a\u6982\u7387\u6fc0\u6d3b\u65b9\u5dee\u8c03\u8282\uff08TPVM\uff09\u6761\u4ef6\u3002\u9996\u5148\uff0cTPVM \u88ab\u8bc1\u660e\u548c\u65e2\u6709\u7684\u5149\u6ed1\u6761\u4ef6\u5728\u591a\u6570\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4e0b\uff08\u5982\u5728\u7ebf\u5e7f\u544a\u3001\u793e\u4ea4\u7f51\u7edc\u7b49\uff09\u540c\u6837\u6210\u7acb\u3002\u5176\u6b21\uff0c\u901a\u8fc7 TPVM \u6761\u4ef6\uff0c\u7814\u7a76\u5458\u4eec\u5f97\u4ee5\u5bf9\u5e26\u6982\u7387\u548c\u65e0\u6982\u7387\u6fc0\u6d3b\u81c2\u7684\u6a21\u578b\u5206\u522b\u8bbe\u8ba1\u57fa\u4e8e\u65b9\u5dee\u5206\u6790\u7684\u65b0\u578b\u7b97\u6cd5\uff0c\u5373 BCUCB-T \u548c SESCB\u3002\u5728\u5e26\u6982\u7387\u6fc0\u6d3b\u81c2\u7684\u6a21\u578b\u4e0b\uff0cBCUCB-T \u7b97\u6cd5\u53ef\u4ee5\u5c06 K \u5bf9\u635f\u5931\u7684\u5f71\u54cd\u4ece\u6b64\u524d\u7684 O(K) \u964d\u4f4e\u4e3a O(log^2 K)\u6216 O(log K)\u3002\u800c\u5728\u65e0\u6982\u7387\u6fc0\u6d3b\u81c2\u7684\u6a21\u578b\u4e0b\uff0cSESCB \u5219\u5c06\u635f\u5931\u4e0a\u754c\u7531\u6b64\u524d\u7684 O(log K)\u964d\u4f4e\u81f3 O(1)\u3002\u6700\u540e\uff0c\u4eff\u771f\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u7814\u7a76\u5458\u4eec\u6240\u63d0\u51fa\u7684\u7b97\u6cd5\u5728\u5f88\u591a\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4e2d\u90fd\u80fd\u8d85\u8d8a\u73b0\u6709\u7b97\u6cd5\u7684\u6548\u679c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"837\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-10-1024x837.png\" alt=\"\u56fe5\uff1a\u5e26\u6982\u7387\u6fc0\u6d3b\u81c2\u7684 CMAB \u6a21\u578b\u4e0b\u672c\u6587\u7ed3\u679c\u4e0e\u6b64\u524d\u7ed3\u679c\u5bf9\u6bd4\u56fe\" class=\"wp-image-1070340\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-10-1024x837.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-10-300x245.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-10-768x628.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-10-220x180.png 220w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-10.png 1203w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\"><em>\u56fe5\uff1a\u5e26\u6982\u7387\u6fc0\u6d3b\u81c2\u7684 CMAB \u6a21\u578b\u4e0b\u672c\u6587\u7ed3\u679c\u4e0e\u6b64\u524d\u7ed3\u679c\u5bf9\u6bd4\u56fe<\/em><\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u52a8\u91cf\u4f1a\u6539\u53d8\u4f18\u5316\u5668\u5728\u53ef\u5206\u6570\u636e\u4e0a\u7684\u9690\u5f0f\u6b63\u5219\u5417\">\u52a8\u91cf\u4f1a\u6539\u53d8\u4f18\u5316\u5668\u5728\u53ef\u5206\u6570\u636e\u4e0a\u7684\u9690\u5f0f\u6b63\u5219\u5417\uff1f<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"807\" height=\"409\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-11.png\" alt=\"paper screenshot\" class=\"wp-image-1070343\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-11.png 807w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-11-300x152.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-11-768x389.