{"id":601,"date":"2025-11-21T12:01:58","date_gmt":"2025-11-21T04:01:58","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=601"},"modified":"2025-11-28T09:29:19","modified_gmt":"2025-11-28T01:29:19","slug":"efficient-transformers-a-survey","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=601","title":{"rendered":"Efficient Transformers: A Survey"},"content":{"rendered":"\n<p><strong>\u4f1a\u8bae<\/strong>\uff1aICML 2025<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong><mark class=\"has-inline-color has-black-color\">\u8bba\u6587\u6982\u8ff0<\/mark><\/strong><\/h2>\n\n\n\n<p>\u968f\u7740Transformer\u67b6\u6784\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u7b49\u9886\u57df\u7684\u5e7f\u6cdb\u5e94\u7528\uff0c\u5176\u8ba1\u7b97\u548c\u5185\u5b58\u5f00\u9500\u6210\u4e3a\u4e86\u9650\u5236\u5176\u8fdb\u4e00\u6b65\u5e94\u7528\u7684\u74f6\u9888\uff0c\u5c24\u5176\u662f\u5728\u957f\u6587\u672c\u6216\u5927\u89c4\u6a21\u6570\u636e\u96c6\u7684\u4efb\u52a1\u4e2d\u3002\u672c\u6587\u63d0\u51fa\u4e86\u51e0\u79cd\u65b0\u7684Transformer\u53d8\u4f53\uff0c\u901a\u8fc7\u4f18\u5316\u8ba1\u7b97\u8def\u5f84\u548c\u6539\u8fdb\u6ce8\u610f\u529b\u673a\u5236\u6765\u63d0\u9ad8\u5176\u8ba1\u7b97\u6548\u7387\uff0c\u65e8\u5728\u964d\u4f4eTransformer\u6a21\u578b\u7684\u65f6\u95f4\u590d\u6742\u5ea6\u548c\u7a7a\u95f4\u590d\u6742\u5ea6\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u80cc\u666f<\/strong><strong><\/strong><\/h2>\n\n\n\n<p>Transformer\u6a21\u578b\u51ed\u501f\u5176\u51fa\u8272\u7684\u6027\u80fd\uff0c\u5c24\u5176\u662f\u5728\u5e8f\u5217\u5efa\u6a21\u4efb\u52a1\u4e2d\uff08\u5982\u673a\u5668\u7ffb\u8bd1\u3001\u6587\u672c\u751f\u6210\u7b49\uff09\uff0c\u5728\u8fd1\u5e74\u6765\u6210\u4e3a\u6df1\u5ea6\u5b66\u4e60\u7684\u4e3b\u6d41\u67b6\u6784\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"830\" height=\"536\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-13.png\"  class=\"wp-image-603\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-13.png 830w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-13-300x194.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-13-768x496.png 768w\" sizes=\"auto, (max-width: 830px) 100vw, 830px\" title=\"Efficient Transformers: A Survey\u63d2\u56fe\" alt=\"Efficient Transformers: A Survey\u63d2\u56fe\" \/><\/figure>\n\n\n\n<p>\u7136\u800c\uff0cTransformer\u7684\u6838\u5fc3\u8ba1\u7b97\u2014\u2014\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09\u5177\u6709<strong>O(n\u00b2)\u7684\u65f6\u95f4\u548c\u7a7a\u95f4\u590d\u6742\u5ea6\uff0c\u5176\u4e2dn<\/strong>\u662f\u8f93\u5165\u5e8f\u5217\u7684\u957f\u5ea6\u3002\u5bf9\u4e8e\u957f\u5e8f\u5217\uff0c\u8fd9\u79cd\u590d\u6742\u5ea6\u5c06\u5bfc\u81f4\u4e25\u91cd\u7684\u8ba1\u7b97\u8d44\u6e90\u548c\u5185\u5b58\