{"id":39414,"date":"2025-03-14T23:02:28","date_gmt":"2025-03-14T15:02:28","guid":{"rendered":"https:\/\/17aitech.com\/?p=39414"},"modified":"2025-03-14T23:02:28","modified_gmt":"2025-03-14T15:02:28","slug":"%e6%b2%a1%e6%9c%89%e5%bd%92%e4%b8%80%e5%8c%96%e5%b1%82%e7%9a%84transformer%ef%bc%81%e5%88%98%e5%a3%ae%e5%b8%a6%e9%98%9f%ef%bc%8c%e4%bd%95%e6%81%ba%e6%98%8e%e3%80%81yann-lecun%e9%83%bd%e5%8f%82","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=39414","title":{"rendered":"\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-14-11\" target=\"_blank\">\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86<\/a><\/p>\n<section>\u4f55\u607a\u660e\u53c8\u53cc\u53d2\u53d5\u53d1\u65b0\u4f5c\u4e86\uff0c\u8fd9\u6b21\u8fd8\u662f\u4e0e\u56fe\u7075\u5956\u5f97\u4e3b Yann LeCun \u5408\u4f5c\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u9879\u7814\u7a76\u7684\u4e3b\u9898\u662f\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684 Transformer\uff08Transformers without Normalization\uff09\uff0c\u5e76\u5df2\u88ab CVPR 2025 \u4f1a\u8bae\u63a5\u6536\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-bccdd1d2c7318f0147f4d3dfbf67de22.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-bccdd1d2c7318f0147f4d3dfbf67de22.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Meta FAIR \u7814\u7a76\u79d1\u5b66\u5bb6\u5218\u58ee\u7684\u63a8\u6587<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u8fc7\u53bb\u5341\u5e74\uff0c\u5f52\u4e00\u5316\u5c42\u5df2\u7ecf\u5de9\u56fa\u4e86\u5176\u4f5c\u4e3a\u73b0\u4ee3\u795e\u7ecf\u7f51\u7edc\u6700\u57fa\u672c\u7ec4\u4ef6\u4e4b\u4e00\u7684\u5730\u4f4d\u3002\u8fd9\u4e00\u5207\u53ef\u4ee5\u8ffd\u6eaf\u5230 2015 \u5e74\u6279\u5f52\u4e00\u5316\uff08batch normalization\uff09\u7684\u53d1\u660e\uff0c\u5b83\u4f7f\u89c6\u89c9\u8bc6\u522b\u6a21\u578b\u7684\u6536\u655b\u901f\u5ea6\u53d8\u5f97\u66f4\u5feb\u3001\u66f4\u597d\uff0c\u5e76\u5728\u968f\u540e\u51e0\u5e74\u4e2d\u83b7\u5f97\u8fc5\u901f\u53d1\u5c55\u3002\u4ece\u90a3\u65f6\u8d77\uff0c\u7814\u7a76\u4eba\u5458\u9488\u5bf9\u4e0d\u540c\u7684\u7f51\u7edc\u67b6\u6784\u6216\u9886\u57df\u63d0\u51fa\u4e86\u8bb8\u591a\u5f52\u4e00\u5316\u5c42\u7684\u53d8\u4f53\u3002<\/section>\n<section><\/section>\n<section>\u5982\u4eca\uff0c\u51e0\u4e4e\u6240\u6709\u73b0\u4ee3\u7f51\u7edc\u90fd\u5728\u4f7f\u7528\u5f52\u4e00\u5316\u5c42\uff0c\u5176\u4e2d\u5c42\u5f52\u4e00\u5316\uff08Layer Norm\uff0cLN\uff09\u662f\u6700\u53d7\u6b22\u8fce\u4e4b\u4e00\uff0c\u7279\u522b\u662f\u5728\u5360\u4e3b\u5bfc\u5730\u4f4d\u7684 Transformer \u67b6\u6784\u4e2d\u3002<\/section>\n<section><\/section>\n<section>\u5f52\u4e00\u5316\u5c42\u7684\u5e7f\u6cdb\u5e94\u7528\u5f88\u5927\u7a0b\u5ea6\u4e0a\u5f97\u76ca\u4e8e\u5b83\u4eec\u5728\u4f18\u5316\u65b9\u9762\u7684\u5b9e\u8bc1\u4f18\u52bf\u3002\u9664\u4e86\u5b9e\u73b0\u66f4\u597d\u7684\u7ed3\u679c\u4e4b\u5916\uff0c\u5f52\u4e00\u5316\u5c42\u8fd8\u6709\u52a9\u4e8e\u52a0\u901f\u548c\u7a33\u5b9a\u6536\u655b\u3002\u968f\u7740\u795e\u7ecf\u7f51\u7edc\u53d8\u5f97\u8d8a\u6765\u8d8a\u5bbd\u3001\u8d8a\u6765\u8d8a\u6df1\uff0c\u5f52\u4e00\u5316\u5c42\u7684\u5fc5\u8981\u6027\u53d8\u5f97\u8d8a\u6765\u8d8a\u91cd\u8981\u3002\u56e0\u6b64\uff0c\u7814\u7a76\u4eba\u5458\u666e\u904d\u8ba4\u4e3a\u5f52\u4e00\u5316\u5c42\u5bf9\u4e8e\u6709\u6548\u8bad\u7ec3\u6df1\u5ea6\u7f51\u7edc\u81f3\u5173\u91cd\u8981\uff0c\u751a\u81f3\u662f\u5fc5\u4e0d\u53ef\u5c11\u7684\u3002\u8fd9\u4e00\u89c2\u70b9\u4e8b\u5b9e\u4e0a\u5f97\u5230\u4e86\u5fae\u5999\u8bc1\u660e\uff1a<strong>\u8fd1\u5e74\u6765\uff0c\u65b0\u67b6\u6784\u7ecf\u5e38\u5bfb\u6c42\u53d6\u4ee3\u6ce8\u610f\u529b\u5c42\u6216\u5377\u79ef\u5c42\uff0c\u4f46\u51e0\u4e4e\u603b\u662f\u4fdd\u7559\u5f52\u4e00\u5316\u5c42\u3002<\/strong><\/section>\n<section><\/section>\n<section>\u672c\u6587\u4e2d\uff0c\u7814\u7a76\u8005\u63d0\u51fa\u4e86 Transformer \u4e2d\u5f52\u4e00\u5316\u5c42\u7684\u4e00\u79cd\u7b80\u5355\u5e73\u66ff\u3002\u4ed6\u4eec\u7684\u63a2\u7d22\u59cb\u4e8e\u4ee5\u4e0b\u89c2\u5bdf\uff1aLN \u5c42\u4f7f\u7528\u7c7b tanh \u7684 S \u5f62\u66f2\u7ebf\u5c06\u5176\u8f93\u5165\u6620\u5c04\u5230\u8f93\u51fa\uff0c\u540c\u65f6\u7f29\u653e\u8f93\u5165\u6fc0\u6d3b\u5e76\u538b\u7f29\u6781\u503c\u3002<\/section>\n<section><\/section>\n<section>\u53d7\u6b64\u542f\u53d1\uff0c<strong>\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u5143\u7d20\u7ea7\u8fd0\u7b97\uff0c\u79f0\u4e3a Dynamic Tanh\uff08DyT\uff09\uff0c\u5b9a\u4e49\u4e3a\uff1aDyT (x) = tanh (\u03b1x)<\/strong>\uff0c\u5176\u4e2d \u03b1 \u662f\u4e00\u4e2a\u53ef\u5b66\u4e60\u53c2\u6570\u3002\u6b64\u8fd0\u7b97\u65e8\u5728\u901a\u8fc7 \u03b1 \u5b66\u4e60\u9002\u5f53\u7684\u7f29\u653e\u56e0\u5b50\u5e76\u901a\u8fc7\u6709\u754c tanh \u51fd\u6570\u538b\u7f29\u6781\u503c\u6765\u6a21\u62df LN \u7684\u884c\u4e3a\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4e0e\u5f52\u4e00\u5316\u5c42\u4e0d\u540c\uff0cDyT \u53ef\u4ee5\u5b9e\u73b0\u8fd9\u4e24\u79cd\u6548\u679c\uff0c\u800c\u65e0\u9700\u8ba1\u7b97\u6fc0\u6d3b\u6570\u636e\u3002<\/section>\n<section><\/section>\n<section>\u8bba\u6587\u4e00\u4f5c Jiachen Zhu \u4e3a\u7ebd\u7ea6\u5927\u5b66\u56db\u5e74\u7ea7\u535a\u58eb\u751f\u3001\u4e8c\u4f5c\u9648\u946b\u78ca\uff08Xinlei Chen\uff09\u4e3a FAIR \u7814\u7a76\u79d1\u5b66\u5bb6\uff0c\u9879\u76ee\u8d1f\u8d23\u4eba\u4e3a\u5218\u58ee\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2e898cf3ba758901416b0bb82b46fc26.