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-11-240x122.png 240w\" sizes=\"auto, (max-width: 807px) 100vw, 807px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/does-momentum-change-the-implicit-regularization-on-separable-data\/<\/p>\n\n\n\n<p>\u4e3a\u63d0\u5347\u8bad\u7ec3\u901f\u5ea6\uff0c\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u4f18\u5316\u5668\u5e7f\u6cdb\u91c7\u7528\u52a8\u91cf\u52a0\u901f\u6280\u672f\u3002\u7136\u800c\uff0c\u76ee\u524d\u5b66\u754c\u4ecd\u672a\u80fd\u5398\u6e05\u52a8\u91cf\u662f\u5982\u4f55\u5f71\u54cd\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u7684\u3002\u672c\u6587\u4ece\u52a8\u91cf\u52a0\u901f\u6280\u672f\u7684\u9690\u5f0f\u6b63\u5219\u6548\u5e94\u5207\u5165\uff0c\u63a2\u7a76\u4e86\u5176\u5bf9\u6cdb\u5316\u80fd\u529b\u7684\u5f71\u54cd\u3002\u5c24\u5176\u662f\uff0c\u672c\u6587\u8bc1\u660e\u4e86\u5728\u5bf9\u7ebf\u6027\u53ef\u5206\u6570\u636e\u4e0a\uff0c\u5e26\u52a8\u91cf\u7684\u68af\u5ea6\u4e0b\u964d\u6cd5\u6536\u655b\u5230\u7684\u70b9\u662f L^2 \u6700\u5927\u95f4\u9694\u95ee\u9898\u7684\u89e3 (L^2 max-margin solution)\uff0c\u4e0e\u4e0d\u4f7f\u7528\u52a8\u91cf\u7684\u68af\u5ea6\u4e0b\u964d\u6cd5\u76f8\u540c\u3002\u8fd9\u610f\u5473\u7740\u5e26\u52a8\u91cf\u7684\u68af\u5ea6\u4e0b\u964d\u6cd5\u5c06\u6536\u655b\u5230\u4e00\u4e2a\u4f4e\u590d\u6742\u5ea6\u7684\u6a21\u578b\uff0c\u4ece\u800c\u4fdd\u8bc1\u4e86\u6a21\u578b\u7684\u6cdb\u5316\u6027\u8d28\u3002<\/p>\n\n\n\n<p>\u672c\u6587\u66f4\u8fdb\u4e00\u6b65\u5206\u6790\u4e86\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u5e26\u968f\u673a\u6027\u548c\u5e26\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7684\u53d8\u79cd(\u5373\u968f\u673a\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u548c\u786e\u5b9a\u6027 Adam \u7b97\u6cd5)\uff0c\u8bc1\u660e\u4e86\u5b83\u4eec\u4e5f\u4f1a\u6536\u655b\u5230 L^2 \u6700\u5927\u95f4\u9694\u95ee\u9898\u7684\u89e3\u3002\u8fd9\u9996\u6b21\u8bc1\u660e\u4e86\u968f\u673a\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\u5728\u4eff\u5c04\u566a\u58f0\u5047\u8bbe\u4e0b\uff0c\u5c06\u6536\u655b\u5230\u9a7b\u70b9\u3002\u8fd9\u4e00\u5047\u8bbe\u76f8\u6bd4\u73b0\u6709\u7814\u7a76\u4e2d\u6709\u754c\u65b9\u5dee\u566a\u58f0\u7684\u5047\u8bbe\uff0c\u9002\u7528\u8303\u56f4\u66f4\u4e3a\u5e7f\u6cdb\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u591a\u4e2a\u573a\u666f\u7684\u6570\u503c\u5b9e\u9a8c\u9a8c\u8bc1\u4e86\u8be5\u7406\u8bba\u7ed3\u679c\uff0c\u8bf7\u67e5\u770b\u8bba\u6587\u539f\u6587\u4e86\u89e3\u66f4\u591a\u7ec6\u8282\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"735\" height=\"669\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-12.png\" alt=\"\u56fe6\uff1a\u8bba\u6587\u300a\u52a8\u91cf\u4f1a\u6539\u53d8\u4f18\u5316\u5668\u5728\u53ef\u5206\u6570\u636e\u4e0a\u7684\u9690\u5f0f\u6b63\u5219\u5417\uff1f\u300b\u7684\u7ed3\u679c\u5c55\u793a\u56fe\" class=\"wp-image-1070346\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-12.png 735w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-12-300x273.