{"id":35,"date":"2025-08-13T17:52:33","date_gmt":"2025-08-13T09:52:33","guid":{"rendered":"https:\/\/alyname.top\/index.php\/2025\/08\/13\/attention-is-all-you-need\/"},"modified":"2025-08-15T16:44:24","modified_gmt":"2025-08-15T08:44:24","slug":"attention-is-all-you-need","status":"publish","type":"post","link":"https:\/\/alyname.top\/index.php\/2025\/08\/13\/attention-is-all-you-need\/","title":{"rendered":"\u7b14\u8bb0 – Attention Is All You Need"},"content":{"rendered":"

\u67d0\u7bc7\u5f00\u521b\u4e86\u4e00\u4e2a\u65f6\u4ee3\uff08\u7b97\u662f\u5427\uff09\u7684\u6587\u7ae0\u3002\u77ed\u77ed\u5355\u680f\u5341\u9875\u3002<\/p>\n

Attention \u673a\u5236\u3001Transformer\u3002<\/p>\n

Transformer \u524d NLP \u9886\u57df\u5360\u7edf\u6cbb\u5730\u4f4d\u7684\u662f\u7f51\u7edc\u7ed3\u6784\u662f\u5404\u79cd RNN \u548c LSTM \u7b49\u6280\u672f\u3002Transformer \u7684\u8fc7\u7a0b\u76f8\u5bf9\u6765\u8bf4\u7b80\u6d01\u5f97\u591a\uff1a\u53ea\u5305\u542b\u4e00\u5bf9 Encoder-Decoder\uff0c\u5176\u4e2d\u7684\u6838\u5fc3\u7ec4\u6210\u90e8\u5206\u662f\u4e00\u5806\u5806\u53e0\u7684 Attention \u5355\u5143\u3002<\/p>\n

\u4e3a\u5565\u53eb\u201cAll you need\u201d\uff1f\u56e0\u4e3a\u8fd9\u7bc7\u6587\u7ae0\u63d0\u51fa\u7684\u7f51\u7edc\u7ed3\u6784\u662f\u7b2c\u4e00\u4e2a\u4ec5\u7528 Attention \u673a\u5236\uff0c\u4e0d\u7528\u4efb\u4f55 RNN \u6216 CNN \u76f8\u5173\u7ed3\u6784\u642d\u5efa\u7684\u7f51\u7edc\u3002<\/p>\n

[1706.03762] Attention Is All You Need<\/a>\uff082017 \u5e74 6 \u6708\uff09<\/p>\n

\n
\u8981\u5e72\u5565<\/h5>\n

Transformer \u80fd\u5e72\u7684\u4e8b\u60c5\u633a\u591a\u7684\uff0c\u51e0\u4e4e\u4e00\u5207\u5e8f\u5217\u76f8\u5173\u7684\u4efb\u52a1\u90fd\u80fd handle\u3002<\/p>\n

\u8bba\u6587\u4e2d\u91cd\u70b9\u63d0\u5230\u7684\u662f\u673a\u5668\u7ffb\u8bd1\u4efb\u52a1\u3002\u8fd9\u91cc\u4e5f\u4ee5\u7ffb\u8bd1\u4e3a\u4f8b\u3002<\/p>\n

Transformer<\/h5>\n

\"Pasted
\n\u597d\u7ecf\u5178\u7684\u56fe\u554a\u3002<\/p>\n

Encoder \u7684\u4f5c\u7528\u662f\u628a\u8f93\u5165\u7684\u8bcd\u5411\u91cf <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>x<\/mi>1<\/mn>\u22ef<\/mo>n<\/mi><\/mrow><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u53d8\u6210\u9ad8\u7ef4\u9690\u7a7a\u95f4\u4e2d\u7684 <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>z<\/mi>1<\/mn>\u22ef<\/mo>n<\/mi><\/mrow><\/msup><\/math><\/mjx-assistive-mml><\/mjx-container>\u3002Decoder \u6839\u636e\u8fd9\u4e9b <\/g><\/g><\/g><\/svg>z<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u81ea\u56de\u5f52\u5730\u751f\u6210\u8f93\u51fa <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>y<\/mi>1<\/mn>\u22ef<\/mo>m<\/mi><\/mrow><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container>\u3002<\/p>\n