u5360\u7528\u95ee\u9898\u3002\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u8bb8\u591a\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u4e0d\u540c\u7684\u9ad8\u6548Transformer\u53d8\u4f53\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u4e3b\u8981\u8d21\u732e<\/strong><\/h2>\n\n\n\n<p>\u8bba\u6587\u7684\u6838\u5fc3\u8d21\u732e\u662f\u63d0\u51fa\u51e0\u79cd\u9ad8\u6548\u7684Transformer\u67b6\u6784\u548c\u76f8\u5e94\u7684\u4f18\u5316\u7b97\u6cd5\uff0c\u65e8\u5728\u51cf\u5c11\u8ba1\u7b97\u548c\u5185\u5b58\u5f00\u9500\uff0c\u540c\u65f6\u4fdd\u6301\u6216\u63d0\u9ad8\u6a21\u578b\u7684\u51c6\u786e\u6027\u3002\u5177\u4f53\u8d21\u732e\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p><strong>1.\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\uff08Sparse Attention Mechanism\uff09<\/strong><\/p>\n\n\n\n<p>\u80cc\u666f\uff1a\u4f20\u7edf\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u5bf9\u6bcf\u5bf9\u8bcd\u7684\u76f8\u4f3c\u6027\u8fdb\u884c\u8ba1\u7b97\uff0c\u5bfc\u81f4\u590d\u6742\u5ea6\u4e3aO(n\u00b2)\uff0c\u5c24\u5176\u662f\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\uff0c\u5185\u5b58\u548c\u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u5de8\u5927\u3002<\/p>\n\n\n\n<p>\u7a00\u758f\u5316\u7b56\u7565\uff1a<\/p>\n\n\n\n<p>\u5c40\u90e8\u6ce8\u610f\u529b\uff08Local Attention\uff09\uff1a\u53ea\u8003\u8651\u76f8\u90bb\u8bcd\u7684\u6ce8\u610f\u529b\uff0c\u51cf\u5c11\u8ba1\u7b97\u91cf\u3002\u8fd9\u79cd\u65b9\u6cd5\u9002\u7528\u4e8e\u5e8f\u5217\u4e2d\u7684\u5c40\u90e8\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/p>\n\n\n\n<p>\u5757\u72b6\u6ce8\u610f\u529b\uff08Block Sparse Attention\uff09\uff1a\u5c06\u8f93\u5165\u5e8f\u5217\u5206\u6210\u591a\u4e2a\u5757\uff0c\u6bcf\u4e2a\u5757\u5185\u7684\u8bcd\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u5757\u4e4b\u95f4\u8fdb\u884c\u8f83\u4f4e\u590d\u6742\u5ea6\u7684\u4ea4\u4e92\u3002\u8fd9\u79cd\u65b9\u6cd5\u9002\u7528\u4e8e\u5177\u6709\u5c40\u90e8\u4f9d\u8d56\u7684\u4efb\u52a1\uff0c\u5982\u6587\u672c\u5206\u7c7b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"831\" height=\"433\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-14.png\"  class=\"wp-image-604\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-14.png 831w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-14-300x156.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-14-768x400.png 768w\" sizes=\"auto, (max-width: 831px) 100vw, 831px\" title=\"Efficient Transformers: A Survey\u63d2\u56fe1\" alt=\"Efficient Transformers: A Survey\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<p>\u56fa\u5b9a\u6a21\u5f0f\u7a00\u758f\uff08Fixed Sparse