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2e898cf3ba758901416b0bb82b46fc26.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aTransformers without Normalization<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2503.10622<\/section>\n<\/li>\n<li>\n<section>\u9879\u76ee\u4e3b\u9875\uff1ahttps:\/\/jiachenzhu.github.io\/DyT\/<\/section>\n<\/li>\n<li>\n<section>GitHub \u5730\u5740\uff1ahttps:\/\/github.com\/jiachenzhu\/DyT<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>DyT \u4f7f\u7528\u8d77\u6765\u975e\u5e38\u7b80\u5355\uff0c\u5982\u4e0b\u56fe 1 \u6240\u793a\uff0c\u7814\u7a76\u8005\u76f4\u63a5\u7528 DyT \u66ff\u6362\u89c6\u89c9\u548c\u8bed\u8a00 Transformer \u7b49\u67b6\u6784\u4e2d\u7684\u73b0\u6709\u5f52\u4e00\u5316\u5c42\u3002\u5b9e\u8bc1\u7ed3\u679c\u8868\u660e\uff0c\u4f7f\u7528 DyT \u7684\u6a21\u578b\u53ef\u4ee5\u5728\u5404\u79cd\u8bbe\u7f6e\u4e2d\u7a33\u5b9a\u8bad\u7ec3\u5e76\u83b7\u5f97\u8f83\u9ad8\u7684\u6700\u7ec8\u6027\u80fd\u3002\u540c\u65f6\uff0cDyT \u901a\u5e38\u4e0d\u9700\u8981\u5728\u539f\u59cb\u67b6\u6784\u4e0a\u8c03\u6574\u8bad\u7ec3\u8d85\u53c2\u6570\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-725fd702bf6ff61ffb80c7f8b79f7fc0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-725fd702bf6ff61ffb80c7f8b79f7fc0.png\"><\/a><\/section>\n<section>DyT \u6a21\u5757\u53ef\u4ee5\u901a\u8fc7\u77ed\u77ed\u51e0\u884c PyTorch \u4ee3\u7801\u6765\u5b9e\u73b0\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1e6bf71f4b8180ba78d5452a10a85758.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1e6bf71f4b8180ba78d5452a10a85758.png\"><\/a><\/section>\n<section>\u8be5\u5de5\u4f5c\u6311\u6218\u4e86\u300c\u5f52\u4e00\u5316\u5c42\u5bf9\u8bad\u7ec3\u73b0\u4ee3\u795e\u7ecf\u7f51\u7edc\u5fc5\u4e0d\u53ef\u5c11\u300d\u8fd9\u4e00\u89c2\u5ff5\uff0c\u5e76\u63d0\u4f9b\u4e86\u6709\u5173\u5f52\u4e00\u5316\u5c42\u5c5e\u6027\u7684\u5b9e\u8bc1\u89c1\u89e3\u3002\u6b64\u5916\uff0c\u521d\u6b65\u7ed3\u679c\u8868\u660e\uff0cDyT \u53ef\u4ee5\u63d0\u5347\u8bad\u7ec3\u548c\u63a8\u7406\u901f\u5ea6\uff0c\u4ece\u800c\u6210\u4e3a\u4ee5\u6548\u7387\u4e3a\u5bfc\u5411\u7684\u7f51\u7edc\u8bbe\u8ba1\u7684\u5019\u9009\u65b9\u6848\u3002<\/section>\n<section><\/section>\n<section>\u5218\u58ee\u53d1\u63a8\u79f0\uff0c\u5bf9\u4ed6\u800c\u8a00\uff0c\u5f52\u4e00\u5316\u5c42\u4e00\u76f4\u662f\u6df1\u5ea6\u5b66\u4e60\u4e2d\u6bd4\u8f83\u795e\u79d8\u7684\u5185\u5bb9\u3002\u8fd9\u9879\u5de5\u4f5c\u8ba9\u4ed6\u5bf9\u5f52\u4e00\u5316\u5c42\u7684\u4f5c\u7528\u6709\u4e86\u66f4\u6df1\u7684\u7406\u89e3\u3002\u53e6\u5916\uff0c\u8003\u8651\u5230\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u9700\u8981\u6570\u5343\u4e07\u7684\u7b97\u529b\u9700\u6c42\uff0cDyT \u6709\u6f5c\u529b\u52a9\u529b\u6210\u672c\u964d\u4f4e\u3002\u4ed6\u5f88\u671f\u5f85\u63a5\u4e0b\u6765 DyT \u7684\u5e94\u7528\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-333676a17ae56c4e3b20cd5b875ca848.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-333676a17ae56c4e3b20cd5b875ca848.png\"><\/a><\/section>\n<section><strong>\u5f52\u4e00\u5316\u5c42\u6709\u4ec0\u4e48\u4f5c\u7528\uff1f<\/strong><\/section>\n<section><\/section>\n<section>\u8981\u53bb\u6389 Transformer \u4e2d\u7684\u5f52\u4e00\u5316\u5c42\uff0c\u9996\u5148\u8981\u505a\u7684\u5f53\u7136\u662f\u4e86\u89e3\u5f52\u4e00\u5316\u5c42\u6709\u4ec0\u4e48\u7528\u3002<\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u901a\u8fc7\u5b9e\u8bc1\u7814\u7a76\u5bf9\u6b64\u8fdb\u884c\u4e86\u5206\u6790\u3002\u4e3a\u6b64\uff0c\u4ed6\u4eec\u4f7f\u7528\u4e86\u4e09\u4e2a\u4e0d\u540c\u7684\u7ecf\u8fc7\u8bad\u7ec3\u7684 Transformer \u6a21\u578b\uff1a\u4e00\u4e2a Vision Transformer\uff08ViT-B\uff09\u3001\u4e00\u4e2a wav2vec 2.0 Large Transformer \u548c\u4e00\u4e2a Diffusion Transformer\uff08DiT-XL\uff09\u3002<\/section>\n<section><\/section>\n<section>\u4ed6\u4eec\u4f7f\u7528\u8fd9\u4e09\u4e2a\u6a21\u578b\u91c7\u6837\u4e86\u4e00\u5c0f\u6279\u6837\u672c\uff0c\u5e76\u8ba9\u5176\u524d\u5411\u901a\u8fc7\u6574\u4e2a\u7f51\u7edc\u3002\u7136\u540e\uff0c\u4ed6\u4eec\u76d1\u6d4b\u4e86\u5176\u4e2d\u5f52\u4e00\u5316\u5c42\u7684\u8f93\u5165\u548c\u8f93\u51fa\uff0c\u5373\u5f52\u4e00\u5316\u64cd\u4f5c\u524d\u540e\u7684\u5f20\u91cf\u3002<\/section>\n<section><\/section>\n<section>\u7531\u4e8e LN \u4f1a\u4fdd\u7559\u8f93\u5165\u5f20\u91cf\u7684\u7ef4\u5ea6\uff0c\u56e0\u6b64\u53ef\u4ee5\u5728\u8f93\u5165\u548c\u8f93\u51fa\u5f20\u91cf\u5143\u7d20\u4e4b\u95f4\u5efa\u7acb\u4e00\u4e00\u5bf9\u5e94\u5173\u7cfb\uff0c\u4ece\u800c\u53ef\u4ee5\u76f4\u63a5\u53ef\u89c6\u5316\u5b83\u4eec\u7684\u5173\u7cfb\u3002\u8fd9\u4e2a\u6620\u5c04\u5173\u7cfb\u89c1\u56fe 2\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d8847e59ba9b29e53ffbcd27a6368a93.