{"id":34299,"date":"2024-12-04T02:56:53","date_gmt":"2024-12-03T18:56:53","guid":{"rendered":"https:\/\/17aitech.com\/?p=34299"},"modified":"2024-12-04T02:56:53","modified_gmt":"2024-12-03T18:56:53","slug":"icml-2024-%e6%8f%ad%e7%a4%ba%e9%9d%9e%e7%ba%bf%e5%bd%a2transformer%e5%9c%a8%e4%b8%8a%e4%b8%8b%e6%96%87%e5%ad%a6%e4%b9%a0%e4%b8%ad%e5%ad%a6%e4%b9%a0%e5%92%8c%e6%b3%9b%e5%8c%96%e7%9a%84%e6%9c%ba","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34299","title":{"rendered":"ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-06-28-7\" target=\"_blank\">ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f2ef263020affc3831b835d31c350f04.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f2ef263020affc3831b835d31c350f04.png\"><\/a><\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>AIxiv\u4e13\u680f\u662f<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/section>\n<\/section>\n<\/blockquote>\n<p><strong>\u672c\u6587\u4f5c\u8005\u674e\u5b8f\u5eb7\uff0c\u7f8e\u56fd\u4f26\u65af\u52d2\u7406\u5de5\u5927\u5b66\u7535\u6c14\u3001\u8ba1\u7b97\u673a\u4e0e\u7cfb\u7edf\u5de5\u7a0b\u7cfb\u5728\u8bfb\u535a\u58eb\u751f\uff0c\u672c\u79d1\u6bd5\u4e1a\u4e8e\u4e2d\u56fd\u79d1\u5b66\u6280\u672f\u5927\u5b66\u3002\u7814\u7a76\u65b9\u5411\u5305\u62ec<mark data-type=\"tech_methods\" data-id=\"01946acc-d031-4c0e-909c-f062643b7273\">\u6df1\u5ea6\u5b66\u4e60<\/mark>\u7406\u8bba\uff0c\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7406\u8bba\uff0c\u7edf\u8ba1<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u7b49\u7b49\u3002\u76ee\u524d\u5df2\u5728 ICLR\/ICML\/Neurips \u7b49 AI \u9876\u4f1a\u53d1\u8868\u591a\u7bc7\u8bba\u6587\u3002<\/strong><\/p>\n<p>\u4e0a\u4e0b\u6587\u5b66\u4e60 (in-context learning, \u7b80\u5199\u4e3a ICL) \u5df2\u7ecf\u5728\u5f88\u591a LLM \u6709\u5173\u7684\u5e94\u7528\u4e2d\u5c55\u73b0\u4e86\u5f3a\u5927\u7684\u80fd\u529b\uff0c\u4f46\u662f\u5bf9\u5176\u7406\u8bba\u7684\u5206\u6790\u4ecd\u7136\u6bd4\u8f83\u6709\u9650\u3002\u4eba\u4eec\u4f9d\u7136\u8bd5\u56fe\u7406\u89e3\u4e3a\u4ec0\u4e48\u57fa\u4e8e Transformer \u67b6\u6784\u7684 LLM \u53ef\u4ee5\u5c55\u73b0\u51fa ICL \u7684\u80fd\u529b\u3002<\/p>\n<p>\u8fd1\u671f\uff0c\u4e00\u4e2a\u6765\u81ea\u7f8e\u56fd\u4f26\u65af\u52d2\u7406\u5de5\u5927\u5b66\u548c <mark data-type=\"institutions\" data-id=\"b86983cf-b858-45b0-86a1-b05cb94cde08\">IBM<\/mark> \u7814\u7a76\u9662\u7684\u56e2\u961f\u4ece\u4f18\u5316\u548c\u6cdb\u5316\u7406\u8bba\u7684\u89d2\u5ea6\u5206\u6790\u4e86\u5e26\u6709\u975e\u7ebf\u6027\u6ce8\u610f\u529b\u6a21\u5757 (attention) \u548c<mark data-type=\"tech_methods\" data-id=\"2130a104-5169-4950-b418-b2aebd0b2d1e\">\u591a\u5c42\u611f\u77e5\u673a<\/mark> (MLP) \u7684 Transformer \u7684 ICL \u80fd\u529b\u3002\u4ed6\u4eec\u7279\u522b\u4ece\u7406\u8bba\u7aef\u8bc1\u660e\u4e86\u5355\u5c42 Transformer \u9996\u5148\u5728 attention \u5c42\u6839\u636e query \u9009\u62e9\u4e00\u4e9b\u4e0a\u4e0b\u6587\u793a\u4f8b\uff0c\u7136\u540e\u5728 MLP \u5c42\u6839\u636e\u6807\u7b7e\u5d4c\u5165\u8fdb\u884c\u9884\u6d4b\u7684 ICL \u673a\u5236\u3002\u8be5\u6587\u7ae0\u5df2\u6536\u5f55\u5728 ICML 2024\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-c25687dd659da305b1bbe3f4b16d7d02.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-c25687dd659da305b1bbe3f4b16d7d02.png\"><\/a><\/p>\n<ul>\n<li>\n<p>\u8bba\u6587\u9898\u76ee\uff1aHow Do Nonlinear Transformers Learn and Generalize in In-Context Learning?