{"id":5788,"date":"2025-03-20T11:49:52","date_gmt":"2025-03-20T03:49:52","guid":{"rendered":"https:\/\/nullthought.net\/?p=5788"},"modified":"2025-10-07T15:47:11","modified_gmt":"2025-10-07T07:47:11","slug":"%e6%97%a0%e5%bd%92%e4%b8%80%e5%8c%96transformer%ef%bc%9a%e7%94%a8dynamic-tanh-dyt%e5%8f%96%e4%bb%a3%e5%b1%82%e5%bd%92%e4%b8%80%e5%8c%96%ef%bc%88layer-normalization-ln%ef%bc%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5788","title":{"rendered":"\u65e0\u5f52\u4e00\u5316Transformer\uff1a\u7528Dynamic Tanh (DyT)\u53d6\u4ee3\u5c42\u5f52\u4e00\u5316\uff08Layer Normalization, LN\uff09"},"content":{"rendered":"\n

\u8bba\u6587Transformers without Normalization<\/a><\/strong>\u7684\u7814\u7a76\u8bc1\u660e\u4e86Transformer\u53ef\u4ee5\u5728\u65e0\u5f52\u4e00\u5316\u7684\u60c5\u51b5\u4e0b\u7a33\u5b9a\u8bad\u7ec3\uff0c\u5e76\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u7684\u66ff\u4ee3\u65b9\u6cd5 DyT<\/strong>\u3002DyT \u901a\u8fc7 \u52a8\u6001\u7f29\u653e tanh<\/strong> \u53d6\u4ee3 LN\uff0c\u6210\u529f\u590d\u73b0\u4e86\u5f52\u4e00\u5316\u5c42\u7684\u4f5c\u7528\uff0c\u5e76\u5728\u591a\u4e2a\u5b9e\u9a8c\u4e2d\u8fbe\u5230\u4e86\u7b49\u540c\u6216\u66f4\u4f18\u7684\u6027\u80fd\u3002\u8be5\u7814\u7a76\u6311\u6218\u4e86\u5f52\u4e00\u5316\u5c42\u201c\u4e0d\u53ef\u6216\u7f3a\u201d\u7684\u4f20\u7edf\u8ba4\u77e5\uff0c\u5e76\u4e3a\u9ad8\u6548\u795e\u7ecf\u7f51\u7edc\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u65b0\u7684\u601d\u8def\u3002<\/p>\n\n\n\n

\u8bba\u6587\u4f5c\u8005\u4e3aJiachen Zhu, Xinlei Chen, Kaiming He\uff08\u4f55\u607a\u660e\uff09<\/a>, Yann LeCun, Zhuang Liu\uff0c\u6765\u81ea\u6765\u81eaFAIR(Meta), New York University, MIT\u548cPrinceton University\u3002\u6768\u7acb\u6606\uff08Yann LeGun\uff09\u6559\u6388<\/a>\u4e5f\u662f\u8bba\u6587\u4f5c\u8005\u4e4b\u4e00\u3002<\/p>\n\n\n\n

\"\"
Transformers without Normalization<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n

\u4e00\u3001\u5f15\u8a00<\/strong><\/p>\n\n\n\n

\u5728\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u7f51\u7edc\u4e2d\uff0c\u5f52\u4e00\u5316\uff08Normalization\uff09<\/a><\/strong>\u5c42\u5df2\u6210\u4e3a\u4e0d\u53ef\u6216\u7f3a\u7684\u7ec4\u4ef6\u3002\u81ea2015\u5e74\u6279\u5f52\u4e00\u5316\uff08Batch Normalization, BN\uff09\u63d0\u51fa\u4ee5\u6765\uff0c\u5f52\u4e00\u5316\u5c42\u5728\u4f18\u5316\u3001\u7a33\u5b9a\u8bad\u7ec3\u548c\u52a0\u901f\u6536\u655b\u7b49\u65b9\u9762\u7684\u4f18\u52bf\u5df2\u88ab\u5e7f\u6cdb\u9a8c\u8bc1\u3002\u7279\u522b\u662f\u5728Transformer\u67b6\u6784\u4e2d\uff0c\u5c42\u5f52\u4e00\u5316\uff08Layer Normalization, LN\uff09\u6210\u4e3a\u9ed8\u8ba4\u9009\u62e9\uff0c\u5e76\u88ab\u5e7f\u6cdb\u7528\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\u53ca\u5176\u4ed6\u4efb\u52a1\u3002\u7136\u800c\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u65b9\u6cd5 \u52a8\u6001\u53cc\u66f2\u6b63\u5207\uff08Dynamic Tanh, DyT\uff09<\/strong>\uff0c\u5b83\u53ef\u4ee5\u5b8c\u5168\u66ff\u4ee3\u5f52\u4e00\u5316\u5c42\uff0c\u4f7fTransformer\u7f51\u7edc\u5728\u65e0\u9700\u5f52\u4e00\u5316\u7684\u60c5\u51b5\u4e0b\u4ecd\u80fd\u7a33\u5b9a\u8bad\u7ec3\u5e76\u53d6\u5f97\u7b49\u540c\u751a\u81f3\u66f4\u4f18\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