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d8847e59ba9b29e53ffbcd27a6368a93.png\"><\/a><\/section>\n<section>\u5177\u6709\u5c42\u5f52\u4e00\u5316\u7684\u7c7b tanh \u6620\u5c04\u3002\u5bf9\u4e8e\u8fd9\u4e09\u4e2a\u6a21\u578b\uff0c\u8be5\u56e2\u961f\u53d1\u73b0\uff0c\u5b83\u4eec\u7684\u65e9\u671f LN \u5c42\uff08\u56fe 2 \u7b2c 1 \u5217\uff09\u7684\u8f93\u5165 &#8211; \u8f93\u51fa\u5173\u7cfb\u57fa\u672c\u4e0a\u662f\u7ebf\u6027\u7684\u3002\u4f46\u662f\uff0c\u66f4\u6df1\u7684 LN \u5c42\u5374\u6709\u66f4\u6709\u8da3\u7684\u8868\u73b0\u3002<\/section>\n<section><\/section>\n<section>\u53ef\u4ee5\u89c2\u5bdf\u5230\uff0c\u8fd9\u4e9b\u66f2\u7ebf\u7684\u5f62\u72b6\u5927\u591a\u4e0e tanh \u51fd\u6570\u8868\u793a\u7684\u5b8c\u6574\u6216\u90e8\u5206 S \u5f62\u66f2\u7ebf\u975e\u5e38\u76f8\u4f3c\uff08\u89c1\u56fe 3\uff09\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2e9cf0bb8c7ae43827d895d05b64c0c8.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2e9cf0bb8c7ae43827d895d05b64c0c8.png\"><\/a><\/section>\n<section>\u4eba\u4eec\u53ef\u80fd\u9884\u671f LN \u5c42\u4f1a\u5bf9\u8f93\u5165\u5f20\u91cf\u8fdb\u884c\u7ebf\u6027\u53d8\u6362\uff0c\u56e0\u4e3a\u51cf\u53bb\u5e73\u5747\u503c\u548c\u9664\u4ee5\u6807\u51c6\u5dee\u90fd\u662f\u7ebf\u6027\u8fd0\u7b97\u3002LN \u4ee5\u6bcf\u4e2a token \u7684\u65b9\u5f0f\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u4ec5\u5bf9\u6bcf\u4e2a token \u7684\u6fc0\u6d3b\u8fdb\u884c\u7ebf\u6027\u53d8\u6362\u3002<\/section>\n<section><\/section>\n<section>\u7531\u4e8e token \u5177\u6709\u4e0d\u540c\u7684\u5e73\u5747\u503c\u548c\u6807\u51c6\u5dee\uff0c\u56e0\u6b64\u8fd9\u79cd\u7ebf\u6027\u5e76\u4e0d\u5bf9\u8f93\u5165\u5f20\u91cf\u7684\u6240\u6709\u6fc0\u6d3b\u90fd\u6210\u7acb\u3002\u5c3d\u7ba1\u5982\u6b64\uff0c\u8be5\u56e2\u961f\u8868\u793a\u4f9d\u7136\u5f88\u60ca\u8bb6\uff1a\u5b9e\u9645\u7684\u975e\u7ebf\u6027\u53d8\u6362\u7adf\u7136\u4e0e\u67d0\u4e2a\u7ecf\u8fc7\u7f29\u653e\u7684 tanh \u51fd\u6570\u9ad8\u5ea6\u76f8\u4f3c\uff01<\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e\u8fd9\u6837\u4e00\u4e2a S \u578b\u66f2\u7ebf\uff0c\u53ef\u4ee5\u770b\u5230\u5176\u4e2d\u5fc3\u90e8\u5206\uff08x \u503c\u63a5\u8fd1\u96f6\u7684\u90e8\u5206\uff09\u4ecd\u7136\u4e3b\u8981\u5448\u7ebf\u6027\u5f62\u72b6\u3002\u5927\u591a\u6570\u70b9\uff08\u7ea6 99%\uff09\u90fd\u5c5e\u4e8e\u8fd9\u4e2a\u7ebf\u6027\u8303\u56f4\u3002\u4f46\u662f\uff0c\u4ecd\u6709\u8bb8\u591a\u70b9\u660e\u663e\u8d85\u51fa\u6b64\u8303\u56f4\uff0c\u8fd9\u4e9b\u70b9\u88ab\u8ba4\u4e3a\u5177\u6709\u300c\u6781\u7aef\u300d\u503c\uff0c\u4f8b\u5982 ViT \u6a21\u578b\u4e2d x \u5927\u4e8e 50 \u6216\u5c0f\u4e8e -50 \u7684\u70b9\u3002<\/section>\n<section><\/section>\n<section>\u5f52\u4e00\u5316\u5c42\u5bf9\u8fd9\u4e9b\u503c\u7684\u4e3b\u8981\u4f5c\u7528\u662f\u5c06\u5b83\u4eec\u538b\u7f29\u4e3a\u4e0d\u592a\u6781\u7aef\u7684\u503c\uff0c\u4ece\u800c\u4e0e\u5927\u591a\u6570\u70b9\u66f4\u52a0\u4e00\u81f4\u3002\u8fd9\u662f\u5f52\u4e00\u5316\u5c42\u65e0\u6cd5\u901a\u8fc7\u7b80\u5355\u7684\u4eff\u5c04\u53d8\u6362\u5c42\u8fd1\u4f3c\u7684\u5730\u65b9\u3002<\/section>\n<section><\/section>\n<section>\u8be5\u56e2\u961f\u5047\u8bbe\uff0c<strong>\u8fd9\u79cd\u5bf9\u6781\u7aef\u503c\u7684\u975e\u7ebf\u6027\u548c\u4e0d\u6210\u6bd4\u4f8b\u7684\u538b\u7f29\u6548\u5e94\u6b63\u662f\u5f52\u4e00\u5316\u5c42\u7684\u5173\u952e\u4e4b\u5904<\/strong>\u3002<\/section>\n<section><\/section>\n<section>\u524d\u6bb5\u65f6\u95f4\u7684\u4e00\u7bc7\u8bba\u6587\u300aOn the Nonlinearity of Layer Normalization\u300b\u540c\u6837\u91cd\u70b9\u6307\u51fa\u4e86 LN \u5c42\u5f15\u5165\u7684\u5f3a\u975e\u7ebf\u6027\uff0c\u5e76\u4e14\u8868\u660e\u8fd9\u79cd\u975e\u7ebf\u6027\u53ef\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u8868\u5f81\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u8fd9\u79cd\u538b\u7f29\u884c\u4e3a\u8fd8\u53cd\u6620\u4e86\u751f\u7269\u795e\u7ecf\u5143\u5bf9\u5927\u8f93\u5165\u7684\u9971\u548c\uff08saturation\uff09\u7279\u6027\uff0c\u8fd9\u79cd\u73b0\u8c61\u5927\u7ea6\u4e00\u4e2a\u4e16\u7eaa\u524d\u5c31\u5df2\u7ecf\u88ab\u89c2\u5bdf\u5230\u3002<\/section>\n<section><\/section>\n<section>token \u548c\u901a\u9053\u7684\u5f52\u4e00\u5316\u3002LN \u5c42\u5982\u4f55\u5bf9\u6bcf\u4e2a token \u6267\u884c\u7ebf\u6027\u53d8\u6362\uff0c\u540c\u65f6\u4ee5\u8fd9\u79cd\u975e\u7ebf\u6027\u65b9\u5f0f\u538b\u7f29\u6781\u7aef\u503c\u5462\uff1f<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u7406\u89e3\u8fd9\u4e00\u70b9\uff0c\u8be5\u56e2\u961f\u5206\u522b\u6309 token \u548c\u901a\u9053\u5bf9\u8fd9\u4e9b\u70b9\u8fdb\u884c\u53ef\u89c6\u5316\u3002\u56fe 4 \u7ed9\u51fa\u4e86 ViT \u7684\u7b2c\u4e8c\u548c\u7b2c\u4e09\u4e2a\u5b50\u56fe\u7684\u60c5\u51b5\uff0c\u4f46\u4e3a\u4e86\u66f4\u6e05\u6670\uff0c\u56fe\u4e2d\u4f7f\u7528\u4e86\u91c7\u6837\u7684\u70b9\u5b50\u96c6\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-63326f4acf31ac483ef400346ed4a2bd.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-63326f4acf31ac483ef400346ed4a2bd.png\"><\/a><\/section>\n<section>\u5728\u56fe 4 \u5de6\u8fb9\u4e24\u4e2a\u5c0f\u56fe\u4e2d\uff0c\u4f7f\u7528\u4e86\u540c\u4e00\u989c\u8272\u6807\u8bb0\u6bcf\u4e2a token \u7684\u6fc0\u6d3b\u3002\u53ef\u4ee5\u89c2\u5bdf\u5230\uff0c\u4efb\u4f55\u5355\u4e2a token \u7684\u6240\u6709\u70b9\u786e\u5b9e\u90fd\u4f1a\u5f62\u6210\u4e00\u6761\u76f4\u7ebf\u3002\u4f46\u662f\uff0c\u7531\u4e8e\u6bcf\u4e2a token \u90fd\u6709\u4e0d\u540c\u7684\u65b9\u5dee\uff0c\u56e0\u6b64\u659c\u7387\u4e5f\u4e0d\u540c\u3002\u8f93\u5165 x \u8303\u56f4\u8f83\u5c0f\u7684 token \u5f80\u5f80\u5177\u6709\u8f83\u5c0f\u7684\u65b9\u5dee\uff0c\u5e76\u4e14\u5f52\u4e00\u5316\u5c42\u5c06\u4f7f\u7528\u8f83\u5c0f\u7684\u6807\u51c6\u504f\u5dee\u6765\u9664\u5b83\u4eec\u7684\u6fc0\u6d3b\uff0c\u4ece\u800c\u8ba9\u76f4\u7ebf\u6709\u8f83\u5927\u7684\u659c\u7387\u3002<\/section>\n<section><\/section>\n<section>\u603b\u7684\u6765\u8bf4\uff0c\u5b83\u4eec\u5f62\u6210\u4e86\u4e00\u6761\u7c7b\u4f3c\u4e8e tanh \u51fd\u6570\u7684 S \u5f62\u66f2\u7ebf\u3002\u5728\u53f3\u4fa7\u7684\u4e24\u4e2a\u5c0f\u56fe\u4e2d\uff0c\u540c\u6837\u4f7f\u7528\u76f8\u540c\u7684\u989c\u8272\u6807\u8bb0\u5404\u4e2a\u901a\u9053\u7684\u6fc0\u6d3b\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u4e0d\u540c\u901a\u9053\u7684\u8f93\u5165\u8303\u56f4\u5f80\u5f80\u5b58\u5728\u5de8\u5927\u5dee\u5f02\uff0c\u53ea\u6709\u5c11\u6570\u901a\u9053\uff08\u4f8b\u5982\u7ea2\u8272\u3001\u7eff\u8272\u548c\u7c89\u8272\uff09\u4f1a\u8868\u73b0\u51fa\u8f83\u5927\u7684\u6781\u7aef\u503c \u2014\u2014 \u800c\u8fd9\u4e9b\u901a\u9053\u6b63\u662f\u88ab\u5f52\u4e00\u5316\u5c42\u538b\u7f29\u5f97\u6700\u5389\u5bb3\u7684\u3002<\/section>\n<section><\/section>\n<section><strong>Dynamic Tanh\uff08DyT\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u65e2\u77e5\u6839\u672c\uff0c\u6b63\u5f53\u521b\u65b0\u3002\u57fa\u4e8e\u5f52\u4e00\u5316\u5c42\u548c\u6269\u5c55\u7248 tanh \u51fd\u6570\u7684\u76f8\u4f3c\u6027\uff0c\u8be5\u56e2\u961f\u63d0\u51fa\u4e86 Dynamic Tanh\uff08DyT\uff09\uff0c\u5e76\u4e14\u8fd9\u53ef\u4ee5\u4f5c\u4e3a\u5f52\u4e00\u5316\u5c42\u7684\u76f4\u63a5\u66ff\u4ee3\u3002<\/section>\n<section><\/section>\n<section>\u7ed9\u5b9a\u4e00\u4e2a\u8f93\u5165\u5f20\u91cf x\uff0cDyT \u5c42\u7684\u5b9a\u4e49\u5982\u4e0b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0e07769c8e1389b41740d31257412bfc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0e07769c8e1389b41740d31257412bfc.png\"><\/a><\/section>\n<section>\u5176\u4e2d \u03b1 \u662f\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u6807\u91cf\u53c2\u6570\uff0c\u5141\u8bb8\u6839\u636e\u8f93\u5165\u7684\u8303\u56f4\u4ee5\u4e0d\u540c\u7684\u65b9\u5f0f\u7f29\u653e\u8f93\u5165\uff0c\u5e76\u4f1a\u8003\u8651\u4e0d\u540c\u7684 x \u5c3a\u5ea6\uff08\u56fe 2\uff09\u3002\u4e5f\u56e0\u6b64\uff0c\u4ed6\u4eec\u5c06\u6574\u4e2a\u64cd\u4f5c\u547d\u540d\u4e3a\u300c\u52a8\u6001\u300dtanh\u3002\u03b3 \u548c \u03b2 \u662f\u53ef\u5b66\u4e60\u7684\u6bcf\u901a\u9053\u5411\u91cf\u53c2\u6570\uff0c\u4e0e\u6240\u6709\u5f52\u4e00\u5316\u5c42\u4e2d\u4f7f\u7528\u7684\u53c2\u6570\u76f8\u540c \u2014\u2014 \u5b83\u4eec\u5141\u8bb8\u8f93\u51fa\u7f29\u653e\u5230\u4efb\u4f55\u5c3a\u5ea6\u3002\u8fd9\u6709\u65f6\u88ab\u89c6\u4e3a\u5355\u72ec\u7684\u4eff\u5c04\u5c42\uff1b\u8fd9\u91cc\uff0c\u8be5\u56e2\u961f\u5c06\u5b83\u4eec\u89c6\u4e3a DyT \u5c42\u7684\u4e00\u90e8\u5206\uff0c\u5c31\u50cf\u5f52\u4e00\u5316\u5c42\u4e5f\u5305\u62ec\u5b83\u4eec\u4e00\u6837\u3002\u7b97\u6cd5 1 \u7ed9\u51fa\u4e86\u7528\u7c7b PyTorch \u7684\u4f2a\u4ee3\u7801\u5b9e\u73b0\u7684 DyT\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-59a1e74bfe76138887b4e564b9ad2255.