Patterns\uff09\uff1a\u4f8b\u5982\uff0c\u91c7\u7528&#8221;\u6ed1\u52a8\u7a97\u53e3&#8221;\u7b56\u7565\uff0c\u6bcf\u4e2a\u8bcd\u53ea\u4e0e\u76f8\u90bb\u7684\u8bcd\u5efa\u7acb\u6ce8\u610f\u529b\uff0c\u800c\u4e0d\u662f\u4e0e\u6240\u6709\u8bcd\u8fdb\u884c\u4ea4\u4e92\u3002<\/p>\n\n\n\n<p>\u6548\u679c\uff1a\u8fd9\u4e9b\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u901a\u8fc7\u51cf\u5c11\u8ba1\u7b97\u91cf\u548c\u5185\u5b58\u6d88\u8017\uff0c\u4f7f\u5f97Transformer\u80fd\u591f\u5904\u7406\u66f4\u957f\u7684\u8f93\u5165\u5e8f\u5217\uff0c\u65f6\u95f4\u590d\u6742\u5ea6\u964d\u5230O(n log n)\u6216O(n)\u3002<\/p>\n\n\n\n<p><strong>2.\u4f4e\u79e9\u8fd1\u4f3c\uff08Low-rank Approximation\uff09<\/strong><\/p>\n\n\n\n<p>\u80cc\u666f\uff1a\u6ce8\u610f\u529b\u77e9\u9635\u662f\u4e00\u4e2a\u5bc6\u96c6\u7684\u77e9\u9635\uff0c\u4f20\u7edf\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u8ba1\u7b97\u91cf\u5927\uff0c\u5c24\u5176\u662f\u5f53\u8f93\u5165\u957f\u5ea6\u5f88\u957f\u65f6\uff0c\u8ba1\u7b97\u5f00\u9500\u4e25\u91cd\u3002<\/p>\n\n\n\n<p>\u4f4e\u79e9\u8fd1\u4f3c\u65b9\u6cd5\uff1a\u901a\u8fc7\u5bf9\u6ce8\u610f\u529b\u77e9\u9635\u8fdb\u884c\u4f4e\u79e9\u8fd1\u4f3c\uff0c\u51cf\u5c11\u77e9\u9635\u7684\u7ef4\u5ea6\uff0c\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6\u548c\u5185\u5b58\u6d88\u8017\u3002\u4f4e\u79e9\u8fd1\u4f3c\u65b9\u6cd5\u901a\u8fc7\u5c06\u6ce8\u610f\u529b\u77e9\u9635\u5206\u89e3\u6210\u4f4e\u79e9\u77e9\u9635\uff0c\u4ece\u800c\u51cf\u5c11\u4e86\u8ba1\u7b97\u91cf\u3002<\/p>\n\n\n\n<p>\u4ee3\u8868\u6027\u6a21\u578b\uff1aLinformer\uff1a\u901a\u8fc7\u7ebf\u6027\u6295\u5f71\u5c06 K \u548c V \u7684\u5e8f\u5217\u957f\u5ea6\u7ef4\u5ea6\u4ece n \u6295\u5f71\u5230\u4e00\u4e2a\u66f4\u4f4e\u7684 k \u7ef4\uff08k &lt;&lt; n\uff09\u3002\u8fd9\u6837\uff0cQK\u1d40 \u7684\u8ba1\u7b97\u5c31\u53d8\u6210\u4e86 (Q)(K\u1d40E)\uff0c\u5f97\u5230\u4e00\u4e2a n x k \u7684\u77e9\u9635\uff0c\u590d\u6742\u5ea6\u964d\u81f3 O(nk)\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"830\" height=\"171\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-15.png\"  class=\"wp-image-605\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-15.png 830w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-15-300x62.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-15-768x158.png 768w\" sizes=\"auto, (max-width: 830px) 100vw, 830px\" title=\"Efficient Transformers: A Survey\u63d2\u56fe2\" alt=\"Efficient Transformers: A Survey\u63d2\u56fe2\" \/><\/figure>\n\n\n\n<p>\u6548\u679c\uff1a\u8fd9\u79cd\u65b9\u6cd5\u51cf\u5c11\u4e86\u591a\u4f59\u7684\u8ba1\u7b97\uff0c\u63d0\u9ad8\u4e86\u957f\u5e8f\u5217\u6570\u636e\u5904\u7406\u7684\u6548\u7387\u3002<\/p>\n\n\n\n<p><strong>3.