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-12-198x180.png 198w\" sizes=\"auto, (max-width: 735px) 100vw, 735px\" \/><figcaption class=\"wp-element-caption\"><em>\u56fe6\uff1a\u8bba\u6587\u300a\u52a8\u91cf\u4f1a\u6539\u53d8\u4f18\u5316\u5668\u5728\u53ef\u5206\u6570\u636e\u4e0a\u7684\u9690\u5f0f\u6b63\u5219\u5417\uff1f\u300b\u7684\u7ed3\u679c\u5c55\u793a\u56fe<\/em><\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u7a33\u5b9a\u7684\u795e\u7ecf\u5143\u54cd\u5e94\u4f1a\u63d0\u5347\u6a21\u578b\u6cdb\u5316\u6027\u80fd\">\u7a33\u5b9a\u7684\u795e\u7ecf\u5143\u54cd\u5e94\u4f1a\u63d0\u5347\u6a21\u578b\u6cdb\u5316\u6027\u80fd<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"690\" height=\"429\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-13.png\" alt=\"paper screenshot\" class=\"wp-image-1070349\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-13.png 690w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-13-300x187.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-13-240x149.png 240w\" sizes=\"auto, (max-width: 690px) 100vw, 690px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/neuron-with-steady-response-leads-to-better-generalization\/<\/p>\n\n\n\n<p>\u5982\u4f55\u63d0\u9ad8\u6a21\u578b\u7684\u6cdb\u5316\u6027\u80fd\uff0c\u4e00\u76f4\u662f\u673a\u5668\u5b66\u4e60\u548c\u6df1\u5ea6\u5b66\u4e60\u7684\u6838\u5fc3\u95ee\u9898\u4e4b\u4e00\u3002\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u4e0d\u65ad\u53d1\u5c55\uff0c\u5404\u79cd\u5404\u6837\u7684\u7f51\u7edc\u7ed3\u6784\u88ab\u5e94\u7528\u5728\u591a\u79cd\u4e0d\u540c\u7684\u4efb\u52a1\u4e2d\u3002\u80fd\u5426\u63a2\u5bfb\u5230\u7edf\u6444\u4e0d\u540c\u4efb\u52a1\u548c\u7f51\u683c\u7ed3\u6784\u7684\u672c\u8d28\u5171\u6027\u6765\u63d0\u9ad8\u591a\u79cd\u7f51\u7edc\u7684\u6cdb\u5316\u6027\u80fd\uff0c\u662f\u672c\u6587\u7684\u7814\u7a76\u5207\u5165\u70b9\u3002<\/p>\n\n\n\n<p>\u7814\u7a76\u5458\u4eec\u4ece\u795e\u7ecf\u5143\u7ea7\u522b\u7684\u7ec6\u7c92\u5ea6\u51fa\u53d1\uff0c\u4ed4\u7ec6\u5206\u6790\u4e86\u5355\u4e2a\u795e\u7ecf\u5143\u5728\u795e\u7ecf\u7f51\u7edc\u8bad\u7ec3\u548c\u6d4b\u8bd5\u4e2d\u7684\u54cd\u5e94\u7279\u6027\uff0c\u53d1\u73b0\u63d0\u5347\u795e\u7ecf\u5143\u5bf9\u540c\u7c7b\u8f93\u5165\u6837\u672c\u54cd\u5e94\u7684\u7a33\u5b9a\u6027\u80fd\u591f\u6709\u6548\u5730\u63d0\u9ad8\u795e\u7ecf\u7f51\u7edc\u7684\u6cdb\u5316\u6027\u80fd\u3002\u636e\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u8bbe\u8ba1\u51fa\u4e86\u4e00\u79cd\u901a\u7528\u7684\u6b63\u5219\u9879\uff0c\u7528\u4e8e\u63a7\u5236\u795e\u7ecf\u5143\u5728\u6fc0\u6d3b\u72b6\u6001\u4e0b\u54cd\u5e94\u7684\u7c7b\u5185\u65b9\u5dee\uff0c\u5e76\u8fdb\u4e00\u6b65\u5206\u6790\u4e86\u5c06\u6b64\u6b63\u5219\u9879\u5e94\u7528\u5728\u4e0d\u540c\u5c42\u795e\u7ecf\u5143\u6240\u5e26\u6765\u7684\u6548\u679c\u5dee\u5f02\u3002\u8be5\u6b63\u5219\u9879\u7b80\u5355\u9ad8\u6548\uff0c\u80fd\u5728\u4e0d\u540c\u9886\u57df\u7684\u591a\u4e2a\u6570\u636e\u96c6\uff08ImageNet, CIFAR10, PubMed, WikiCS)\u4ee5\u53ca\u591a\u79cd\u7f51\u7edc\u7ed3\u6784 (MLP, CNN, GNN)\u4e0a\u666e\u904d\u5730\u63d0\u5347\u6a21\u578b\u7684\u6cdb\u5316\u6027\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"258\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-14-1024x258.png\" alt=\"\u56fe7\uff1aMLP \u6a21\u578b\u5728 MNIST \u6570\u636e\u96c6\u4e0a\u7684\u8bad\u7ec3\u8fc7\u7a0b\u56fe\u3002\u5176\u4e2d\uff0c\u7ea2\u7ebf\u8868\u793a\u539f\u59cb\u6a21\u578b\u7684\u8bad\u7ec3\u66f2\u7ebf\uff0c\u84dd\u7ebf\u662f\u52a0\u5165\u5168\u65b0\u6b63\u5219\u5316\u6280\u672f\u540e\u7684\u65b0\u6a21\u578b\u8bad\u7ec3\u66f2\u7ebf\u3002\u6700\u53f3\u7684\u56fe\u7247\u663e\u793a\uff0c\u5728\u539f\u59cb\u6a21\u578b\u4e2d\uff0c\u795e\u7ecf\u5143\u5bf9\u540c\u7c7b\u6837\u672c\u54cd\u5e94\u7684\u65b9\u5dee\u4f1a\u968f\u7740\u8bad\u7ec3\u4e0d\u65ad\u589e\u5927\uff0c\u800c\u7814\u7a76\u5458\u4eec\u6240\u63d0\u51fa\u7684\u6b63\u5219\u5316\u6280\u672f\u80fd\u591f\u5927\u5e45\u964d\u4f4e\u795e\u7ecf\u5143\u54cd\u5e94\u7684\u7c7b\u5185\u65b9\u5dee\u3002\u76f8\u5e94\u5730\uff0c\u539f\u59cb\u6a21\u578b\u7684\u8bad\u7ec3\u635f\u5931\u867d\u7136\u6bd4\u65b0\u6a21\u578b\u8981\u4f4e\uff08\u89c1\u7b2c\u4e8c\u5e45\u56fe\uff09\uff0c\u4f46\u65b0\u6a21\u578b\u5728\u6d4b\u8bd5\u96c6\u4e0a\u7684\u8bc6\u522b\u51c6\u786e\u7387\u5374\u6bd4\u539f\u59cb\u6a21\u578b\u6709\u663e\u8457\u63d0\u9ad8\uff08\u89c1\u7b2c\u4e00\u5e45\u56fe\uff09\uff0c\u56e0\u6b64\u7814\u7a76\u5458\u4eec\u6240\u63d0\u51fa\u7684\u6b63\u5219\u5316\u6280\u672f\u6709\u6548\u5730\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u6cdb\u5316\u6027\u80fd\u3002\" class=\"wp-image-1070352\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-14-1024x258.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-14-300x76.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-14-768x194.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-14-240x61.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-14.png 1268w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\"><em>\u56fe7\uff1aMLP \u6a21\u578b\u5728 MNIST \u6570\u636e\u96c6\u4e0a\u7684\u8bad\u7ec3\u8fc7\u7a0b\u56fe\u3002\u5176\u4e2d\uff0c\u7ea2\u7ebf\u8868\u793a\u539f\u59cb\u6a21\u578b\u7684\u8bad\u7ec3\u66f2\u7ebf\uff0c\u84dd\u7ebf\u662f\u52a0\u5165\u5168\u65b0\u6b63\u5219\u5316\u6280\u672f\u540e\u7684\u65b0\u6a21\u578b\u8bad\u7ec3\u66f2\u7ebf\u3002\u6700\u53f3\u7684\u56fe\u7247\u663e\u793a\uff0c\u5728\u539f\u59cb\u6a21\u578b\u4e2d\uff0c\u795e\u7ecf\u5143\u5bf9\u540c\u7c7b\u6837\u672c\u54cd\u5e94\u7684\u65b9\u5dee\u4f1a\u968f\u7740\u8bad\u7ec3\u4e0d\u65ad\u589e\u5927\uff0c\u800c\u7814\u7a76\u5458\u4eec\u6240\u63d0\u51fa\u7684\u6b63\u5219\u5316\u6280\u672f\u80fd\u591f\u5927\u5e45\u964d\u4f4e\u795e\u7ecf\u5143\u54cd\u5e94\u7684\u7c7b\u5185\u65b9\u5dee\u3002\u76f8\u5e94\u5730\uff0c\u539f\u59cb\u6a21\u578b\u7684\u8bad\u7ec3\u635f\u5931\u867d\u7136\u6bd4\u65b0\u6a21\u578b\u8981\u4f4e\uff08\u89c1\u7b2c\u4e8c\u5e45\u56fe\uff09\uff0c\u4f46\u65b0\u6a21\u578b\u5728\u6d4b\u8bd5\u96c6\u4e0a\u7684\u8bc6\u522b\u51c6\u786e\u7387\u5374\u6bd4\u539f\u59cb\u6a21\u578b\u6709\u663e\u8457\u63d0\u9ad8\uff08\u89c1\u7b2c\u4e00\u5e45\u56fe\uff09\uff0c\u56e0\u6b64\u7814\u7a76\u5458\u4eec\u6240\u63d0\u51fa\u7684\u6b63\u5219\u5316\u6280\u672f\u6709\u6548\u5730\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u6cdb\u5316\u6027\u80fd\u3002<\/em><\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u7b49\u7ea7\u5f3a\u5316\u5b66\u4e60-\u60b2\u89c2\u9762\u5bf9\u4e0d\u786e\u5b9a\u6027\u4e0e-\u5e38\u6570regret\">\u7b49\u7ea7\u5f3a\u5316\u5b66\u4e60\uff1a\u60b2\u89c2\u9762\u5bf9\u4e0d\u786e\u5b9a\u6027\u4e0e\u201c\u5e38\u6570regret\u201d<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"678\" height=\"162\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-15.png\" alt=\"paper