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-59a1e74bfe76138887b4e564b9ad2255.png\"><\/a><\/section>\n<section>\u8981\u60f3\u5c06 DyT \u5c42\u96c6\u6210\u5230\u73b0\u6709\u67b6\u6784\u4e2d\uff0c\u65b9\u6cd5\u5f88\u7b80\u5355\uff1a\u76f4\u63a5\u7528\u4e00\u4e2a DyT \u5c42\u66ff\u6362\u4e00\u4e2a\u5f52\u4e00\u5316\u5c42\uff08\u89c1\u56fe 1\uff09\u3002\u8fd9\u9002\u7528\u4e8e\u6ce8\u610f\u529b\u5757\u3001FFN \u5757\u548c\u6700\u7ec8\u5f52\u4e00\u5316\u5c42\u5185\u7684\u5f52\u4e00\u5316\u5c42\u3002<\/section>\n<section><\/section>\n<section>\u5c3d\u7ba1 DyT \u53ef\u80fd\u770b\u8d77\u6765\u50cf\u6216\u53ef\u88ab\u89c6\u4e3a\u6fc0\u6d3b\u51fd\u6570\uff0c\u4f46\u672c\u7814\u7a76\u4ec5\u4f7f\u7528\u5b83\u6765\u66ff\u6362\u5f52\u4e00\u5316\u5c42\uff0c\u800c\u4e0d\u4f1a\u6539\u53d8\u539f\u59cb\u67b6\u6784\u4e2d\u6fc0\u6d3b\u51fd\u6570\u7684\u4efb\u4f55\u90e8\u5206\uff0c\u4f8b\u5982 GELU \u6216 ReLU\u3002\u7f51\u7edc\u7684\u5176\u4ed6\u90e8\u5206\u4e5f\u4fdd\u6301\u4e0d\u53d8\u3002\u8be5\u56e2\u961f\u8fd8\u89c2\u5bdf\u5230\uff0c\u51e0\u4e4e\u4e0d\u9700\u8981\u8c03\u6574\u539f\u59cb\u67b6\u6784\u4f7f\u7528\u7684\u8d85\u53c2\u6570\u5373\u53ef\u4f7f DyT \u8868\u73b0\u826f\u597d\u3002<\/section>\n<section><\/section>\n<section>\u5c3d\u7ba1 DyT \u53ef\u80fd\u770b\u8d77\u6765\u50cf\u6216\u53ef\u88ab\u89c6\u4e3a\u6fc0\u6d3b\u51fd\u6570\uff0c\u4f46\u672c\u7814\u7a76\u4ec5\u4f7f\u7528\u5b83\u6765\u66ff\u6362\u5f52\u4e00\u5316\u5c42\uff0c\u800c\u4e0d\u4f1a\u6539\u53d8\u539f\u59cb\u67b6\u6784\u4e2d\u6fc0\u6d3b\u51fd\u6570\u7684\u4efb\u4f55\u90e8\u5206\uff0c\u4f8b\u5982 GELU \u6216 ReLU\u3002\u7f51\u7edc\u7684\u5176\u4ed6\u90e8\u5206\u4e5f\u4fdd\u6301\u4e0d\u53d8\u3002\u8be5\u56e2\u961f\u8fd8\u89c2\u5bdf\u5230\uff0c\u51e0\u4e4e\u4e0d\u9700\u8981\u8c03\u6574\u539f\u59cb\u67b6\u6784\u4f7f\u7528\u7684\u8d85\u53c2\u6570\u5373\u53ef\u4f7f DyT \u8868\u73b0\u826f\u597d\u3002<\/section>\n<section><\/section>\n<section>\u5173\u4e8e\u7f29\u653e\u53c2\u6570\u3002\u5728\u8fd9\u91cc\uff0c\u603b\u662f\u7b80\u5355\u5730\u5c06 \u03b3 \u521d\u59cb\u5316\u4e3a\u5168\u4e00\u5411\u91cf\uff0c\u5c06 \u03b2 \u521d\u59cb\u5316\u4e3a\u5168\u96f6\u5411\u91cf\uff0c\u540e\u63a5\u5f52\u4e00\u5316\u5c42\u3002\u5bf9\u4e8e scaler \u53c2\u6570 \u03b1\uff0c\u9664\u4e86 LLM \u8bad\u7ec3\u5916\uff0c\u9ed8\u8ba4\u521d\u59cb\u5316\u4e3a 0.5 \u901a\u5e38\u5c31\u8db3\u591f\u4e86\u3002\u9664\u975e\u53e6\u6709\u660e\u786e\u8bf4\u660e\uff0c\u5426\u5219\u5728\u540e\u7eed\u7684\u5b9e\u9a8c\u4e2d\uff0c\u03b1 \u5747\u88ab\u521d\u59cb\u5316\u4e3a 0.5\u3002<\/section>\n<section><\/section>\n<section>\u8bf4\u660e\u3002DyT \u5e76\u975e\u4e00\u79cd\u65b0\u578b\u7684\u5f52\u4e00\u5316\u5c42\uff0c\u56e0\u4e3a\u5b83\u5728\u524d\u5411\u4f20\u9012\u8fc7\u7a0b\u4e2d\u4f1a\u72ec\u7acb\u5730\u5bf9\u5f20\u91cf\u4e2d\u7684\u6bcf\u4e2a\u8f93\u5165\u5143\u7d20\u8fdb\u884c\u64cd\u4f5c\uff0c\u800c\u65e0\u9700\u8ba1\u7b97\u7edf\u8ba1\u6570\u636e\u6216\u5176\u4ed6\u7c7b\u578b\u7684\u805a\u5408\u3002\u4f46\u5b83\u786e\u5b9e\u4fdd\u7559\u4e86\u5f52\u4e00\u5316\u5c42\u7684\u6548\u679c\uff0c\u5373\u4ee5\u975e\u7ebf\u6027\u65b9\u5f0f\u538b\u7f29\u6781\u7aef\u503c\uff0c\u540c\u65f6\u5bf9\u8f93\u5165\u7684\u4e2d\u5fc3\u90e8\u5206\u6267\u884c\u8fd1\u4e4e\u7ebf\u6027\u7684\u53d8\u6362\u3002<\/section>\n<section><\/section>\n<section><strong>DyT \u5728\u5b9e\u9a8c\u4e2d\u7684\u8868\u73b0<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u9a8c\u8bc1 DyT \u7684\u6548\u679c\uff0c\u7814\u7a76\u56e2\u961f\u5728\u4e0d\u540c\u4efb\u52a1\u548c\u9886\u57df\u4e2d\u6d4b\u8bd5\u4e86 Transformer \u53ca\u5176\u4ed6\u67b6\u6784\uff0c\u5c06\u539f\u59cb\u67b6\u6784\u4e2d\u7684 LN \u6216 RMSNorm \u66ff\u6362\u4e3a DyT \u5c42\uff0c\u5e76\u6309\u7167\u5b98\u65b9\u5f00\u6e90\u65b9\u6848\u8fdb\u884c\u8bad\u7ec3\u548c\u6d4b\u8bd5\u3002<\/section>\n<section><\/section>\n<section><strong>\u89c6\u89c9\u76d1\u7763\u5b66\u4e60<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u56e2\u961f\u5728 ImageNet-1K \u5206\u7c7b\u4efb\u52a1\u4e0a\u8bad\u7ec3\u4e86 Base \u548c Large \u4e24\u79cd\u89c4\u6a21\u7684 Vision Transformer\uff08ViT\uff09\u548c ConvNeXt \u6a21\u578b\u3002<\/section>\n<section><\/section>\n<section>\u9009\u62e9 ViT \u548c ConvNeXt \u662f\u56e0\u4e3a\u5b83\u4eec\u65e2\u5177\u4ee3\u8868\u6027\uff0c\u53c8\u5206\u522b\u91c7\u7528\u4e0d\u540c\u673a\u5236\uff1aViT \u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236\uff0cConvNeXt \u57fa\u4e8e\u5377\u79ef\u64cd\u4f5c\u3002\u4ece\u8868 1 \u7684 Top-1 \u5206\u7c7b\u51c6\u786e\u7387\u6765\u770b\uff0cDyT \u5728\u4e24\u79cd\u67b6\u6784\u548c\u4e0d\u540c\u89c4\u6a21\u6a21\u578b\u4e0a\u5747\u4f18\u4e8e LN\u3002\u56fe 5 \u4e2d\u5c55\u793a\u7684 ViT-B \u548c ConvNeXt-B \u7684\u8bad\u7ec3\u635f\u5931\u66f2\u7ebf\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e9e7ce5cfcb8d33c4f00edafb04cced1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e9e7ce5cfcb8d33c4f00edafb04cced1.