\u56e0\u5f0f\u5206\u89e3\u81ea\u6ce8\u610f\u529b\uff08Factorized Self-attention\uff09<\/strong><\/p>\n\n\n\n<p>\u80cc\u666f\uff1a\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u6d89\u53ca\u5bf9\u6240\u6709\u8bcd\u4e4b\u95f4\u7684\u4ea4\u4e92\u8fdb\u884c\u8ba1\u7b97\uff0c\u8fd9\u5bfc\u81f4\u9ad8\u6602\u7684\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n\n\n\n<p>\u56e0\u5f0f\u5206\u89e3\u65b9\u6cd5\uff1a\u901a\u8fc7\u5c06\u6ce8\u610f\u529b\u77e9\u9635\u5206\u89e3\u6210\u591a\u4e2a\u56e0\u5b50\u6765\u7b80\u5316\u8ba1\u7b97\u3002\u4f8b\u5982\uff0c\u91c7\u7528Kronecker\u79ef\uff08Kronecker product\uff09\u6216\u8005\u5176\u4ed6\u56e0\u5f0f\u5206\u89e3\u6280\u672f\uff0c\u5c06\u8ba1\u7b97\u4efb\u52a1\u5206\u89e3\u6210\u66f4\u5c0f\u7684\u5757\u6765\u52a0\u901f\u8ba1\u7b97\u3002<\/p>\n\n\n\n<p>\u6548\u679c\uff1a\u8fd9\u79cd\u65b9\u6cd5\u80fd\u591f\u51cf\u5c11\u8ba1\u7b97\u91cf\uff0c\u540c\u65f6\u4e0d\u5f71\u54cd\u6a21\u578b\u7684\u8868\u73b0\uff0c\u5c24\u5176\u662f\u5728\u5904\u7406\u8f83\u957f\u6587\u672c\u65f6\uff0c\u8868\u73b0\u5c24\u4e3a\u7a81\u51fa\u3002<\/p>\n\n\n\n<p><strong>4.\u4f4e\u79e9\u5f20\u91cf\u5206\u89e3\uff08Low-rank Tensor Decomposition\uff09<\/strong><\/p>\n\n\n\n<p>\u80cc\u666f\uff1a\u4e3a\u4e86\u8fdb\u4e00\u6b65\u4f18\u5316\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4f4e\u79e9\u5f20\u91cf\u5206\u89e3\u7684\u6280\u672f\uff0c\u5c06\u591a\u7ef4\u6ce8\u610f\u529b\u77e9\u9635\u5206\u89e3\u6210\u4f4e\u79e9\u5f20\u91cf\uff0c\u4ee5\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n\n\n\n<p>\u5e94\u7528\uff1a\u4f4e\u79e9\u5f20\u91cf\u5206\u89e3\u5728\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u7684\u540c\u65f6\uff0c\u4fdd\u7559\u4e86\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\uff0c\u5c24\u5176\u662f\u5728\u591a\u6a21\u6001\u4efb\u52a1\uff08\u5982\u89c6\u89c9-\u8bed\u8a00\u8054\u5408\u4efb\u52a1\uff09\u4e2d\uff0c\u5f20\u91cf\u5206\u89e3\u80fd\u591f\u6709\u6548\u5730\u51cf\u5c11\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n\n\n\n<p><strong>5.\u7a00\u758f\u77e9\u9635\u4e58\u6cd5\uff08Sparse Matrix Multiplication\uff09<\/strong><\/p>\n\n\n\n<p>\u80cc\u666f\uff1aTransformer\u6a21\u578b\u9700\u8981\u5927\u91cf\u7684\u77e9\u9635\u4e58\u6cd5\u64cd\u4f5c\uff0c\u5c24\u5176\u662f\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\uff0c\u4f20\u7edf\u7684\u5bc6\u96c6\u77e9\u9635\u4e58\u6cd5\u4f1a\u6d88\u8017\u5927\u91cf\u7684\u8ba1\u7b97\u8d44\u6e90\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"831\" height=\"562\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-16.png\"  class=\"wp-image-607\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-16.png 831w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-16-300x203.