<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2402.15607<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u80cc\u666f\u4ecb\u7ecd<\/strong><\/p>\n<p><strong>\u4e0a\u4e0b\u6587\u5b66\u4e60 in context learning (ICL)<\/strong><\/p>\n<p>\u4e0a\u4e0b\u6587\u5b66\u4e60 (ICL) \u662f\u4e00\u79cd\u65b0\u7684\u5b66\u4e60\u8303\u5f0f\uff0c\u5728\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark> (LLM) \u4e2d\u975e\u5e38\u6d41\u884c\u3002\u5b83\u5177\u4f53\u662f\u6307\u5728\u6d4b\u8bd5<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark> (testing query)<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f55a8e6edd8fb5af4dcba702256ad350.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f55a8e6edd8fb5af4dcba702256ad350.png\"><\/a>\u524d\u6dfb\u52a0 N \u4e2a\u6d4b\u8bd5\u6837\u672c testing examples (\u4e0a\u4e0b\u6587)\uff0c\u5373\u6d4b\u8bd5\u8f93\u5165<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-80cf4bc99839dbb401ee553dd1ac5c40.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-80cf4bc99839dbb401ee553dd1ac5c40.png\"><\/a>\u548c\u6d4b\u8bd5\u8f93\u51fa<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-8e69a3c55a93d1c9dfe35ed6348f4e02.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-8e69a3c55a93d1c9dfe35ed6348f4e02.png\"><\/a>\u7684\u7ec4\u5408\uff0c\u4ece\u800c\u6784\u6210\u4e00\u4e2a testing prompt:<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-a4a9779905a3282b43372c81630ba3ed.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-a4a9779905a3282b43372c81630ba3ed.png\"><\/a>\uff0c\u4f5c\u4e3a\u6a21\u578b\u7684\u8f93\u5165\u4ee5\u5f15\u5bfc\u6a21\u578b\u4f5c\u51fa\u6b63\u786e\u7684\u63a8\u65ad\u3002\u8fd9\u79cd\u65b9\u5f0f\u4e0d\u540c\u4e8e\u7ecf\u5178\u7684\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u7684\u65b9\u5f0f\uff0c\u5b83\u4e0d\u9700\u8981\u6539\u53d8\u6a21\u578b\u7684<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\uff0c\u4ece\u800c\u66f4\u52a0\u7684\u9ad8\u6548\u3002<\/p>\n<p><strong>ICL \u7406\u8bba\u5de5\u4f5c\u7684\u8fdb\u5c55<\/strong><\/p>\n<p>\u8fd1\u671f\u7684\u5f88\u591a\u7406\u8bba\u5de5\u4f5c\u90fd\u662f\u57fa\u4e8e [1] \u6240\u63d0\u51fa\u7684\u7814\u7a76\u6846\u67b6\uff0c\u5373\u4eba\u4eec\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528 prompt \u7684\u683c\u5f0f\u6765\u5bf9 Transformer \u8fdb\u884c\u8bad\u7ec3 (\u8fd9\u4e00\u6b65\u4e5f\u53ef\u4ee5\u7406\u89e3\u4e3a\u5728\u6a21\u62df\u4e00\u79cd\u7b80\u5316\u7684 LLM \u9884\u8bad\u7ec3\u6a21\u5f0f)\uff0c\u4ece\u800c\u4f7f\u5f97\u6a21\u578b\u5177\u6709 ICL \u80fd\u529b\u3002\u5df2\u6709\u7684\u7406\u8bba\u5de5\u4f5c\u805a\u7126\u4e8e\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b (expressive power) \u7684\u89d2\u5ea6 [2]\u3002\u4ed6\u4eec\u53d1\u73b0\uff0c\u4eba\u4eec\u80fd\u591f\u627e\u5230\u4e00\u4e2a\u6709\u7740 \u201c\u5b8c\u7f8e\u201d \u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u7684 Transformer \u53ef\u4ee5\u901a\u8fc7\u524d\u5411\u8fd0\u7b97\u6267\u884c ICL\uff0c\u751a\u81f3\u9690\u542b\u5730\u6267\u884c<mark data-type=\"tech_methods\" data-id=\"847e8a6a-524c-4d34-a50e-b17cc5dd84d4\">\u68af\u5ea6\u4e0b\u964d<\/mark>\u7b49\u7ecf\u5178<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u7b97\u6cd5\u3002\u4f46\u662f\u8fd9\u4e9b\u5de5\u4f5c\u65e0\u6cd5\u56de\u7b54\u4e3a\u4ec0\u4e48 Transformer \u53ef\u4ee5\u88ab\u8bad\u7ec3\u6210\u8fd9\u6837 \u201c\u5b8c\u7f8e\u201d \u7684\uff0c\u5177\u6709 ICL \u80fd\u529b\u7684<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u3002\u56e0\u6b64\uff0c\u8fd8\u6709\u4e00\u4e9b\u5de5\u4f5c\u8bd5\u56fe\u4ece Transformer \u7684\u8bad\u7ec3\u6216\u6cdb\u5316\u7684\u89d2\u5ea6\u7406\u89e3 ICL \u673a\u5236 [3,4]\u3002\u4e0d\u8fc7\uff0c\u53d7\u5236\u4e8e\u5206\u6790 Transformer \u7ed3\u6784\u7684\u590d\u6742\u6027\uff0c\u8fd9\u4e9b\u5de5\u4f5c\u76ee\u524d\u6b62\u6b65\u4e8e\u7814\u7a76<mark data-type=\"tech_methods\" data-id=\"7d29606b-791a-4280-844c-0f5d88782dfb\">\u7ebf\u6027\u56de\u5f52<\/mark>\u4efb\u52a1\uff0c\u800c\u6240\u8003\u8651\u7684\u6a21\u578b\u901a\u5e38\u4f1a\u7565\u53bb Transformer \u4e2d\u7684\u975e\u7ebf\u5f62\u90e8\u5206\u3002<\/p>\n<p>\u672c\u6587\u4ece\u4f18\u5316\u548c\u6cdb\u5316\u7406\u8bba\u7684\u89d2\u5ea6\u5206\u6790\u4e86\u5e26\u6709\u975e\u7ebf\u6027 attention \u548c MLP \u7684 Transformer \u7684 ICL \u80fd\u529b\u548c\u673a\u5236\uff1a<\/p>\n<ul>\n<li>\n<p>\u57fa\u4e8e\u4e00\u4e2a\u7b80\u5316\u7684\u5206\u7c7b\u6a21\u578b\uff0c\u672c\u6587\u5177\u4f53<mark data-type=\"tech_tasks\" data-id=\"e1abaa20-0000-4ff4-ad42-f97ca7fca4b1\">\u91cf\u5316<\/mark>\u4e86\u6570\u636e\u7684\u7279\u5f81\u5982\u4f55\u5f71\u54cd\u4e86\u4e00\u5c42\u5355\u5934 Transformer \u7684\u57df\u5185 (in-domain) \u548c\u57df\u5916 (out-of-domain, OOD) \u7684 ICL \u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p>\u672c\u6587\u8fdb\u4e00\u6b65\u9610\u91ca\u4e86 ICL \u662f\u5982\u4f55\u901a\u8fc7\u88ab\u8bad\u7ec3\u7684 Transformer \u6765\u5b9e\u73b0\u4e86\u3002<\/p>\n<\/li>\n<li>\n<p>\u57fa\u4e8e\u88ab\u8bad\u7ec3\u7684 Transformer \u7684\u7279\u70b9\uff0c\u672c\u6587\u8fd8\u5206\u6790\u4e86\u5728 ICL \u63a8\u65ad\u7684\u65f6\u5019\u4f7f\u7528\u57fa\u4e8e\u5e45\u503c\u7684\u6a21\u578b<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark> (magnitude-based pruning) \u7684\u53ef\u884c\u6027\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u7406\u8bba\u90e8\u5206<\/strong><\/p>\n<p><strong>\u95ee\u9898\u63cf\u8ff0<\/strong><\/p>\n<p>\u672c\u6587\u8003\u8651\u4e00\u4e2a\u4e8c<mark data-type=\"concepts\" data-id=\"aaeb1c15-66f2-4822-91bb-b441607f9ecf\">\u5206\u7c7b\u95ee\u9898<\/mark>\uff0c\u5373\u5c06<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-62e04a2d01e9c7c7cf9c14634444dbab.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-62e04a2d01e9c7c7cf9c14634444dbab.png\"><\/a>\u901a\u8fc7\u4e00\u4e2a\u4efb\u52a1<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-4c0b28ee8fbb04e6a35b4588bac5b147.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-4c0b28ee8fbb04e6a35b4588bac5b147.png\"><\/a><mark data-type=\"concepts\" data-id=\"8ec6a68f-ad96-4b85-ab72-6f8931886922\">\u6620\u5c04<\/mark>\u5230<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-19e5fa5a6d6b57cffdcbb40299cda90f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-19e5fa5a6d6b57cffdcbb40299cda90f.png\"><\/a>\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u6837\u7684\u4e00\u4e2a\u95ee\u9898\uff0c\u672c\u6587\u6784\u5efa\u4e86 prompt \u6765\u8fdb\u884c\u5b66\u4e60\u3002\u8fd9\u91cc\u7684 prompt \u88ab\u8868\u793a\u4e3a\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-3f684541484b8b3323b0e50bfddd0b12.