\n

\u4ec0\u4e48\u662f\u81ea\u56de\u5f52\uff1fDecoder \u4e00\u4e2a\u4e00\u4e2a\u8f93\u51fa <\/g><\/g><\/g><\/svg>y<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u5143\u7d20\uff0c\u6bcf\u6b21\u6839\u636e <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>y<\/mi>1<\/mn>\u22ef<\/mo>i<\/mi>\u2212<\/mo>1<\/mn><\/mrow><\/mrow><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u548c <\/g><\/g><\/g><\/svg>z<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u751f\u6210\u5f53\u524d\u7684 <\/g><\/g><\/g><\/g><\/g><\/svg>y<\/mi>i<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container>\u3002<\/p>\n<\/blockquote>\n

\u6bcf\u4e2a\u7f16\u7801\u5668\u5206\u4e3a <\/g><\/g><\/g><\/svg>N<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u5c42\u3002\u5bf9 Encoder\uff0c\u4e00\u4e2a\u5c42\u7531\u4e00\u4e2a Attention \u5c42\u548c\u4e00\u4e2a\u5168\u8fde\u63a5\u6784\u6210\uff1b\u5bf9 Decoder\uff0c\u4e00\u4e2a\u5c42\u7531\u4e00\u4e2a Attention\uff0c\u4e00\u4e2a\u63a5\u53d7 Encoder \u8f93\u51fa\u7684 Attention \u548c\u4e00\u4e2a\u5168\u8fde\u63a5\u6784\u6210\u3002\u5404\u5c42\u4e4b\u95f4\u90fd\u6709\u6b8b\u5dee\u8fde\u63a5\u7b49\u7ed3\u6784\u3002<\/p>\n

\u6587\u4e2d\u7684 Attention \u5c42\u90fd\u662f\u591a\u5934\u7684\uff0c\u4e00\u4e2a Multi-Head Attention \u5c42\u7531 <\/g><\/g><\/g><\/svg>h<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u5934\u6784\u6210\uff0c\u6bcf\u4e2a\u5934\u662f\u4e00\u4e2a Attention \u5355\u5143\u3002\u4e00\u4e2a Attention \u5355\u5143\u63a5\u53d7\u4e24\u4e2a <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>n<\/mi>\u00d7<\/mo>d<\/mi>k<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u77e9\u9635 <\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u4e00\u4e2a <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>n<\/mi>\u00d7<\/mo>d<\/mi>v<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u77e9\u9635 <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u8f93\u51fa\u4e00\u4e2a <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>n<\/mi>\u00d7<\/mo>d<\/mi>v<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u77e9\u9635\u3002<\/p>\n

Attention \u5c42\u63a5\u53d7\u4e09\u4e2a\u957f\u4e3a <\/g><\/g><\/g><\/svg>n<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684 <\/g><\/g><\/g><\/svg>d<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7ef4\u5411\u91cf\u5e8f\u5217\uff08\u5176\u5b9e\u957f\u5ea6\u4e0d\u540c\u4e5f\u884c\u5427\uff09<\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>0<\/mn><\/msub>,<\/mo>K<\/mi>0<\/mn><\/msub>,<\/mo>V<\/mi>0<\/mn><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u5bf9\u6bcf\u4e2a\u5934\u628a\u5b83\u4eec\u5206\u522b\u4e58\u4ee5\u77e9\u9635 <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>W<\/mi>i<\/mi>Q<\/mi><\/msubsup>,<\/mo>W<\/mi>i<\/mi>K<\/mi><\/msubsup>,<\/mo>W<\/mi>i<\/mi>V<\/mi><\/msubsup><\/math><\/mjx-assistive-mml><\/mjx-container> \u53d8\u6210\u8f93\u5165\u5230\u8be5\u5934\u7684 <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi>,<\/mo>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\uff08\u7ef4\u5ea6\u5206\u522b\u4e3a <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>k<\/mi><\/msub>,<\/mo>d<\/mi>k<\/mi><\/msub>,<\/mo>d<\/mi>v<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container>\uff09\u3002\u6240\u6709\u5934\u5904\u7406\u5b8c\u4e4b\u540e\u628a\u6240\u6709 <\/g><\/g><\/g><\/svg>h<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a <\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>v<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7ef4\u8f93\u51fa\u62fc\u8d77\u6765\uff0c\u4e58\u4ee5 <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>h<\/mi>d<\/mi>v<\/mi><\/msub>\u00d7<\/mo>d<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7ef4\u77e9\u9635 <\/g><\/g><\/g><\/g><\/g><\/svg>W<\/mi>O<\/mi><\/msup><\/math><\/mjx-assistive-mml><\/mjx-container> \u5f97\u5230 <\/g><\/g><\/g><\/svg>d<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7ef4\u8f93\u51fa\u3002<\/p>\n