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/11\/image-16-768x519.png 768w\" sizes=\"auto, (max-width: 831px) 100vw, 831px\" title=\"Efficient Transformers: A Survey\u63d2\u56fe3\" alt=\"Efficient Transformers: A Survey\u63d2\u56fe3\" \/><\/figure>\n\n\n\n<p>\u7a00\u758f\u77e9\u9635\u4f18\u5316\uff1a\u4f7f\u7528\u7a00\u758f\u77e9\u9635\u4e58\u6cd5\u6765\u4f18\u5316\u8ba1\u7b97\u8fc7\u7a0b\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u5177\u6709\u7a00\u758f\u6ce8\u610f\u529b\u7684Transformer\u6a21\u578b\u3002\u901a\u8fc7\u4f18\u5316\u7a00\u758f\u77e9\u9635\u7684\u5b58\u50a8\u548c\u8ba1\u7b97\u65b9\u5f0f\uff0c\u80fd\u591f\u663e\u8457\u51cf\u5c11\u5185\u5b58\u5360\u7528\u548c\u8ba1\u7b97\u65f6\u95f4\u3002<\/p>\n\n\n\n<p>\u6548\u679c\uff1a\u8fd9\u79cd\u4f18\u5316\u65b9\u6cd5\u52a0\u901f\u4e86\u77e9\u9635\u8fd0\u7b97\uff0c\u51cf\u5c11\u4e86\u5185\u5b58\u6d88\u8017\uff0c\u7279\u522b\u662f\u5728\u8ba1\u7b97\u5927\u89c4\u6a21Transformer\u65f6\u663e\u8457\u63d0\u9ad8\u4e86\u6548\u7387\u3002<\/p>\n\n\n\n<p><strong>6.\u56e0\u679c\u5377\u79ef\uff08Causal Convolutions\uff09<\/strong><\/p>\n\n\n\n<p>\u80cc\u666f\uff1a\u5728\u4e00\u4e9b\u4efb\u52a1\u4e2d\uff0c\u8f93\u5165\u5e8f\u5217\u662f\u6709\u65f6\u95f4\u987a\u5e8f\u7684\uff0c\u5c24\u5176\u662f\u5728NLP\u548c\u65f6\u95f4\u5e8f\u5217\u4efb\u52a1\u4e2d\u3002<\/p>\n\n\n\n<p>\u56e0\u679c\u5377\u79ef\uff1a\u91c7\u7528\u56e0\u679c\u5377\u79ef\u4ee3\u66ff\u5168\u8fde\u63a5\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4ece\u800c\u51cf\u5c11\u8ba1\u7b97\u590d\u6742\u5ea6\u3002\u56e0\u679c\u5377\u79ef\u5728\u65f6\u5e8f\u6570\u636e\u4e2d\u5e7f\u6cdb\u5e94\u7528\uff0c\u80fd\u591f\u5229\u7528\u524d\u5e8f\u4fe1\u606f\u6765\u751f\u6210\u5f53\u524d\u7684\u8f93\u51fa\uff0c\u907f\u514d\u4e86\u5168\u5c40\u8ba1\u7b97\u3002<\/p>\n\n\n\n<p>\u6548\u679c\uff1a\u8be5\u65b9\u6cd5\u6709\u6548\u51cf\u5c11\u4e86\u4e0d\u5fc5\u8981\u7684\u8ba1\u7b97\uff0c\u540c\u65f6\u63d0\u9ad8\u4e86\u65f6\u5e8f\u4efb\u52a1\u7684\u6548\u7387\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u7b97\u6cd5\u548c\u4f18\u5316<\/h2>\n\n\n\n<p>\u4f5c\u8005\u9488\u5bf9\u6bcf\u79cd\u6539\u8fdb\u63d0\u51fa\u4e86\u8be6\u7ec6\u7684\u7b97\u6cd5\u548c\u4f18\u5316\u6d41\u7a0b\uff0c\u5305\u62ec\u5982\u4f55\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u52a8\u6001\u5730\u9009\u62e9\u7a00\u758f\u6a21\u5f0f\uff0c\u5982\u4f55\u5229\u7528GPU\u786c\u4ef6\u52a0\u901f\u7a00\u758f\u77e9\u9635\u7684\u8ba1\u7b97\u7b49\u3002\u6b64\u5916\uff0c\u8bba\u6587\u8fd8\u63d0\u4f9b\u4e86\u9ad8\u6548Transformer\u67b6\u6784\u7684\u5b9e\u73b0\u4ee3\u7801\uff0c\u5e76\u5c55\u793a\u4e86\u5176\u5728\u591a\u4e2a\u6807\u51c6\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5b9e\u9a8c\u7ed3\u679c<\/h2>\n\n\n\n<p>\u4f5c\u8005\u901a\u8fc7\u4e00\u7cfb\u5217\u5b9e\u9a8c\u5bf9\u8fd9\u4e9b\u65b0\u67b6\u6784\u8fdb\u884c\u4e86\u9a8c\u8bc1\uff0c\u7ed3\u679c\u8868\u660e\uff0c\u63d0\u51fa\u7684\u9ad8\u6548Transformer\u67