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-3f684541484b8b3323b0e50bfddd0b12.png\"><\/a><\/p>\n<p>\u8bad\u7ec3\u7f51\u7edc\u4e3a\u4e00\u4e2a\u5355\u5c42\u5355\u5934 Transformer\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-02fc52763f6817a3359098b14b595d76.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-02fc52763f6817a3359098b14b595d76.png\"><\/a><\/p>\n<p>\u9884\u8bad\u7ec3\u8fc7\u7a0b\u662f\u6c42\u89e3\u4e00\u4e2a\u5bf9\u6240\u6709\u8bad\u7ec3\u4efb\u52a1\u7684\u7ecf\u9a8c\u98ce\u9669\u6700\u5c0f\u5316 (empirical risk minimization)\u3002<mark data-type=\"concepts\" data-id=\"4c38563a-2d9b-439e-bfb4-21d209eeff3e\">\u635f\u5931\u51fd\u6570<\/mark>\u4f7f\u7528\u7684\u662f\u9002\u5408\u4e8c<mark data-type=\"concepts\" data-id=\"aaeb1c15-66f2-4822-91bb-b441607f9ecf\">\u5206\u7c7b\u95ee\u9898<\/mark>\u7684 Hinge loss\uff0c\u8bad\u7ec3\u7b97\u6cd5\u662f\u968f\u673a<mark data-type=\"tech_methods\" data-id=\"847e8a6a-524c-4d34-a50e-b17cc5dd84d4\">\u68af\u5ea6\u4e0b\u964d<\/mark>\u3002<\/p>\n<p>\u672c\u6587\u5b9a\u4e49\u4e86\u4e24\u79cd ICL \u6cdb\u5316\u7684\u60c5\u51b5\u3002\u4e00\u4e2a\u662f in-domain \u7684\uff0c\u5373\u6cdb\u5316\u7684\u65f6\u5019\u6d4b\u8bd5\u6570\u636e\u7684\u5206\u5e03\u548c\u8bad\u7ec3\u6570\u636e\u4e00\u6837\uff0c\u6ce8\u610f\u8fd9\u4e2a\u60c5\u51b5\u91cc\u9762\u6d4b\u8bd5\u4efb\u52a1\u4e0d\u5fc5\u548c\u8bad\u7ec3\u4efb\u52a1\u4e00\u6837\uff0c\u5373\u8fd9\u91cc\u5df2\u7ecf\u8003\u8651\u4e86\u5bf9\u672a\u89c1\u4efb\u52a1 (unseen task) \u7684\u6cdb\u5316\u3002\u53e6\u4e00\u4e2a\u662f out-of-domain \u7684\uff0c\u5373\u6d4b\u8bd5\u3001\u8bad\u7ec3\u6570\u636e\u5206\u5e03\u4e0d\u4e00\u6837\u3002<\/p>\n<p>\u672c\u6587\u8fd8\u6d89\u53ca\u4e86\u5728 ICL \u63a8\u65ad\u7684\u65f6\u5019\u8fdb\u884c magnitude-based pruning \u7684\u5206\u6790\uff0c\u8fd9\u91cc\u7684<mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u65b9\u5f0f\u662f\u6307\u5bf9\u4e8e\u8bad\u7ec3\u5f97\u5230\u7684\u4e2d\u7684\u5404\u4e2a<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\uff0c\u6839\u636e\u5176\u5e45\u503c\u5927\u5c0f\uff0c\u8fdb\u884c\u4ece\u5c0f\u5230\u5927\u7684\u5220\u9664\u3002<\/p>\n<p><strong>\u5bf9\u6570\u636e\u548c\u4efb\u52a1\u7684\u6784\u5efa<\/strong><\/p>\n<p>\u8fd9\u4e00\u90e8\u5206\u8bf7\u53c2\u8003\u539f\u6587\u7684 Section 3.2\uff0c\u8fd9\u91cc\u53ea\u505a\u4e00\u4e2a\u6982\u8ff0\u3002\u672c\u6587\u7684\u7406\u8bba\u5206\u6790\u662f\u57fa\u4e8e\u6700\u8fd1\u6bd4\u8f83\u706b\u70ed\u7684 feature learning \u8def\u7ebf\uff0c\u5373\u901a\u5e38\u5c06\u6570\u636e\u5047\u8bbe\u4e3a\u53ef\u5206\uff08\u901a\u5e38\u662f\u6b63\u4ea4\uff09\u7684 pattern\uff0c\u4ece\u800c\u63a8\u5bfc\u51fa\u57fa\u4e8e\u4e0d\u540c pattern \u7684\u68af\u5ea6\u53d8\u5316\u3002\u672c\u6587\u9996\u5148\u5b9a\u4e49\u4e86\u4e00\u7ec4 in-domain-relevant (IDR) pattern \u7528\u4e8e\u51b3\u5b9a in-domain \u4efb\u52a1\u7684\u5206\u7c7b\uff0c\u548c\u4e00\u7ec4\u4e0e\u4efb\u52a1\u65e0\u5173\u7684 in-domain-irrelevant (IDI) pattern\uff0c\u8fd9\u4e9b pattern \u4e4b\u95f4\u4e92\u76f8\u6b63\u4ea4\u3002IDR pattern \u6709<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-0d3041e6145b462e1a7b247390aee1f0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-0d3041e6145b462e1a7b247390aee1f0.png\"><\/a>\u4e2a\uff0cIDI pattern \u6709<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-cdb42068a8c931eefe31905c34b9e92b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-cdb42068a8c931eefe31905c34b9e92b.png\"><\/a>\u4e2a\u3002\u4e00\u4e2a<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-8d16319a37e80280c6c465d25ad9a060.