screenshot\" class=\"wp-image-1070355\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-15.png 678w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-15-300x72.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-15-240x57.png 240w\" sizes=\"auto, (max-width: 678px) 100vw, 678px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/tiered-reinforcement-learning-pessimism-in-the-face-of-uncertainty-and-constant-regret\/<\/p>\n\n\n\n<p>\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u5728\u8bb8\u591a\u7528\u6237\u4ea4\u4e92\u578b\u5e94\u7528\u4e2d\u90fd\u53d6\u5f97\u4e86\u6210\u529f\uff0c\u6bd4\u5982\u533b\u7597\u9886\u57df\u3001\u63a8\u8350\u7cfb\u7edf\u7b49\u3002\u5176\u4e2d\uff0c\u75c5\u4eba\/\u5ba2\u6237\u626e\u6f14\u7684\u662f\u73af\u5883\u7684\u89d2\u8272\uff0c\u6cbb\u7597\u65b9\u6848\/\u63a8\u8350\u7b97\u6cd5\u5219\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u51b3\u7b56\uff0c\u5176\u5177\u4f53\u505a\u6cd5\u5c31\u662f\u90e8\u7f72\u51b3\u7b56\u3001\u6536\u96c6\u6570\u636e\u3001\u5e76\u7528\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u63d0\u5347\u76f4\u5230\u63a5\u8fd1\u6700\u4f18\u3002<\/p>\n\n\n\n<p>\u7531\u4e8e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7b97\u6cd5\u4e0d\u53ef\u907f\u514d\u5730\u4f1a\u7ed9\u7528\u6237\u63d0\u4f9b\u9519\u8bef\u7684\u51b3\u7b56\uff0c\u4f46\u4e0d\u540c\u7684\u7528\u6237\u5bf9\u4e8e\u627f\u62c5\u6216\u63a5\u6536\u51b3\u7b56\u9519\u8bef\u6240\u9020\u6210\u7684\u635f\u5931\u7684\u80fd\u529b\u4e5f\u4e0d\u540c\uff0c\u6545\u9700\u5177\u4f53\u60c5\u51b5\u5177\u4f53\u5206\u6790\u3002\u7136\u800c\uff0c\u73b0\u6709\u6846\u67b6\u5ffd\u7565\u4e86\u51b3\u7b56\u9519\u8bef\u5bf9\u4e2a\u4f53\u7528\u6237\u5f71\u54cd\u7684\u7279\u6b8a\u6027\u3002\u56e0\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u201c\u7b49\u7ea7\u5f3a\u5316\u5b66\u4e60\uff08Tiered RL\uff09\u201d\u7684\u65b0\u6846\u67b6\uff0c\u5176\u6839\u636e\u98ce\u9669\u7684\u627f\u53d7\u80fd\u529b\u5bf9\u7528\u6237\u8fdb\u884c\u7b49\u7ea7\u533a\u5206\uff0c\u627f\u53d7\u80fd\u529b\u8d8a\u5f31\u7b49\u7ea7\u8d8a\u9ad8\u3002<\/p>\n\n\n\n<p>\u5728\u5bf9 gap \u4e0d\u505a\u5047\u8bbe\u7684\u60c5\u51b5\u4e0b\uff0c\u7814\u7a76\u5458\u4eec\u8bc1\u660e\u4e86\u4e0e online setting \u76f8\u540c\u7684 O(\u221a(SAH^3 K)) \u7684\u6781\u5c0f\u5316\u6781\u5927\u4e0b\u754c\uff08minimax-lower bound\uff09\uff0c\u63ed\u793a\u4e86\u4e00\u822c\u60c5\u5f62\u4e0b\u8fd9\u4e2a\u95ee\u9898\u7684\u56f0\u96be\u5ea6\u3002\u5728\u5f15\u5165\u6700\u5c0f gap \u5047\u8bbe\u7684 gap-dependent setting \u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u8bbe\u8ba1\u4e86\u65b0\u7684\u7b97\u6cd5\uff0c\u5728\u4fdd\u8bc1\u4f4e\u7b49\u7ea7\u7528\u6237\u7684\u201cregret\u201d\u4ecd\u7136\u662f\u6700\u4f18\u7684\u524d\u63d0\u4e0b\uff0c\u9ad8\u7b49\u7ea7\u7528\u6237\u627f\u62c5\u7684\u201cregret \u4e0e\u4ea4\u4e92\u6b21\u6570 K \u65e0\u5173\uff0c\u6253\u7834\u4e86\u666e\u901a online learning \u7684 O(log K) \u4e0b\u754c\uff0c\u4ece\u800c\u8bc1\u660e\u4e86\u65b0\u6846\u67b6\u548c\u7b97\u6cd5\u7684\u4f18\u8d8a\u6027\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"988\" height=\"436\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-16.png\" alt=\"\u56fe8\uff1a\u7b49\u7ea7\u5f3a\u5316\u5b66\u4e60\uff08Tiered RL\uff09\u6846\u67b6\u56fe\" class=\"wp-image-1070358\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-16.png 988w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-16-300x132.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-16-768x339.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/neurips-2022-the-big-convergence-of-ai-theory-16-240x106.png 240w\" sizes=\"auto, (max-width: 988px) 100vw, 988px\" \/><figcaption class=\"wp-element-caption\"><em>\u56fe8\uff1a\u7b49\u7ea7\u5f3a\u5316\u5b66\u4e60\uff08Tiered RL\uff09\u6846\u67b6\u56fe<\/em><\/figcaption><\/figure>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7f16\u8005\u6309\uff1a\u4f5c\u4e3a\u76ee\u524d\u5168\u7403\u6700\u8d1f\u76db\u540d\u7684\u4eba\u5de5\u667a\u80fd\u76db\u4f1a\u4e4b\u4e00\uff0cNeurIPS (Conference on Neural Information Processing Systems) \u5728\u6bcf\u5e74\u5e74\u672b\u90fd\u662f\u8ba1\u7b97\u673a\u79d1\u5b66\u9886\u57df\u77a9\u76ee\u7684\u7126\u70b9\u3002\u88ab NeurIPS \u63a5\u6536\u7684\u8bba\u6587\uff0c\u4ee3\u8868\u7740\u5f53\u4eca\u795e\u7ecf\u79d1\u5b66\u548c\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u7684\u6700\u9ad8\u6c34\u5e73\u3002\u4eca\u5e74\u7684 NeurIPS \u5927\u4f1a\u5c06\u4e8e11\u670828\u65e5\u81f312\u67089\u65e5\u4e3e\u884c\uff0c\u672c\u5c4a\u5927\u4f1a\u5171\u6536\u523010411\u7bc7\u6709\u6548\u6295\u7a3f\uff0c\u5176\u4e2d2672\u7bc7\u83b7\u63a5\u6536\uff0c\u6700\u7ec8\u63a5\u6536\u7387\u4e3a25.6%\u3002\u76f8\u6bd4\u53bb\u5e74\uff0c\u6295\u7a3f\u6570\u91cf\u7ee7\u7eed\u589e\u52a0\u3002 \u5728\u672c\u5c4a\u5927\u4f1a\u4e2d\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e5f\u6709\u8bf8\u591a\u8bba\u6587\u5165\u9009\uff0c\u5185\u5bb9\u4e3b\u8981\u6db5\u76d6\u4eba\u5de5\u667a\u80fd\u4e94\u5927\u70ed\u70b9\u8bdd\u9898\uff1a\u4eba\u5de5\u667a\u80fd\u8d70\u5411\u5927\u4e00\u7edf\u3001\u8ba1\u7b97\u673a\u7406\u8bba\u3001\u8d4b\u80fd\u4ea7\u4e1a\u754c\u7684\u4eba\u5de5\u667a\u80fd\u3001\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u3001\u4eba\u5de5\u667a\u80fd\u8d4b\u80fd\u5185\u5bb9\u4e0e\u8bbe\u8ba1\u751f\u6210\u3002\u5728\u63a5\u4e0b\u6765\u7684\u51e0\u5468\u91cc\uff0c\u6211\u4eec\u5c06\u6309\u8bdd\u9898\u4e0e\u5927\u5bb6\u5206\u4eab\u76f8\u5173\u9886\u57df\u7684\u5b66\u672f\u6700\u524d\u6cbf\uff01\u4eca\u5929\uff0c\u8ba9\u6211\u4eec\u5148\u4ece\u201c\u4eba\u5de5\u667a\u80fd\u8d70\u5411\u5927\u4e00\u7edf\u201d\u548c\u201c\u8ba1\u7b97\u673a\u7406\u8bba\u201d\u8bdd\u9898\u4e0b\u76848\u7bc7\u8bba\u6587\u7cbe\u534e\u5f00\u59cb\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/mask-based-latent-reconstruction-for-reinforcement-learning\/ \u89c6\u89c9\u72b6\u6001\u8868\u5f81\u7684\u8d28\u91cf\u5bf9\u57fa\u4e8e\u89c6\u89c9\u7684\u5f3a\u5316\u5b66\u4e60\uff08vision-based reinforcement learning\uff09\u81f3\u5173\u91cd\u8981\u3002\u4e3a\u4e86\u5b66\u4e60\u9ad8\u6548\u7684\u72b6\u6001\u8868\u5f81\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u521b\u65b0\u6027\u5730\u5c06\u57fa\u4e8e\u63a9\u7801\u7684\u5efa\u6a21\u6280\u672f\uff08mask-based modeling\uff09\u5e94\u7528\u5230\u5f3a\u5316\u5b66\u4e60\u4e2d\uff0c\u4ee5\u4fc3\u8fdb\u5176\u72b6\u6001\u8868\u5f81\u5b66\u4e60\u3002\u6b64\u524d\u57fa\u4e8e\u63a9\u7801\u7684\u5efa\u6a21\u6280\u672f\u5df2\u7ecf\u5728 CV \u548c NLP \u9886\u57df\u4e2d\u5927\u653e\u5f02\u5f69\uff0c\u800c\u8fd9\u9879\u5de5\u4f5c\u662f\u5c06\u5176\u5e94\u7528\u5230\u5f3a\u5316\u5b66\u4e60\u9886\u57df\u5e2e\u52a9\u7b56\u7565\u5b66\u4e60\u7684\u9996\u6b21\u63a2\u7d22\u3002 \u5177\u4f53\u5730\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u800c\u6709\u6548\u7684\u81ea\u76d1\u7763\u65b9\u6cd5\uff0c\u5373\u57fa\u4e8e\u63a9\u7801\u7684\u9690\u7a7a\u95f4\u91cd\u5efa (mask-based latent reconstruction\uff0c\u7b80\u79f0\u4e3a MLR)\u3002MLR \u901a\u8fc7\u4ece\u5177\u6709\u65f6\u7a7a\u63a9\u7801\u7684\u89c6\u89c9\u72b6\u6001\u4e2d\u9884\u6d4b\u5176\u5728\u9690\u7a7a\u95f4\u4e2d\u7684\u5b8c\u6574\u8868\u5f81\uff0c\u4ece\u800c\u4f7f\u795e\u7ecf\u7f51\u7edc\u5728\u5b66\u4e60\u72b6\u6001\u8868\u5f81\u65f6\u80fd\u591f\u66f4\u597d\u5730\u5229\u7528\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u7f16\u7801\u66f4\u591a\u7b56\u7565\u5b66\u4e60\u6240\u9700\u8981\u7684\u8bed\u4e49\u4fe1\u606f\u3002\u5927\u91cf\u57fa\u51c6\u5b9e\u9a8c\u8868\u660e\uff0cMLR \u663e\u8457\u63d0\u9ad8\u4e86\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u7684\u6837\u672c\u6548\u7387\uff08sample efficiency\uff09\uff0c\u5728\u591a\u4e2a\u8fde\u7eed\u548c\u79bb\u6563\u7684\u5f3a\u5316\u5b66\u4e60\u73af\u5883\u4e2d\u53d6\u5f97\u4e86 SOTA \u7684\u6027\u80fd\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/transcormer-transformer-for-sentence-scoring-with-sliding-language-modeling\/ \u53e5\u5b50\u8bc4\u5206\u65e8\u5728\u8bc4\u4f30\u4e00\u4e2a\u53e5\u5b50\u7684\u6700\u5927\u4f3c\u7136\u4f30\u8ba1\uff0c\u88ab\u5e7f\u6cdb\u5e94\u7528\u4e8e\u8bb8\u591a\u81ea\u7136\u8bed\u8a00\u4efb\u52a1\u7684\u573a\u666f\u4e2d\uff0c\u5305\u62ec\u91cd\u6392\u5e8f\u3001\u8bed\u8a00\u53ef\u63a5\u53d7\u6027\u7b49\u3002\u8fc7\u53bb\u7528\u4e8e\u89e3\u51b3\u53e5\u5b50\u8bc4\u5206\u7684\u5de5\u4f5c\u4e3b\u8981\u4ee5\u4e24\u79cd\u7ecf\u5178\u8bed\u8a00\u6a21\u578b\u4e3a\u4e3b\uff1a\u56e0\u679c\u8bed\u8a00\u6a21\u578b\uff08causal language modeling, CLM\uff09\u548c\u63a9\u7801\u8bed\u8a00\u6a21\u578b\uff08masked language modeling, MLM\uff09\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u5de5\u4f5c\u90fd\u5b58\u5728\u4e00\u5b9a\u7684\u74f6\u9888\uff1aCLM \u867d\u7136\u53ea\u9700\u8981\u8ba1\u7b97\u4e00\u6b21\u4f46\u5374\u53ea\u5229\u7528\u4e86\u5355\u5411\u4fe1\u606f\uff1bMLM \u80fd\u591f\u5229\u7528\u53cc\u5411\u8bed\u4e49\uff0c\u4f46\u6bcf\u6b21\u53ea\u80fd\u9884\u6d4b\u90e8\u5206\u5355\u8bcd\u800c\u4e0d\u5f97\u4e0d\u9700\u8981\u591a\u6b21\u63a8\u7406\u3002 \u56e0\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u6ed1\u52a8\u8bed\u8a00\u6a21\u578b\u7684 Transformer \u6a21\u578b Transcormer\uff0c\u5e76\u5728\u5176\u4e2d\u8bbe\u8ba1\u4e86\u4e00\u79cd\u4e09\u6d41\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7528\u4e8e\u7ef4\u62a4\u6ed1\u52a8\u8bed\u8a00\u6a21\u578b\u3002\u5229\u7528\u8fd9\u6837\u7684\u8bbe\u8ba1\uff0cTranscormer \u53ef\u4ee5\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u5229\u7528\u53cc\u5411\u4fe1\u606f\u8fdb\u884c\u9884\u6d4b\u7684\u540c\u65f6\uff0c\u53ea\u9700\u4e00\u6b21\u8ba1\u7b97\u5373\u53ef\u5f97\u5230\u6240\u6709\u5355\u8bcd\u7684\u6982\u7387\u3002\u6ed1\u52a8\u8bed\u8a00\u6a21\u578b\u5728\u8ba1\u7b97\u53e5\u5b50\u8bc4\u5206\u65f6\uff0cTranscormer \u8fd8\u53ef\u4ee5\u907f\u514d CLM \u53ea\u80fd\u5229\u7528\u5355\u5411\u4fe1\u606f\u7684\u7f3a\u70b9\u4ee5\u53ca MLM \u9700\u8981\u591a\u6b21\u8ba1\u7b97\u7684\u4e0d\u8db3\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cTranscormer \u5728\u53e5\u5b50\u8bc4\u5206\u4efb\u52a1\u4e0a\u80fd\u591f\u53d6\u5f97\u6bd4\u5176\u4ed6\u65b9\u6cd5\u66f4\u597d\u7684\u7ed3\u679c\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/newed.any0.dpdns.org\/en-us\/research\/publication\/peripheral-vision-transformer\/ \u4eba\u7c7b\u62e5\u6709\u5468\u8fb9\u89c6\u89c9\u8fd9\u79cd\u7279\u6b8a\u7684\u89c6\u89c9\u5904\u7406\u7cfb\u7edf\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u7684\u6574\u4e2a\u89c6\u91ce\u53ef\u4ee5\u6839\u636e\u5230\u51dd\u89c6\u4e2d\u5fc3\u7684\u8ddd\u79bb\u88ab\u5212\u5206\u4e3a\u591a\u4e2a\u8f6e\u5ed3\u533a\u57df\uff0c\u800c\u5468\u8fb9\u89c6\u89c9\u4f7f\u6211\u4eec\u80fd\u591f\u611f\u77e5\u4e0d\u540c\u533a\u57df\u7684\u5404\u79cd\u89c6\u89c9\u7279\u5f81\u3002\u53d7\u8be5\u751f\u7269\u5b66\u542f\u53d1\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5f00\u59cb\u63a2\u7d22\u5728\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u4e2d\u6a21\u62df\u5468\u8fb9\u89c6\u89c9\u8fdb\u884c\u89c6\u89c9\u8bc6\u522b\u7684\u65b9\u6cd5\u3002 [&hellip;]<\/p>\n","protected":false},"author":34512,"featured_media":1070310,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1069932","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1069932","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":2,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1069932\/revisions"}],"predecessor-version":[{"id":1070361,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1069932\/revisions\/1070361"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media\/1070310"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1069932"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1069932"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1069932"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1069932"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}