\u89c2\u5bdf\u4e0a\u9762\u7684\u56fe\u53ef\u4ee5\u53d1\u73b0\uff0cEncoder \u7528\u7684 Attention \u5c42\u548c Decoder \u7528\u7684\u7b2c\u4e00\u4e2a Attention \u5c42\u5206\u522b\u7528\u4e86\u4e09\u4e2a\u4e00\u6837\u7684\u8f93\u5165\uff0c\u79f0\u8fd9\u79cd Attention \u5c42\u662f Self-Attention\u3002Decoder \u7528\u7684\u7b2c\u4e8c\u4e2a Attention \u5c42\u91c7\u7528\u4e86\u6765\u81ea Encoder \u7684\u8f93\u51fa\u4f5c\u4e3a <\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u6765\u6e90\uff0c\u79f0\u8fd9\u662f Cross-Attention\u3002<\/p>\n

\u5728\u4e0b\u6587\u4e2d\u53ef\u4ee5\u770b\u5230\uff0cAttention \u5355\u5143\u5185\u90e8\u7684\u64cd\u4f5c\u6d41\u7a0b\u662f\u56fa\u5b9a\u7684\uff0c\u6ca1\u6709\u53ef\u8bad\u7ec3\u7684\u53c2\u6570\u3002Attention \u5c42\u4e2d\u5b9e\u9645\u53ef\u8bad\u7ec3\u7684\u53c2\u6570\u662f\u4e0a\u8ff0\u7684\u51e0\u4e2a <\/g><\/g><\/g><\/svg>W<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u5b83\u4eec\u63a7\u5236\u4e86\u4ece\u539f\u6765\u7684\u5e8f\u5217\u91cc\u9762\u201c\u63d0\u53d6\u7279\u5f81\u201d <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi>,<\/mo>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u8fc7\u7a0b\u3002<\/p>\n

Attention \u673a\u5236<\/h5>\n

\u8003\u8651\u8fd9\u6837\u4e00\u4e2a\u95ee\u9898\uff1a\u6709\u4e00\u7ec4 <\/g><\/g><\/g><\/svg>n<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u952e\u503c\u5bf9 <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>{<\/mo>(<\/mo>k<\/mi>i<\/mi><\/msub>,<\/mo>v<\/mi>i<\/mi><\/msub>)<\/mo>}<\/mo><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u952e\u548c\u503c\u90fd\u662f\u5411\u91cf\uff0c\u73b0\u5728\u7ed9\u4e86\u4e00\u4e2a\u8be2\u95ee\u5411\u91cf <\/g><\/g><\/g><\/svg>q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u5e0c\u671b\u627e\u5230\u6240\u6709\u4e0e\u8be2\u95ee\u5411\u91cf\u76f8\u4f3c<\/strong>\u7684\u952e\u5bf9\u5e94\u7684\u503c\uff0c\u6216\u8005\u628a\u6bcf\u4e2a\u503c\u6309\u7167\u952e\u4e0e\u8be2\u95ee\u7684\u76f8\u4f3c\u5ea6\u7ed9\u4e2a\u52a0\u6743\uff1b<\/g><\/g><\/g><\/g><\/g><\/svg>n<\/mi>=<\/mo>1<\/mn><\/math><\/mjx-assistive-mml><\/mjx-container> \u65f6\u5e0c\u671b\u628a <\/g><\/g><\/g><\/svg>v<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u6309 <\/g><\/g><\/g><\/g><\/g><\/svg>q<\/mi>,<\/mo>k<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u76f8\u4f3c\u5ea6\u8fdb\u884c\u67d0\u79cd\u52a0\u6743\u6c42\u51fa\u5bf9\u5e94\u503c\u3002<\/p>\n