b6\u6784\u5728\u591a\u79cd\u4efb\u52a1\u4e0a\uff08\u5982\u673a\u5668\u7ffb\u8bd1\u3001\u6587\u672c\u751f\u6210\u3001\u957f\u6587\u672c\u5904\u7406\uff09\u5747\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\uff0c\u5e76\u4e14\u8ba1\u7b97\u6548\u7387\u5927\u5e45\u63d0\u9ad8\u3002\u5177\u4f53\u5b9e\u9a8c\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p>\u957f\u6587\u672c\u5904\u7406\uff1a\u5728\u5904\u7406\u957f\u5ea6\u4e3a5000\u7684\u957f\u6587\u672c\u65f6\uff0c\u91c7\u7528\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u7684Transformer\u6a21\u578b\u76f8\u8f83\u4e8e\u6807\u51c6Transformer\uff0c\u8bad\u7ec3\u65f6\u95f4\u51cf\u5c11\u4e86\u8fd150%\uff0c\u4e14\u5185\u5b58\u6d88\u8017\u51cf\u5c11\u4e8660%\u3002<\/p>\n\n\n\n<p>\u673a\u5668\u7ffb\u8bd1\uff1a\u5728WMT-14\u82f1\u5fb7\u7ffb\u8bd1\u4efb\u52a1\u4e0a\uff0c\u57fa\u4e8e\u7a00\u758f\u6ce8\u610f\u529b\u7684Transformer\u6a21\u578b\u7684BLEU\u5206\u6570\u4ec5\u7a0d\u900a\u4e8e\u4f20\u7edfTransformer\uff0c\u4f46\u8bad\u7ec3\u901f\u5ea6\u548c\u63a8\u7406\u901f\u5ea6\u63d0\u5347\u4e8640%\u4ee5\u4e0a\u3002<\/p>\n\n\n\n<p>\u6587\u672c\u751f\u6210\uff1a\u5728GPT-2\u7684\u6587\u672c\u751f\u6210\u4efb\u52a1\u4e2d\uff0c\u63d0\u51fa\u7684\u9ad8\u6548\u67b6\u6784\u4e0d\u4ec5\u63d0\u9ad8\u4e86\u751f\u6210\u901f\u5ea6\uff0c\u800c\u4e14\u4fdd\u6301\u4e86\u4e0e\u539f\u59cbGPT-2\u76f8\u5f53\u7684\u751f\u6210\u8d28\u91cf\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u521b\u65b0\u6027\u4e0e\u610f\u4e49<\/h2>\n\n\n\n<p>\u9ad8\u6548\u7684\u8ba1\u7b97\uff1a\u901a\u8fc7\u7a00\u758f\u6ce8\u610f\u529b\u548c\u4f4e\u79e9\u8fd1\u4f3c\u7b49\u6280\u672f\uff0c\u663e\u8457\u51cf\u5c11\u4e86Transformer\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u7279\u522b\u662f\u5728\u5904\u7406\u957f\u6587\u672c\u548c\u5927\u89c4\u6a21\u6570\u636e\u96c6\u65f6\u8868\u73b0\u5c24\u4e3a\u7a81\u51fa\u3002<\/p>\n\n\n\n<p>\u53ef\u6269\u5c55\u6027\uff1a\u8fd9\u4e9b\u4f18\u5316\u4e0d\u4ec5\u80fd\u591f\u5728\u73b0\u6709\u4efb\u52a1\u4e2d\u63d0\u5347\u6548\u7387\uff0c\u800c\u4e14\u5177\u6709\u5f88\u597d\u7684\u53ef\u6269\u5c55\u6027\uff0c\u9002\u7528\u4e8e\u66f4\u591a\u7684\u4e0b\u6e38\u4efb\u52a1\u548c\u5e94\u7528\u573a\u666f\u3002<\/p>\n\n\n\n<p>\u786c\u4ef6\u53cb\u597d\uff1a\u901a\u8fc7\u9488\u5bf9\u7a00\u758f\u77e9\u9635\u548c\u8ba1\u7b97\u8fc7\u7a0b\u7684\u4f18\u5316\uff0c\u8fd9\u4e9b\u65b0\u67b6\u6784\u53ef\u4ee5\u6709\u6548\u5229\u7528\u73b0\u4ee3GPU\u548cTPU\u786c\u4ef6\uff0c\u63d0\u5347\u6a21\u578b\u5728\u5b9e\u9645\u90e8\u7f72\u4e2d\u7684\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u7ed3\u8bba<\/h2>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587\u4e3a\u9ad8\u6548Transformer\u67b6\u6784\u7684\u53d1\u5c55\u63d0\u4f9b\u4e86\u65b0\u7684\u65b9\u5411\uff0c\u901