DyT \u7684\u6838\u5fc3\u601d\u60f3\u6765\u6e90\u4e8e\u5bf9LN\u5c42\u7684\u6df1\u5165\u5206\u6790\u3002\u7814\u7a76\u53d1\u73b0\uff0cLN\u7684\u8f93\u5165-\u8f93\u51fa\u5173\u7cfb\u5448\u73b0S\u5f62\u66f2\u7ebf\uff0c\u4e0e\u53cc\u66f2\u6b63\u5207\uff08tanh\uff09\u51fd\u6570\u6781\u4e3a\u76f8\u4f3c\u3002\u56e0\u6b64\uff0cDyT\u901a\u8fc7 \u53ef\u5b66\u4e60\u7684\u7f29\u653e\u56e0\u5b50 \u03b1<\/strong> \u63a7\u5236 tanh(x) \u7684\u5c3a\u5ea6\u53d8\u5316\uff0c\u4ece\u800c\u5728\u4e0d\u8ba1\u7b97\u8f93\u5165\u7edf\u8ba1\u7279\u6027\u7684\u60c5\u51b5\u4e0b\uff0c\u5b9e\u73b0\u7c7b\u4f3cLN\u7684\u5f52\u4e00\u5316\u4f5c\u7528\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u4f7f\u7528DyT\u7684Transformer\u53ef\u4ee5\u5728\u5404\u79cd\u4efb\u52a1\u4e2d\u8fbe\u5230\u751a\u81f3\u8d85\u8fc7\u4f20\u7edfLN\u7684\u6027\u80fd\uff0c\u5e76\u4e14\u65e0\u9700\u8c03\u6574\u8d85\u53c2\u6570\u3002<\/p>\n\n\n\n

\u672c\u7814\u7a76\u6311\u6218\u4e86\u5f52\u4e00\u5316\u5c42\u5728\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u4e2d\u201c\u4e0d\u53ef\u6216\u7f3a\u201d\u7684\u4f20\u7edf\u8ba4\u77e5\uff0c\u5e76\u4e3a\u6df1\u5165\u7406\u89e3\u5f52\u4e00\u5316\u5c42\u7684\u4f5c\u7528\u63d0\u4f9b\u4e86\u65b0\u89c1\u89e3\u3002\u540c\u65f6\uff0cDyT \u5728\u8ba1\u7b97\u6548\u7387\u4e0a\u4e5f\u5c55\u73b0\u51fa\u4f18\u52bf\uff0c\u6709\u671b\u63a8\u52a8\u66f4\u9ad8\u6548\u7684\u7f51\u7edc\u8bbe\u8ba1\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u80cc\u666f\uff1a\u5f52\u4e00\u5316\u5c42\u7684\u4f5c\u7528<\/strong><\/p>\n\n\n\n

\u5f52\u4e00\u5316\u5c42\u7684\u4e3b\u8981\u4f5c\u7528\u662f\u8c03\u6574\u8f93\u5165\u5206\u5e03\uff0c\u4ee5\u6539\u5584\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u6536\u655b\u901f\u5ea6\u3002LN\u7684\u6807\u51c6\u8ba1\u7b97\u65b9\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

LN\u7684\u8ba1\u7b97\u72ec\u7acb\u4e8e\u6279\u6b21\u5c3a\u5bf8\uff08batch size\uff09\uff0c\u4ec5\u5728 \u6bcf\u4e2atoken<\/strong> \u7ef4\u5ea6\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u56e0\u6b64\u7279\u522b\u9002\u7528\u4e8eTransformer\u67b6\u6784\u3002\u6b64\u5916\uff0c\u53e6\u4e00\u79cd\u5e38\u89c1\u7684\u5f52\u4e00\u5316\u65b9\u6cd5\u662f RMSNorm<\/strong>\uff0c\u5176\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