\u9996\u5148\u5bb9\u6613\u60f3\u5230\u8861\u91cf\u4e24\u4e2a\u5411\u91cf <\/g><\/g><\/g><\/g><\/g><\/svg>q<\/mi>,<\/mo>k<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u76f8\u4f3c\u5ea6\u7684\u65b9\u6cd5\u53ef\u4ee5\u662f\u67d0\u79cd\u5e26\u5f52\u4e00\u5316\u7684\u70b9\u79ef\uff1a\u5982\u679c <\/g><\/g><\/g><\/g><\/g><\/svg>q<\/mi>\u22c5<\/mo>k<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u5f88\u5c0f\u7684\u8bdd\u610f\u5473\u7740\u5b83\u4eec\u51e0\u4e4e\u6b63\u4ea4\u4ece\u800c\u80af\u5b9a\u4e0d\u76f8\u4f3c\uff1b\u628a <\/g><\/g><\/g><\/svg>q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u548c <\/g><\/g><\/g><\/svg>k<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u5206\u522b\u5f52\u4e00\u5316\u5230\u5355\u4f4d\u5411\u91cf\u540e\u5b83\u4eec\u70b9\u79ef\u8d8a\u5927\u5728\u5355\u4f4d\u7403\u9762\u4e0a\u7684\u4f4d\u7f6e\u8d8a\u63a5\u8fd1\u3002<\/p>\n

\u56e0\u4e3a\u8fd9\u4e9b\u4e1c\u897f\u90fd\u4f1a\u5728\u795e\u7ecf\u7f51\u7edc\u7684\u8f93\u51fa\u4e0a\u8fdb\u884c\uff0c\u5982\u679c\u7f51\u7edc\u6ca1\u70b8\u7684\u8bdd\u53ef\u4ee5\u8ba4\u4e3a <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>q<\/mi>,<\/mo>k<\/mi>,<\/mo>v<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u6bcf\u4e2a\u5143\u7d20\u90fd\u662f\u72ec\u7acb\u540c\u5206\u5e03\u7684\u4e14\u5747\u503c\u662f <\/g><\/g><\/g><\/svg>0<\/mn><\/math><\/mjx-assistive-mml><\/mjx-container> \u65b9\u5dee\u662f <\/g><\/g><\/g><\/svg>1<\/mn><\/math><\/mjx-assistive-mml><\/mjx-container>\uff08\u59d1\u4e14\u8fd9\u4e48\u8ba4\u4e3a\u5427\uff09\uff0c\u6240\u4ee5\u5148\u4e0d\u8003\u8651\u5411\u91cf\u7ea7\u522b\u7684\u5f52\u4e00\u5316\u3002\u4e3a\u4e86\u7a81\u51fa\u90a3\u51e0\u4e2a\u6700\u76f8\u4f3c\u7684 <\/g><\/g><\/g><\/g><\/g><\/svg>k<\/mi>i<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u91c7\u7528 Softmax\uff1a<\/p>\n

<\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>Softmax<\/mtext>(<\/mo>z<\/mi>)<\/mo>i<\/mi><\/msub>=<\/mo>e<\/mi>z<\/mi>i<\/mi><\/msub><\/mrow><\/msup>\u2211<\/mo>j<\/mi><\/munder>e<\/mi>z<\/mi>j<\/mi><\/msub><\/mrow><\/msup><\/mrow><\/mfrac><\/math><\/mjx-assistive-mml><\/mjx-container><\/p>\n

\u5047\u8bbe <\/g><\/g><\/g><\/svg>q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u548c\u6bcf\u4e2a <\/g><\/g><\/g><\/g><\/g><\/svg>k<\/mi>i<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u90fd\u662f <\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>k<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7ef4\u5411\u91cf\uff0c<\/g><\/g><\/g><\/g><\/g><\/svg>v<\/mi>i<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u90fd\u662f <\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>v<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7ef4\u7684\uff0c\u6709 <\/g><\/g><\/g><\/svg>n<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u952e\u503c\u5bf9\u3002\u628a <\/g><\/g><\/g><\/svg>n<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u952e\u5e76\u5728\u4e00\u8d77\u6784\u6210\u4e00\u4e2a <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>n<\/mi>\u00d7<\/mo>d<\/mi>k<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u77e9\u9635 <\/g><\/g><\/g><\/svg>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u8003\u8651 <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>Softmax<\/mtext>(<\/mo>q<\/mi>K<\/mi>T<\/mi><\/msup>)<\/mo><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u8fd9\u4e2a <\/g><\/g><\/g><\/svg>n<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7ef4\u5411\u91cf\u6070\u5982\u5176\u5206\u5730\u8868\u793a\u4e86 <\/g><\/g><\/g><\/svg>q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u548c\u6bcf\u4e2a <\/g><\/g><\/g><\/g><\/g><\/svg>k<\/mi>i<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u76f8\u4f3c\u5ea6\u3002<\/p>\n