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-8d16319a37e80280c6c465d25ad9a060.png\"><\/a>\u88ab\u8868\u793a\u4e3a\u4e00\u4e2a IDR pattern \u548c\u4e00\u4e2a IDI pattern \u7684\u548c\u3002\u4e00\u4e2a in-domain \u4efb\u52a1\u5c31\u88ab\u5b9a\u4e49\u4e3a\u57fa\u4e8e\u67d0\u4e24\u4e2a IDR pattern \u7684<mark data-type=\"concepts\" data-id=\"aaeb1c15-66f2-4822-91bb-b441607f9ecf\">\u5206\u7c7b\u95ee\u9898<\/mark>\u3002<\/p>\n<p>\u7c7b\u4f3c\u5730\uff0c\u672c\u6587\u901a\u8fc7\u5b9a\u4e49 out-of-domain-relevant (ODR) pattern \u548c out-of-domain-irrelevant (ODI) pattern\uff0c\u53ef\u4ee5\u523b\u753b OOD \u6cdb\u5316\u65f6\u5019\u7684\u6570\u636e\u548c\u4efb\u52a1\u3002<\/p>\n<p>\u672c\u6587\u5bf9 prompt \u7684\u8868\u793a\u53ef\u4ee5\u7528\u4e0b\u56fe\u7684\u4f8b\u5b50\u6765\u9610\u8ff0\uff0c\u5176\u4e2d<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-d2a904bff3e827bcca3f35260b889ff7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-d2a904bff3e827bcca3f35260b889ff7.png\"><\/a>\u662f IDR pattern\uff0c<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-c315f052e3db2befe0b1a8c6a1e9a53a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-c315f052e3db2befe0b1a8c6a1e9a53a.png\"><\/a>\u662f IDI pattern\u3002\u8fd9\u91cc\u5728\u505a\u7684\u4efb\u52a1\u662f\u57fa\u4e8e x \u4e2d\u7684<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-755ce9c75737ff308f21433c46100d4a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-755ce9c75737ff308f21433c46100d4a.png\"><\/a>\u505a\u5206\u7c7b\uff0c\u5982\u679c\u662f<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-d67bbb3653984ebd9be47c27a075af8b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-d67bbb3653984ebd9be47c27a075af8b.png\"><\/a>\u90a3\u4e48\u5176\u6807\u7b7e\u4e3a + 1\uff0c\u5bf9\u5e94\u4e8e +q\uff0c\u5982\u679c\u662f<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-9d86ae590192e46b9898707068637c7d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-9d86ae590192e46b9898707068637c7d.png\"><\/a>\u90a3\u4e48\u5176\u6807\u7b7e\u4e3a &#8211; 1\uff0c\u5bf9\u5e94\u4e8e -q\u3002\u03b1\uff0c\u03b1&#8217;\u00a0\u5206\u522b\u88ab\u5b9a\u4e49\u4e3a\u8bad\u7ec3\u548c\u6d4b\u8bd5 prompt \u4e2d\u8ddf query \u7684 IDR\/ODR pattern \u4e00\u6837\u7684\u4e0a\u4e0b\u6587\u793a\u4f8b\u3002\u4e0b\u56fe\u4e2d\u7684\u4f8b\u5b50\u91cc\u9762\uff0c<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-271d1f4161bcf10e66e58440c5bbe2af.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-271d1f4161bcf10e66e58440c5bbe2af.png\"><\/a>\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-bdc621e78209b181abc58adc23e77b7d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-bdc621e78209b181abc58adc23e77b7d.png\"><\/a><\/p>\n<p><strong>\u7406\u8bba\u7ed3\u679c<\/strong><\/p>\n<p>\u9996\u5148\uff0c\u5bf9\u4e8e in-domain \u7684\u60c5\u51b5\uff0c\u672c\u6587\u5148\u7ed9\u4e86\u4e00\u4e2a condition 3.2 \u6765\u89c4\u5b9a\u8bad\u7ec3\u4efb\u52a1\u9700\u8981\u6ee1\u8db3\u7684\u6761\u4ef6\uff0c\u5373\u8bad\u7ec3\u4efb\u52a1\u9700\u8981\u8986\u76d6\u6240\u6709\u7684 IDR pattern \u548c\u6807\u7b7e\u3002\u7136\u540e in-domain \u7684\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-be29af6528a7e0a23f8851149dd6911e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-be29af6528a7e0a23f8851149dd6911e.png\"><\/a><\/p>\n<p>\u8fd9\u91cc\u8868\u660e\uff1a1\uff0c\u8bad\u7ec3\u4efb\u52a1\u7684\u6570\u91cf\u53ea\u9700\u8981\u5728\u5168\u90e8\u4efb\u52a1\u4e2d\u5360\u6bd4\u8fbe\u5230\u6ee1\u8db3 condition 3.2 \u7684\u5c0f\u6bd4\u4f8b\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5bf9 unseen task \u5b9e\u73b0\u5f88\u597d\u7684\u6cdb\u5316\uff1b2\uff0c\u8ddf\u5f53\u524d\u4efb\u52a1\u76f8\u5173\u7684 IDR pattern \u5728 prompt \u4e2d\u7684\u6bd4\u4f8b\u8d8a\u9ad8\uff0c\u5c31\u53ef\u4ee5\u4ee5\u66f4\u5c11\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u8bad\u7ec3\u8fed\u4ee3\u6b21\u6570\uff0c\u4ee5\u53ca\u66f4\u77ed\u7684 training\/testing prompt \u5b9e\u73b0\u7406\u60f3\u7684\u6cdb\u5316\u3002<\/p>\n<p>\u63a5\u4e0b\u6765\u662f out-of-domain \u6cdb\u5316\u7684\u7ed3\u679c\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-272bdba64ee7a827f123281e09b43475.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-272bdba64ee7a827f123281e09b43475.png\"><\/a><\/p>\n<p>\u8fd9\u91cc\u8bf4\u660e\uff0c\u5982\u679c ODR pattern \u662f IDR pattern \u7684\u7ebf\u6027\u7ec4\u5408\u4e14\u7cfb\u6570\u548c\u5927\u4e8e 1\uff0c\u90a3\u4e48\u6b64\u65f6 OOD ICL \u6cdb\u5316\u53ef\u4ee5\u8fbe\u5230\u7406\u60f3\u7684\u6548\u679c\u3002\u8fd9\u4e2a\u7ed3\u679c\u7ed9\u51fa\u4e86\u5728 ICL \u7684\u6846\u67b6\u4e0b\uff0c\u597d\u7684 OOD \u6cdb\u5316\u6240\u9700\u8981\u7684\u8bad\u7ec3\u548c\u6d4b\u8bd5\u6570\u636e\u4e4b\u95f4\u7684\u5185\u5728\u8054\u7cfb\u3002\u8be5\u5b9a\u7406\u4e5f\u901a\u8fc7 <mark data-type=\"tech_methods\" data-id=\"e8e4513e-695d-49e6-9155-31ae453de372\">GPT-2<\/mark> \u7684\u5b9e\u9a8c\u5f97\u5230\u4e86\u9a8c\u8bc1\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5f53 (12) \u4e2d\u7684\u7cfb\u6570\u548c<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-b92c881965b4b79d9c891f5f65208207.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-b92c881965b4b79d9c891f5f65208207.png\"><\/a>\u5927\u4e8e 1 \u7684\u65f6\u5019\uff0cOOD \u5206\u7c7b\u53ef\u4ee5\u8fbe\u5230\u7406\u60f3\u7684\u7ed3\u679c\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5f53<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-80e614fb328c20bed52b034a71b0bb44.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-80e614fb328c20bed52b034a71b0bb44.png\"><\/a>\uff0c\u5373 prompt \u4e2d\u548c\u5206\u7c7b\u4efb\u52a1\u76f8\u5173\u7684 ODR\/IDR pattern \u6bd4\u4f8b\u8d8a\u9ad8\u7684\u65f6\u5019\uff0c\u6240\u9700\u8981\u7684 context \u957f\u5ea6\u8d8a\u5c0f\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-426f2344cec605cf18841539c3b44738.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-426f2344cec605cf18841539c3b44738.png\"><\/a><\/p>\n<p>\u7136\u540e\uff0c\u672c\u6587\u7ed9\u51fa\u4e86\u5e26\u6709 magnitude-based pruning \u7684 ICL \u6cdb\u5316\u7ed3\u679c\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-064f44e17bd937a34efc4b71196b5fa9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-064f44e17bd937a34efc4b71196b5fa9.png\"><\/a><\/p>\n<p>\u8fd9\u4e2a\u7ed3\u679c\u8868\u660e\uff0c\u9996\u5148\uff0c\u8bad\u7ec3\u5f97\u5230\u7684<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-c24afb751f05620aa41318b55411ba23.