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u8868 1\uff1aImageNet-1K \u4e0a\u7684\u76d1\u7763\u5206\u7c7b\u51c6\u786e\u7387\u3002DyT \u5728\u4e24\u79cd\u67b6\u6784\u548c\u4e0d\u540c\u6a21\u578b\u89c4\u6a21\u4e0a\u5747\u5b9e\u73b0\u4e86\u4f18\u4e8e\u6216\u7b49\u540c\u4e8e LN \u7684\u6027\u80fd\u8868\u73b0\u3002<\/sup><\/em><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-7b6b5bd460e7b671ee05beb206c374ae.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-7b6b5bd460e7b671ee05beb206c374ae.png\"><\/a><\/section>\n<section><strong>\u89c6\u89c9\u81ea\u76d1\u7763\u5b66\u4e60<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u56e2\u961f\u6d4b\u8bd5\u4e86\u4e24\u79cd\u6d41\u884c\u7684\u89c6\u89c9\u81ea\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\uff1a\u4f55\u607a\u660e\u7684 MAE \u548c DINO\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u4e24\u79cd\u65b9\u6cd5\u90fd\u9ed8\u8ba4\u4f7f\u7528 Vision Transformer \u4f5c\u4e3a\u9aa8\u5e72\u7f51\u7edc\uff0c\u4f46\u8bad\u7ec3\u76ee\u6807\u4e0d\u540c\u3002MAE \u4f7f\u7528\u91cd\u5efa\u635f\u5931\u8fdb\u884c\u8bad\u7ec3\uff0c\u800c DINO \u5219\u4f7f\u7528\u8054\u5408\u5d4c\u5165\u635f\u5931\u3002\u7814\u7a76\u56e2\u961f\u5148\u5728 ImageNet-1K \u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u65e0\u6807\u7b7e\u9884\u8bad\u7ec3\uff0c\u7136\u540e\u6dfb\u52a0\u5206\u7c7b\u5c42\u5e76\u7528\u6807\u7b7e\u6570\u636e\u5fae\u8c03\u6765\u6d4b\u8bd5\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u8868 2 \u5c55\u793a\u4e86\u5fae\u8c03\u7684\u7ed3\u679c\u3002\u5728\u81ea\u76d1\u7763\u5b66\u4e60\u4efb\u52a1\u4e2d\uff0cDyT \u548c LN \u7684\u8868\u73b0\u57fa\u672c\u6301\u5e73\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0579eb892173e0650bb8544ba44b3455.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0579eb892173e0650bb8544ba44b3455.png\"><\/a><\/section>\n<section><strong>\u6269\u6563\u6a21\u578b<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u5728 ImageNet-1K \u4e0a\u8bad\u7ec3\u4e86\u4e09\u4e2a\u5c3a\u5bf8\u5206\u522b\u4e3a B\u3001L \u548c XL \u7684 DiT \u6a21\u578b\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u5728 DiT \u4e2d\uff0cLN \u5c42\u7684\u4eff\u5c04\u53c2\u6570\u7528\u4e8e\u7c7b\u8c03\u8282\uff0cDyT \u5b9e\u9a8c\u4e2d\u4e5f\u4fdd\u7559\u4e86\u8fd9\u4e00\u53c2\u6570\uff0c\u53ea\u662f\u7528 tanh (\u03b1x) \u51fd\u6570\u66ff\u6362\u4e86\u5f52\u4e00\u5316\u8fc1\u79fb\u3002\u8bad\u7ec3\u7ed3\u675f\uff0c\u5982\u8868 3 \u6240\u793a\uff0c\u4e0e LN \u76f8\u6bd4\uff0cDyT \u7684 FID \u503c\u76f8\u5f53\u6216\u6709\u6240\u63d0\u9ad8\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-33e8ae9dddcd720c8e71fcba46f19728.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-33e8ae9dddcd720c8e71fcba46f19728.png\"><\/a><\/section>\n<section><strong>LLM<\/strong><\/section>\n<section><\/section>\n<section>\u8fd9\u4e9b\u6a21\u578b\u662f\u6309\u7167 LLaMA \u4e2d\u6982\u8ff0\u7684\u539f\u59cb\u914d\u65b9\u5728\u5e26\u6709 200B tokens \u7684 The Pile \u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\u7684\u3002\u5728\u5e26\u6709 DyT \u7684 LLaMA \u4e2d\uff0c\u7814\u7a76\u8005\u5728\u521d\u59cb\u5d4c\u5165\u5c42\u4e4b\u540e\u6dfb\u52a0\u4e86\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u6807\u91cf\u53c2\u6570\uff0c\u5e76\u8c03\u6574\u4e86 \u03b1 \u7684\u521d\u59cb\u503c\uff08\u7b2c 7 \u8282\uff09\u3002\u4e0b\u8868 4 \u62a5\u544a\u4e86\u8bad\u7ec3\u540e\u7684\u635f\u5931\u503c\uff0c\u5e76\u6309\u7167 OpenLLaMA \u7684\u65b9\u6cd5\uff0c\u5728 lm-eval \u7684 15 \u4e2a\u96f6\u6837\u672c\u4efb\u52a1\u4e0a\u5bf9\u6a21\u578b\u8fdb\u884c\u4e86\u57fa\u51c6\u6d4b\u8bd5\u3002\u5982\u8868 4 \u6240\u793a\uff0c\u5728\u6240\u6709\u56db\u79cd\u89c4\u6a21\u7684\u6a21\u578b\u4e2d\uff0cDyT \u7684\u8868\u73b0\u4e0e RMSNorm \u76f8\u5f53\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-8781a5a7d3ecf87c8e9569f9499faf6c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-8781a5a7d3ecf87c8e9569f9499faf6c.png\"><\/a><\/section>\n<section>\u56fe 6 \u5c55\u793a\u4e86\u635f\u5931\u66f2\u7ebf\uff0c\u663e\u793a\u4e86\u6240\u6709\u6a21\u578b\u5927\u5c0f\u7684\u76f8\u4f3c\u8d8b\u52bf\uff0c\u8bad\u7ec3\u635f\u5931\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u90fd\u975e\u5e38\u63a5\u8fd1\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-959fb93dc424756c840911e6f35e2216.