\u7136\u540e\u628a\u6240\u6709 <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e5f\u5e76\u8d77\u6765\uff0c<\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>Softmax<\/mtext>(<\/mo>q<\/mi>K<\/mi>T<\/mi><\/msup>)<\/mo>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u5c31\u662f\u6309\u76f8\u4f3c\u5ea6\u52a0\u6743\u4e4b\u540e\u7684\u5168\u4f53\u503c\u5411\u91cf\u4e4b\u548c<\/strong>\uff0c\u8fd9\u662f\u4e00\u4e2a <\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>v<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7ef4\u7684\u5411\u91cf\u3002<\/p>\n

\u73b0\u5728\u5047\u8bbe\u6709 <\/g><\/g><\/g><\/svg>m<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u8be2\u95ee\uff0c\u5b83\u4eec\u5e76\u6210\u4e86\u4e00\u4e2a\u77e9\u9635 <\/g><\/g><\/g><\/svg>Q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\u3002Attention \u5355\u5143\u63a5\u53d7\u4e09\u4e2a\u8f93\u5165\u77e9\u9635 <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi>,<\/mo>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u8f93\u51fa<\/p>\n

<\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>Attn<\/mtext>(<\/mo>Q<\/mi>,<\/mo>K<\/mi>,<\/mo>V<\/mi>)<\/mo>=<\/mo>Softmax<\/mtext>(<\/mo>Q<\/mi>K<\/mi>T<\/mi><\/msup><\/mrow>d<\/mi>k<\/mi><\/msub><\/msqrt><\/mfrac>)<\/mo>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container><\/p>\n

\n

\u4e3a\u5565\u8981\u9664\u4e00\u4e2a <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>k<\/mi><\/msub><\/msqrt><\/math><\/mjx-assistive-mml><\/mjx-container>\uff1f\u8fd9\u662f\u7531\u4e8e <\/g><\/g><\/g><\/g><\/g><\/g><\/svg>q<\/mi>K<\/mi>T<\/mi><\/msup><\/math><\/mjx-assistive-mml><\/mjx-container>\uff08\u7531\u4e8e\u6d89\u53ca\u5230 <\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>k<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u65b9\u5dee\u4e3a <\/g><\/g><\/g><\/svg>1<\/mn><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u5143\u7d20\u76f8\u52a0\uff09\u7684\u65b9\u5dee\u4f1a\u53d8\u4e3a <\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>k<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c<\/g><\/g><\/g><\/g><\/g><\/svg>d<\/mi>k<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u8f83\u5927\u65f6\u8fd9\u4e2a\u70b9\u79ef\u5bb9\u6613\u53d8\u5f97\u5f88\u5927\u800c\u6574\u4f53\u8dd1\u5230 Softmax \u7684\u4e24\u7aef\u53bb\uff0c\u4ece\u800c\u9664\u4e00\u4e2a\u6807\u51c6\u5dee\u8fdb\u884c\u5f52\u4e00\u5316\u3002<\/p>\n

\u597d\u8349\u7387\u7684\u64cd\u4f5c\u554a\u2026\u2026\u4f46\u786e\u5b9e\u662f\u90a3\u4e48\u56de\u4e8b\u2026\u2026<\/s><\/p>\n<\/blockquote>\n

<\/g><\/g><\/g><\/svg>Q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u88ab\u79f0\u4e3a\u201c\u67e5\u8be2\u201d\uff08Query\uff09\uff0c<\/g><\/g><\/g><\/svg>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u88ab\u79f0\u4e3a\u201c\u952e\u201d\uff08Key\uff09\uff0c<\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u88ab\u79f0\u4e3a\u201c\u503c\u201d\uff08Value\uff09\u3002Attention \u5355\u5143\uff08\u5168\u79f0 Scaled Dot-Production Attention\uff09\u5b9e\u73b0\u7684\u4e8b\u60c5\uff1a\u6c42\u51fa\u6309 <\/g><\/g><\/g><\/svg>q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u548c <\/g><\/g><\/g><\/g><\/g><\/svg>k<\/mi>i<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u76f8\u4f3c\u5ea6\u52a0\u6743\u540e\u7684\u5168\u4f53 <\/g><\/g><\/g><\/g><\/g><\/svg>v<\/mi>i<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e4b\u548c\uff0c\u5bf9 <\/g><\/g><\/g><\/svg>m<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u8be2\u95ee\u6279\u91cf\u6c42\u89e3\u8f93\u51fa\u3002<\/p>\n