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-c24afb751f05620aa41318b55411ba23.png\"><\/a>\u4e2d\u6709\u4e00\u90e8\u5206\uff08\u5e38\u6570\u6bd4\u4f8b\uff09<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u7684\u5e45\u503c\u5f88\u5c0f\uff0c\u800c\u5269\u4e0b\u7684\u76f8\u5bf9\u6bd4\u8f83\u5927\uff08\u516c\u5f0f 14\uff09\u3002\u5f53\u6211\u4eec\u53ea\u679d\u526a\u5c0f<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u7684\u65f6\u5019\uff0c\u5bf9\u6cdb\u5316\u7ed3\u679c\u57fa\u672c\u6ca1\u6709\u5f71\u54cd\uff0c\u800c\u5f53\u679d\u526a\u6bd4\u4f8b\u589e\u52a0\u5230\u8981\u526a\u5927<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u7684\u65f6\u5019\uff0c\u6cdb\u5316\u8bef\u5dee\u4f1a\u968f\u4e4b\u663e\u8457\u53d8\u5927\uff08\u516c\u5f0f 15\uff0c16\uff09\u3002\u4ee5\u4e0b\u5b9e\u9a8c\u9a8c\u8bc1\u4e86\u5b9a\u7406 3.7\u3002\u4e0b\u56fe A \u4e2d\u6d45\u84dd\u8272\u7684\u7ad6\u7ebf\u8868\u793a\u8bad\u7ec3\u5f97\u5230\u7684<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-c53b5415846fc9804626ad97e1233ccc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-c53b5415846fc9804626ad97e1233ccc.png\"><\/a>\u5448\u73b0\u51fa\u4e86\u516c\u5f0f 14 \u7684\u7ed3\u679c\u3002\u800c\u5bf9\u5c0f<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u8fdb\u884c\u679d\u526a\u4e0d\u4f1a\u4f7f\u6cdb\u5316\u53d8\u5dee\uff0c\u8fd9\u4e2a\u7ed3\u679c\u7b26\u5408\u7406\u8bba\u3002\u56fe B \u53cd\u6620\u51fa\u5f53 prompt \u4e2d\u548c\u4efb\u52a1\u76f8\u5173\u7684\u4e0a\u4e0b\u6587\u8d8a\u591a\u7684\u65f6\u5019\uff0c\u6211\u4eec\u53ef\u4ee5\u5141\u8bb8\u66f4\u5927\u7684\u679d\u526a\u6bd4\u4f8b\u4ee5\u8fbe\u5230\u76f8\u540c\u7684\u6cdb\u5316\u6027\u80fd\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f626413d04a253b03aff180acc6c757f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f626413d04a253b03aff180acc6c757f.png\"><\/a><\/p>\n<p><strong>ICL \u673a\u5236<\/strong><\/p>\n<p>\u901a\u8fc7\u5bf9\u9884\u8bad\u7ec3\u8fc7\u7a0b\u7684\u523b\u753b\uff0c\u672c\u6587\u5f97\u5230\u4e86\u5355\u5c42\u5355\u5934\u975e\u7ebf\u6027 Transformer \u505a ICL \u7684\u5185\u5728\u673a\u5236\uff0c\u8fd9\u4e00\u90e8\u5206\u5728\u539f\u6587\u7684 Section 4\u3002\u8be5\u8fc7\u7a0b\u53ef\u4ee5\u7528\u4e0b\u56fe\u8868\u793a\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-5f061f1cea8cacbb5229ee324c767122.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-5f061f1cea8cacbb5229ee324c767122.png\"><\/a><\/p>\n<p>\u7b80\u800c\u8a00\u4e4b\uff0cattention \u5c42\u4f1a\u9009\u62e9\u548c query \u7684 ODR\/IDR pattern \u4e00\u6837\u7684\u4e0a\u4e0b\u6587\uff0c\u8d4b\u4e88\u5b83\u4eec\u51e0\u4e4e\u5168\u90e8 attention <mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\uff0c\u7136\u540e MLP \u5c42\u4f1a\u91cd\u70b9\u6839\u636e attention \u5c42\u8f93\u51fa\u4e2d\u7684\u6807\u7b7e\u5d4c\u5165\u6765\u4f5c\u51fa\u6700\u540e\u7684\u5206\u7c7b\u3002<\/p>\n<p><strong>\u603b\u7ed3<\/strong><\/p>\n<p>\u672c\u6587\u8bb2\u89e3\u4e86\u5728 ICL \u5f53\u4e2d\uff0c\u975e\u7ebf\u6027 Transformer \u7684\u8bad\u7ec3\u673a\u5236\uff0c\u4ee5\u53ca\u5bf9\u4e8e\u65b0\u4efb\u52a1\u548c\u5206\u5e03\u504f\u79fb\u6570\u636e\u7684\u6cdb\u5316\u80fd\u529b\u3002\u7406\u8bba\u7ed3\u679c\u5bf9\u4e8e\u8bbe\u8ba1 prompt \u9009\u62e9\u7b97\u6cd5\u548c LLM <mark data-type=\"tech_tasks\" data-id=\"30ffd95b-36fa-4e6f-b863-926410835b10\">\u526a\u679d<\/mark>\u7b97\u6cd5\u6709\u4e00\u5b9a\u5b9e\u9645\u610f\u4e49\u3002<\/p>\n<p><em><sup>\u53c2\u8003\u6587\u732e<\/sup><\/em><\/p>\n<p><sup><em>[1] Garg, et al., Neurips 2022. &#8220;What can transformers learn in-context? a case study of simple function classes.