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-959fb93dc424756c840911e6f35e2216.png\"><\/a><\/section>\n<section>\u8bed\u97f3\u81ea\u76d1\u7763\u5b66\u4e60\u3002\u7814\u7a76\u8005\u5728 LibriSpeech \u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u4e86\u4e24\u4e2a wav2vec 2.0 Transformer \u6a21\u578b\u3002\u8868 5 \u62a5\u544a\u4e86\u6700\u7ec8\u7684\u9a8c\u8bc1\u635f\u5931\u3002\u5728\u4e24\u79cd\u6a21\u578b\u89c4\u6a21\u4e0b\uff0cDyT \u7684\u8868\u73b0\u90fd\u4e0e LN \u76f8\u5f53\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e2ea45d7dd22b572e9c190723ed2d08d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e2ea45d7dd22b572e9c190723ed2d08d.png\"><\/a><\/section>\n<section><\/section>\n<section><strong>DNA \u5e8f\u5217\u5efa\u6a21<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u957f\u7a0b DNA \u5e8f\u5217\u5efa\u6a21\u4efb\u52a1\u4e2d\uff0c\u7814\u7a76\u8005\u5bf9 HyenaDNA \u6a21\u578b\u548c Caduceus \u6a21\u578b\u8fdb\u884c\u4e86\u9884\u8bad\u7ec3\u3002\u7ed3\u679c\u5982\u8868 6\uff0c\u5728\u8fd9\u9879\u4efb\u52a1\u4e2d\uff0cDyT \u4fdd\u6301\u4e86\u4e0e LN \u76f8\u5f53\u7684\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-15337adef15221f6866687ddc10ec289.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-15337adef15221f6866687ddc10ec289.png\"><\/a><\/section>\n<section><strong>\u03b1 \u521d\u59cb\u5316<\/strong><\/section>\n<section><\/section>\n<section><strong>\u975e LLM \u6a21\u578b\u7684 \u03b1 \u521d\u59cb\u5316<\/strong><\/section>\n<section><\/section>\n<section>\u975e LLM \u6a21\u578b\u5bf9 \u03b1_0 \u76f8\u5bf9\u4e0d\u654f\u611f\u3002\u56fe 9 \u5c55\u793a\u4e86\u5728\u4e0d\u540c\u4efb\u52a1\u4e2d\u6539\u53d8 \u03b1_0 \u5bf9\u9a8c\u8bc1\u6027\u80fd\u7684\u5f71\u54cd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-13c26a5f19f1c6529a66a28afabbebf8.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-13c26a5f19f1c6529a66a28afabbebf8.png\"><\/a><\/section>\n<section>\u03b1_0 \u8d8a\u5c0f\uff0c\u8bad\u7ec3\u8d8a\u7a33\u5b9a\u3002\u56fe 10 \u5c55\u793a\u4e86\u4f7f\u7528 ImageNet-1K \u6570\u636e\u96c6\u5bf9\u6709\u76d1\u7763 ViT \u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u6d88\u51cf\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-32acc9f3ed7e5039f1159878a0b06596.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-32acc9f3ed7e5039f1159878a0b06596.png\"><\/a><\/section>\n<section>\u5c06 \u03b1_0 = 0.5 \u8bbe\u4e3a\u9ed8\u8ba4\u503c\u3002\u6839\u636e\u7814\u7a76\u7ed3\u679c\uff0c\u7814\u7a76\u8005\u5c06 \u03b1_0 = 0.5 \u8bbe\u7f6e\u4e3a\u6240\u6709\u975e LLM \u6a21\u578b\u7684\u9ed8\u8ba4\u503c\u3002\u8fd9\u79cd\u8bbe\u7f6e\u65e2\u80fd\u63d0\u4f9b\u4e0e LN \u76f8\u5f53\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\uff0c\u53c8\u80fd\u4fdd\u6301\u5f3a\u5927\u7684\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section><strong>LLM \u6a21\u578b\u7684 \u03b1 \u521d\u59cb\u5316<\/strong><\/section>\n<section><\/section>\n<section>\u8c03\u6574 \u03b1_0 \u53ef\u4ee5\u63d0\u9ad8 LLM \u6027\u80fd\u3002\u5982\u524d\u6240\u8ff0\uff0c\u9ed8\u8ba4\u8bbe\u7f6e \u03b1_0 = 0.5 \u5728\u5927\u591a\u6570\u4efb\u52a1\u4e2d\u8868\u73b0\u826f\u597d\u3002\u7136\u800c\uff0c\u7814\u7a76\u8005\u53d1\u73b0\u8c03\u6574 \u03b1_0 \u53ef\u4ee5\u5927\u5e45\u63d0\u9ad8 LLM \u6027\u80fd\u3002\u4ed6\u4eec\u5bf9\u6bcf\u4e2a LLaMA \u6a21\u578b\u90fd\u8fdb\u884c\u4e86 30B tokens \u7684\u9884\u8bad\u7ec3\uff0c\u5e76\u6bd4\u8f83\u4e86\u5b83\u4eec\u7684\u8bad\u7ec3\u635f\u5931\uff0c\u4ece\u800c\u8c03\u6574\u4e86\u5b83\u4eec\u7684 \u03b1_0\u3002<\/section>\n<section><\/section>\n<section>\u8868 11 \u603b\u7ed3\u4e86\u6bcf\u4e2a\u6a21\u578b\u7684\u8c03\u6574\u540e \u03b1_0 \u503c\uff0c\u5176\u4e2d\u6709\u4e24\u4e2a\u91cd\u8981\u53d1\u73b0\uff1a<\/section>\n<section><\/section>\n<section>1. \u8f83\u5927\u7684\u6a21\u578b\u9700\u8981\u8f83\u5c0f\u7684 \u03b1_0 \u503c\u3002\u4e00\u65e6\u786e\u5b9a\u4e86\u8f83\u5c0f\u6a21\u578b\u7684\u6700\u4f73 \u03b1_0 \u503c\uff0c\u5c31\u53ef\u4ee5\u76f8\u5e94\u5730\u7f29\u5c0f\u8f83\u5927\u6a21\u578b\u7684\u641c\u7d22\u7a7a\u95f4\uff1b<\/section>\n<section>2. \u6ce8\u610f\u529b\u5757\u7684 \u03b1_0 \u503c\u8d8a\u9ad8\uff0c\u6027\u80fd\u8d8a\u597d\u3002\u5bf9\u6ce8\u610f\u529b\u5757\u4e2d\u7684 DyT \u5c42\u521d\u59cb\u5316\u8f83\u9ad8\u7684 \u03b1 \u503c\uff0c\u800c\u5bf9\u5176\u4ed6\u4f4d\u7f6e\uff08\u5373 FFN \u533a\u5757\u5185\u6216\u6700\u7ec8\u7ebf\u6027\u6295\u5f71\u4e4b\u524d\uff09\u7684 DyT \u5c42\u521d\u59cb\u5316\u8f83\u4f4e\u7684 \u03b1 \u503c\uff0c\u53ef\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-c71bbc54f1191acc7636e3bd5a1e1e67.