\n

\u90a3\u4e3a\u5565\u53eb\u201c\u6ce8\u610f\u529b\u201d\u673a\u5236\uff1f<\/g><\/g><\/g><\/svg>Q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u76f8\u5f53\u4e8e\u662f\u9700\u8981\u6ce8\u610f\u7684\u5730\u65b9\uff0c<\/g><\/g><\/g><\/g><\/g><\/svg>K<\/mi>\u2212<\/mo>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u5bf9\u662f\u6240\u6709\u5730\u65b9\u4ee5\u53ca\u5b83\u4eec\u5b9e\u9645\u5b58\u5728\u7684\u4e1c\u897f\u3002<\/p>\n<\/blockquote>\n

Why Attention?<\/h5>\n
\n

\u6587\u4e2d\u5c45\u7136\u4e13\u5f00\u4e00\u8282\u8bb2\u4e86 Why\uff01\u597d\u611f\u52a8\u597d\u611f\u52a8<\/p>\n

\u2026\u2026\u8bf6\u600e\u4e48\u611f\u89c9\u8bb2\u7684\u5168\u662f\u5e9f\u8bdd<\/p>\n<\/blockquote>\n

\u9996\u5148\u8003\u8651\u4e3a\u5565\u8981\u6709 Self-Attention \u548c Cross-Attention\u3002\u4ee5\u7ffb\u8bd1\u4efb\u52a1\u4e3a\u4f8b\uff0cEncoder \u4f1a\u628a\u539f\u6765\u7684\u53e5\u5b50\u7f16\u7801\u6210\u4e00\u4e32\u9ad8\u7ef4\u5411\u91cf <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>z<\/mi>1<\/mn><\/msup>,<\/mo>\u22ef<\/mo>,<\/mo>z<\/mi>n<\/mi><\/msup><\/math><\/mjx-assistive-mml><\/mjx-container>\uff0c\u7528 Self-Attention \u4fdd\u8bc1\u53ef\u4ee5\u5173\u6ce8\u5230\u53e5\u5b50\u4e2d\u6bcf\u5bf9\u4f4d\u7f6e\u7684\u8bcd\u7684\u5173\u7cfb\uff1bDecoder \u9700\u8981\u8f93\u51fa\u7ffb\u8bd1\u540e\u7684\u53e5\u5b50\uff0c\u5176\u4e2d\u6bcf\u4e2a\u8bcd\u7684\u6765\u6e90\u662f\u4e4b\u524d\u7ffb\u8bd1\u51fa\u7684\u6240\u6709\u8bcd\u548c\u539f\u6765\u7684\u53e5\u5b50\uff0c\u56e0\u6b64\u8981\u7528 Cross-Attention \u628a\u8bf8 <\/g><\/g><\/g><\/svg>z<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u62ff\u8fdb\u6765\u3002<\/p>\n

\u518d\u8003\u8651\u4e3a\u5565\u662f Attention \u8fd9\u4e2a\u7ed3\u6784\u3002\u4e00\u4e9b\u539f\u56e0\uff1a<\/p>\n