&#8221;<\/em><\/sup><\/p>\n<p><sup><em>[2] Von Oswald et al., ICML 2023. &#8220;Transformers learn in-context by gradient descent.&#8221;<\/em><\/sup><\/p>\n<p><sup><em>[3] Zhang et al., JMLR 2024. &#8220;Trained transformers learn linear models in-context.&#8221;<\/em><\/sup><\/p>\n<p><em><sup>[4] Huang et al., ICML 2024. &#8220;In-context convergence of transformers.&#8221;<\/sup><\/em><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-06-28-7\" target=\"_blank\">ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:ICML 2024 | [&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":[55,65],"class_list":["post-34299","post","type-post","status-publish","format-standard","hentry","category-news","tag-55","tag-65"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236 - \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=34299\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34299\",\"url\":\"https:\/\/17aitech.com\/?p=34299\",\"name\":\"ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34299#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34299#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f2ef263020affc3831b835d31c350f04.png\",\"datePublished\":\"2024-12-03T18:56:53+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34299#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34299\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34299#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f2ef263020affc3831b835d31c350f04.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f2ef263020affc3831b835d31c350f04.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34299#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236\"}]},{\"@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":"ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236 - \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=34299","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34299","url":"https:\/\/17aitech.com\/?p=34299","name":"ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34299#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34299#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f2ef263020affc3831b835d31c350f04.png","datePublished":"2024-12-03T18:56:53+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34299#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34299"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34299#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f2ef263020affc3831b835d31c350f04.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-f2ef263020affc3831b835d31c350f04.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34299#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"ICML 2024 | \u63ed\u793a\u975e\u7ebf\u5f62Transformer\u5728\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e2d\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u673a\u5236"}]},{"@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\/34299","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=34299"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34299\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34299"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34299"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34299"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}