RMSNorm\u53bb\u9664\u4e86\u5747\u503c\u5f52\u4e00\u5316\u64cd\u4f5c\uff0c\u5728\u4e00\u4e9b\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08\u5982LLaMA, T5, Qwen\u7b49\uff09\u4e2d\u88ab\u91c7\u7528\u3002<\/p>\n\n\n\n

\u672c\u7814\u7a76\u9996\u5148\u5206\u6790\u4e86LN\u7684\u5b9e\u9645\u4f5c\u7528\uff0c\u53d1\u73b0\u5176\u4e3b\u8981\u529f\u80fd\u5305\u62ec\uff1a<\/p>\n\n\n\n

    \n
  1. \u7ebf\u6027\u7f29\u653e\u8f93\u5165\uff0c\u5e76\u5728\u6781\u7aef\u503c\u5904\u65bd\u52a0\u975e\u7ebf\u6027\u7ea6\u675f\u3002<\/li>\n\n\n\n
  2. \u901a\u8fc7\u5f52\u4e00\u5316\u4e0d\u540c\u901a\u9053\u7684\u65b9\u5dee\uff0c\u9650\u5236\u6570\u503c\u7684\u6781\u7aef\u589e\u957f\u3002<\/li>\n\n\n\n
  3. \u5176\u8f93\u51fa\u5f62\u72b6\u4e0etanh\u51fd\u6570\u6781\u4e3a\u76f8\u4f3c\uff0c\u8fd9\u4e3aDyT\u7684\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u7075\u611f\u3002<\/li>\n<\/ol>\n\n\n\n

    \u4e09\u3001\u52a8\u6001\u53cc\u66f2\u6b63\u5207\uff08DyT\uff09\u65b9\u6cd5<\/strong><\/p>\n\n\n\n

    DyT\u7684\u5b9a\u4e49\u5982\u4e0b\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    DyT \u76f4\u63a5\u66ff\u6362 Transformer \u4e2d\u7684 LN \u5c42\uff0c\u5e76\u4fdd\u6301\u6240\u6709\u5176\u4ed6\u7ec4\u4ef6\uff08\u5982\u81ea\u6ce8\u610f\u529b\u3001\u524d\u9988\u7f51\u7edc\u3001\u6fc0\u6d3b\u51fd\u6570\u7b49\uff09\u4e0d\u53d8\u3002\u4e0eLN\u76f8\u6bd4\uff0cDyT \u5177\u6709\u4ee5\u4e0b\u4f18\u70b9\uff1a<\/p>\n\n\n\n

      \n
    1. \u8ba1\u7b97\u66f4\u7b80\u5355<\/strong>\uff1a\u4e0d\u9700\u8981\u8ba1\u7b97\u5747\u503c\u548c\u65b9\u5dee\uff0c\u51cf\u5c11\u4e86\u8ba1\u7b97\u5f00\u9500\u3002<\/li>\n\n\n\n
    2. \u4fdd\u6301\u5f52\u4e00\u5316\u6548\u679c<\/strong>\uff1atanh(x) \u672c\u8eab\u5177\u6709\u8f93\u5165\u503c\u9650\u5236\u529f\u80fd\uff0c\u53ef\u6a21\u62dfLN\u5bf9\u6781\u7aef\u503c\u7684\u201c\u6324\u538b\u201d\u4f5c\u7528\u3002<\/li>\n\n\n\n
    3. \u65e0\u987b\u8c03\u6574\u8d85\u53c2\u6570<\/strong>\uff1a\u5927\u591a\u6570\u60c5\u51b5\u4e0b\uff0cDyT\u53ef\u76f4\u63a5\u7528\u4e8e\u6807\u51c6Transformer\u67b6\u6784\uff0c\u800c\u4e0d\u9700\u8981\u91cd\u65b0\u641c\u7d22\u8d85\u53c2\u6570\u3002<\/li>\n<\/ol>\n\n\n\n

      \u56db\u3001\u5b9e\u9a8c\u9a8c\u8bc1<\/strong><\/p>\n\n\n\n

      \u4e3a\u4e86\u9a8c\u8bc1DyT\u7684\u6709\u6548\u6027\uff0c\u4f5c\u8005\u5728\u591a\u4e2a\u9886\u57df\u7684\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u5305\u62ec\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8bed\u97f3\u5904\u7406\u548c\u57fa\u56e0\u5e8f\u5217\u5efa\u6a21\u7b49\u3002<\/p>\n\n\n\n

        \n
      1. \u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\uff08ImageNet-1K\u5206\u7c7b\uff09<\/strong>\n