    \n
  • NLP \u7684\u4efb\u52a1\u5f80\u5f80\u662f\u57fa\u4e8e\u5e8f\u5217\u7684\uff0c\u63d0\u53d6\u5e8f\u5217\u4e2d\u957f\u8ddd\u79bb\u5143\u7d20\u4e4b\u95f4\u7684\u5173\u7cfb\u662f\u7f51\u7edc\u7684\u4e00\u4e2a\u5173\u952e\u80fd\u529b\u3002\u6211\u4eec\u8003\u8651\u7f51\u7edc\u4e2d\u7684\u6570\u636e\u6d41\u5411\uff0c\u5173\u6ce8\u4e24\u4e2a\u957f\u8ddd\u79bb\u5143\u7d20\u7b2c\u4e00\u6b21\u88ab\u5173\u8054\u5230\u4e00\u8d77\u65f6\u7ecf\u8fc7\u7684\u8def\u5f84\u957f\u5ea6\u6709\u591a\u957f\uff0c\u8def\u5f84\u957f\u5ea6\u8d8a\u77ed\uff0c\u7ecf\u8fc7\u7684\u53ef\u8bad\u7ec3\u53c2\u6570\u5c31\u8d8a\u5c11\uff0c\u8fd9\u4e24\u4e2a\u5143\u7684\u5173\u8054\u7d20\u4fbf\u8d8a\u5bb9\u6613\u88ab\u5b66\u4e60\u5230\u3002\u53ef\u4ee5\u53d1\u73b0\uff0cRNN \u8fde\u63a5\u957f\u8ddd\u79bb\u5143\u7d20\u6240\u9700\u8981\u7ecf\u8fc7\u7684\u8def\u5f84\u957f\u5ea6\u662f <\/g><\/g><\/g><\/g><\/g><\/g><\/svg>O<\/mi>(<\/mo>n<\/mi>)<\/mo><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\uff1b\u4f46 Attention \u53ef\u4ee5 <\/g><\/g><\/g><\/g><\/g><\/g><\/svg>O<\/mi>(<\/mo>1<\/mn>)<\/mo><\/math><\/mjx-assistive-mml><\/mjx-container> \u8fde\u63a5\u5e8f\u5217\u4e2d\u4e24\u4e2a\u4efb\u610f\u4f4d\u7f6e\u3002<\/li>\n
  • \u5e76\u884c\u53cb\u5584\u3002<\/li>\n
  • \u7528 Self-Attention \u642d\u8d77\u6765\u7684\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u66f4\u5f3a\u3002\u5728\u7814\u7a76 Transformer \u80fd\u529b\u7684\u8bba\u6587\u4e2d\u7ecf\u5e38\u80fd\u770b\u5230\u8fd9\u79cd\u4f8b\u5b50\u56fe\uff1a
    \n\"Pasted<\/li>\n
  • \u8fd9\u79cd\u56fe\u662f\u600e\u4e48\u6765\u7684\uff1f\u5bf9\u4e00\u4e2a\u957f\u4e3a <\/g><\/g><\/g><\/svg>n<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u53e5\u5b50\uff0c\u8003\u5bdf\u7f51\u7edc\u4e2d\u67d0\u4e2a\u7279\u5b9a Self-Attention \u5934\uff0c\u5b83\u7684 <\/g><\/g><\/g><\/svg>Q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u548c <\/g><\/g><\/g><\/svg>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u90fd\u662f <\/g><\/g><\/g><\/g><\/g><\/g><\/g><\/svg>n<\/mi>\u00d7<\/mo>d<\/mi>k<\/mi><\/msub><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u77e9\u9635\uff0c\u5bf9\u5e94\u4e8e <\/g><\/g><\/g><\/svg>n<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e2a\u8bcd\u7684\u4e24\u79cd\u7f16\u7801\u8868\u793a\u3002\u628a\u6bcf\u4e2a\u8bcd\u7684 <\/g><\/g><\/g><\/svg>q<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u548c\u6bcf\u4e2a\u8bcd\u7684 <\/g><\/g><\/g><\/svg>k<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u70b9\u79ef\uff0c\u5f97\u5230\u4e00\u4e2a <\/g><\/g><\/g><\/g><\/g><\/svg>n<\/mi>\u00d7<\/mo>n<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u7ed3\u679c\u8868\u793a\u6bcf\u4e2a\u8bcd\u5728\u4ee5\u6bcf\u4e2a\u8bcd\u4e3a\u8be2\u95ee\u65f6\u6709\u591a\u91cd\u8981\uff0c\u8fd9\u5c31\u53ef\u4ee5\u753b\u6210\u4e0a\u9762\u7684\u4e8c\u5206\u56fe\u4e86\u3002\u8fd9\u4e00\u822c\u88ab\u79f0\u4e3a Attention \u53ef\u89c6\u5316\uff0c\u53ef\u4ee5\u89c2\u5bdf\u6a21\u578b\u662f\u600e\u4e48\u7406\u89e3\u53e5\u5b50\u7ed3\u6784\u7684\u3002\u5bf9 Cross-Attention \u4e5f\u53ef\u4ee5\u753b\u51fa\u7c7b\u4f3c\u7684\u56fe\u3002<\/li>\n<\/ul>\n
    \n