a\u8fc7\u521b\u65b0\u7684\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u548c\u4f4e\u79e9\u8fd1\u4f3c\u65b9\u6cd5\uff0c\u663e\u8457\u63d0\u9ad8\u4e86Transformer\u5728\u957f\u5e8f\u5217\u5904\u7406\u4e2d\u7684\u6548\u7387\uff0c\u51cf\u5c11\u4e86\u8ba1\u7b97\u8d44\u6e90\u7684\u6d88\u8017\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u8f83\u9ad8\u7684\u6027\u80fd\u3002\u8fd9\u4e9b\u6280\u672f\u7684\u63d0\u51fa\uff0c\u4e0d\u4ec5\u6709\u52a9\u4e8e\u52a0\u901fTransformer\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\uff0c\u8fd8\u6709\u52a9\u4e8e\u6269\u5c55\u5176\u5e94\u7528\u8303\u56f4\uff0c\u7279\u522b\u662f\u5728\u9700\u8981\u5904\u7406\u8d85\u957f\u5e8f\u5217\u7684\u4efb\u52a1\u4e2d\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f1a\u8bae\uff1aICML 2025 \u8bba\u6587\u6982\u8ff0 \u968f\u7740Transformer\u67b6\u6784\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u7b49\u9886\u57df\u7684\u5e7f\u6cdb\u5e94\u7528\uff0c\u5176\u8ba1\u7b97\u548c\u5185\u5b58\u5f00\u9500\u6210\u4e3a\u4e86\u9650\u5236\u5176\u8fdb\u4e00\u6b65\u5e94\u7528\u7684\u74f6\u9888\uff0c\u5c24\u5176\u662f\u5728\u957f\u6587\u672c\u6216\u5927\u89c4\u6a21\u6570\u636e\u96c6\u7684\u4efb\u52a1\u4e2d\u3002\u672c\u6587\u63d0\u51fa\u4e86\u51e0\u79cd\u65b0\u7684Transformer\u53d8\u4f53\uff0c\u901a\u8fc7\u4f18\u5316\u8ba1\u7b97\u8def\u5f84\u548c\u6539\u8fdb\u6ce8\u610f\u529b\u673a\u5236\u6765\u63d0\u9ad8\u5176\u8ba1\u7b97\u6548\u7387\uff0c\u65e8\u5728\u964d\u4f4eTransformer\u6a21\u578b\u7684\u65f6\u95f4\u590d\u6742\u5ea6\u548c\u7a7a\u95f4\u590d\u6742\u5ea6\u3002 \u80cc\u666f Transformer\u6a21\u578b\u51ed\u501f\u5176\u51fa\u8272\u7684\u6027\u80fd\uff0c\u5c24\u5176\u662f\u5728\u5e8f\u5217\u5efa\u6a21\u4efb\u52a1\u4e2d\uff08\u5982\u673a\u5668\u7ffb\u8bd1\u3001\u6587\u672c\u751f\u6210\u7b49\uff09\uff0c\u5728\u8fd1\u5e74\u6765\u6210\u4e3a\u6df1\u5ea6\u5b66\u4e60\u7684\u4e3b\u6d41\u67b6 &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=601\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":603,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-601","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rengongzhineng"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/601","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=601"}],"version-history":[{"count":3,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/601\/revisions"}],"predecessor-version":[{"id":618,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/601\/revisions\/618"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/603"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=601"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=601"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=601"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}