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-c71bbc54f1191acc7636e3bd5a1e1e67.png\"><\/a><\/section>\n<section>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u8bf4\u660e \u03b1_0 \u8c03\u6574\u7684\u5f71\u54cd\uff0c\u56fe 11 \u5c55\u793a\u4e86\u4e24\u4e2a LLaMA \u6a21\u578b\u635f\u5931\u503c\u7684\u70ed\u56fe\u3002\u8fd9\u4e24\u4e2a\u6a21\u578b\u90fd\u53d7\u76ca\u4e8e\u6ce8\u610f\u529b\u5757\u4e2d\u8f83\u9ad8\u7684 \u03b1_0\uff0c\u4ece\u800c\u51cf\u5c11\u4e86\u8bad\u7ec3\u635f\u5931\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-4a5bf05917d914025bbba4c324bc585d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-4a5bf05917d914025bbba4c324bc585d.png\"><\/a><\/section>\n<section>\u6a21\u578b\u5bbd\u5ea6\u4e3b\u8981\u51b3\u5b9a\u4e86 \u03b1_0 \u7684\u9009\u62e9\u3002\u6211\u4eec\u8fd8\u7814\u7a76\u4e86\u6a21\u578b\u5bbd\u5ea6\u548c\u6df1\u5ea6\u5bf9\u6700\u4f18 \u03b1_0 \u7684\u5f71\u54cd\u3002\u7814\u7a76\u8005\u53d1\u73b0\uff0c\u6a21\u578b\u5bbd\u5ea6\u5bf9\u786e\u5b9a\u6700\u4f18 \u03b1_0 \u81f3\u5173\u91cd\u8981\uff0c\u800c\u6a21\u578b\u6df1\u5ea6\u7684\u5f71\u54cd\u5219\u5fae\u4e4e\u5176\u5fae\u3002\u8868 12 \u663e\u793a\u4e86\u4e0d\u540c\u5bbd\u5ea6\u548c\u6df1\u5ea6\u4e0b\u7684\u6700\u4f73 \u03b1_0 \u503c\uff0c\u8868\u660e\u8f83\u5bbd\u7684\u7f51\u7edc\u53ef\u4ee5\u4ece\u8f83\u5c0f\u7684 \u03b1_0 \u503c\u4e2d\u83b7\u76ca\uff0c\u4ece\u800c\u83b7\u5f97\u6700\u4f73\u6027\u80fd\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u6a21\u578b\u6df1\u5ea6\u5bf9 \u03b1_0 \u7684\u9009\u62e9\u5f71\u54cd\u5fae\u4e4e\u5176\u5fae\u3002<\/section>\n<section><\/section>\n<section>\u4ece\u8868 12 \u4e2d\u53ef\u4ee5\u770b\u51fa\uff0c\u7f51\u7edc\u8d8a\u5bbd\uff0c\u300c\u6ce8\u610f\u529b\u300d\u548c\u300c\u5176\u4ed6\u300d\u6240\u9700\u7684\u521d\u59cb\u5316\u5c31\u8d8a\u4e0d\u5747\u8861\u3002\u7814\u7a76\u8005\u5047\u8bbe\uff0cLLM \u7684 \u03b1 \u521d\u59cb\u5316\u7684\u654f\u611f\u5ea6\u4e0e\u5176\u4ed6\u6a21\u578b\u76f8\u6bd4\u8fc7\u5927\u7684\u5bbd\u5ea6\u6709\u5173\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-8eaf6e7b4168fd413a2d9ce2ef9728da.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-8eaf6e7b4168fd413a2d9ce2ef9728da.png\"><\/a><\/section>\n<section>\u66f4\u591a\u7814\u7a76\u7ec6\u8282\uff0c\u53ef\u53c2\u8003\u539f\u8bba\u6587\u3002<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-14-11\" target=\"_blank\">\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Tran [&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":[68,81,65],"class_list":["post-39414","post","type-post","status-publish","format-standard","hentry","category-news","tag-68","tag-81","tag-65"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86 - \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=39414\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=39414\",\"url\":\"https:\/\/17aitech.com\/?p=39414\",\"name\":\"\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=39414#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=39414#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-bccdd1d2c7318f0147f4d3dfbf67de22.png\",\"datePublished\":\"2025-03-14T15:02:28+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=39414#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=39414\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=39414#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-bccdd1d2c7318f0147f4d3dfbf67de22.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-bccdd1d2c7318f0147f4d3dfbf67de22.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=39414#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86\"}]},{\"@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":"\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86 - \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=39414","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=39414","url":"https:\/\/17aitech.com\/?p=39414","name":"\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=39414#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=39414#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-bccdd1d2c7318f0147f4d3dfbf67de22.png","datePublished":"2025-03-14T15:02:28+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=39414#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=39414"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=39414#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-bccdd1d2c7318f0147f4d3dfbf67de22.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-bccdd1d2c7318f0147f4d3dfbf67de22.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=39414#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u6ca1\u6709\u5f52\u4e00\u5316\u5c42\u7684Transformer\uff01\u5218\u58ee\u5e26\u961f\uff0c\u4f55\u607a\u660e\u3001Yann LeCun\u90fd\u53c2\u4e0e\u4e86"}]},{"@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\/39414","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=39414"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39414\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=39414"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=39414"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=39414"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}