    \u4e3a\u4ec0\u4e48 Encoder \u5f80 Decoder \u4f20\u7684\u6570\u636e\u88ab\u653e\u5230\u4e86\u4e2d\u95f4\u5c42 <\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u4f4d\u7f6e\uff0c\u800c\u975e <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container>\uff1f<\/p>\n

    \u8fd9\u53ef\u4ee5\u4ece\u4e0a\u9762\u7684 Attention \u53ef\u89c6\u5316\u56fe\u5165\u624b\u3002<\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u5171\u540c\u63a7\u5236\u4e86\u751f\u6210\u7684 <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u5e94\u8be5\u4ece\u4e4b\u524d\u7684\u54ea\u4e9b <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u52a0\u6743\u5f97\u5230\uff0c\u800c <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u624d\u662f\u6700\u7ec8 Decode \u53d8\u6210\u8bcd\u7684\u4e1c\u897f\u3002\u4ee5\u7ffb\u8bd1\u4efb\u52a1\u4e3a\u4f8b\uff0c<\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u4fe1\u606f\u662f\u5173\u4e8e\u53e5\u5b50\u7ed3\u6784\u7684\uff0c\u662f\u8de8\u8bed\u8a00\u5b58\u5728\u7684\uff0c\u56e0\u6b64\u53ef\u4ee5\u7528\u4e8e\u6307\u5bfc\u7ffb\u8bd1\u540e\u7684\u53e5\u5b50\u5e94\u8be5\u957f\u5565\u6837\uff0c\u800c <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7684\u4fe1\u606f\u662f\u5bf9\u6bcf\u4e2a\u8bed\u8a00 specified \u7684\uff0c\u628a\u8f93\u5165\u7684\u4e2d\u6587\u7684 <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e22\u5230\u8f93\u51fa\u7684\u82f1\u6587\u7684 <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u91cc\u9762\u663e\u7136\u4f1a\u5e72\u6270\u751f\u6210\u3002<\/p>\n

    \u4e8b\u5b9e\u4e0a Transformer \u5728\u522b\u7684\u9886\u57df\u7684\u5e94\u7528\u4e5f\u9075\u5faa\u8fd9\u4e2a\u539f\u5219\u3002\u6709\u7bc7\u505a\u56fe\u50cf\u98ce\u683c\u8fc1\u79fb\u7684\u8bba\u6587\uff08TODO\uff09\u4e5f\u662f\u628a\u56fe\u50cf\u7ed3\u6784\u4fe1\u606f\u4f5c\u4e3a <\/g><\/g><\/g><\/g><\/g><\/svg>Q<\/mi>,<\/mo>K<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u7136\u540e\u98ce\u683c\u4fe1\u606f\u4f20 <\/g><\/g><\/g><\/svg>V<\/mi><\/math><\/mjx-assistive-mml><\/mjx-container> \u4e86<\/p>\n

    \u8fd9\u4e48\u4e00\u770b\u597d\u50cf\u98ce\u683c\u8fc1\u79fb\u4efb\u52a1\u548c\u7ffb\u8bd1\u4efb\u52a1\u633a\u50cf\u7684\u2026\u2026\uff1f<\/p>\n<\/blockquote>\n

    \n

    \u8fd8\u6709\u6ca1\u6709\u522b\u7684\u7ed3\u6784\u5177\u6709\u548c Attention \u4e00\u6837\u597d\u751a\u81f3\u66f4\u597d\u7684\u6027\u8d28\uff1f\u2026\u2026\u4e0d\u77e5\u9053\uff0c\u6211\u4e0d\u60f3\u7814\u7a76 NLP \u554a\u2026\u2026<\/p>\n<\/blockquote>\n

    <\/div>","protected":false},"excerpt":{"rendered":"

    \u67d0\u7bc7\u5f00\u521b\u4e86\u4e00\u4e2a\u65f6\u4ee3\uff08\u7b97\u662f\u5427\uff09\u7684\u6587\u7ae0\u3002\u77ed\u77ed\u5355\u680f\u5341\u9875\u3002 Attention \u673a\u5236\u3001Transformer\u3002 Tra […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-35","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/posts\/35","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/comments?post=35"}],"version-history":[{"count":2,"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/posts\/35\/revisions"}],"predecessor-version":[{"id":133,"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/posts\/35\/revisions\/133"}],"wp:attachment":[{"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/media?parent=35"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/categories?post=35"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/alyname.top\/index.php\/wp-json\/wp\/v2\/tags?post=35"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}