{"id":7546,"date":"2023-02-09T22:13:44","date_gmt":"2023-02-09T13:13:44","guid":{"rendered":"https:\/\/agirobots.com\/?p=7546"},"modified":"2025-07-30T17:35:51","modified_gmt":"2025-07-30T08:35:51","slug":"multi-head-attention","status":"publish","type":"post","link":"https:\/\/developers.agirobots.com\/jp\/multi-head-attention\/","title":{"rendered":"\u3010Transformer\u306e\u57fa\u790e\u3011Multi-Head Attention\u306e\u4ed5\u7d44\u307f"},"content":{"rendered":"\n

\u672c\u8a18\u4e8b\u3067\u306f\u3001Transformer\u306e\u57fa\u790e\u3068\u3057\u3066\u3001Multi-Head Attention\u306e\u4ed5\u7d44\u307f\u3092\u5206\u304b\u308a\u3084\u3059\u304f\u89e3\u8aac\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u672c\u8a18\u4e8b\u306e\u69cb\u6210\u306f\u3001\u306f\u3058\u3081\u306bTransformer\u304a\u3088\u3073Transformer Block\u306b\u3064\u3044\u3066\u7d39\u4ecb\u3057\u3001Transformer\u306b\u304a\u3051\u308bMulti-Head Attention\u306e\u91cd\u8981\u6027\u306b\u3064\u3044\u3066\u8aac\u660e\u3057\u307e\u3059\u3002\u305d\u306e\u5f8c\u306b\u3001Multi-Head Attention\u304c\u3069\u306e\u3088\u3046\u306a\u4ed5\u7d44\u307f\u3067\u5b9f\u73fe\u3055\u308c\u3066\u3044\u308b\u306e\u304b\u3092\u8aac\u660e\u3057\u307e\u3059\u3002\u672c\u8a18\u4e8b\u3092\u8aad\u3081\u3070Multi-Head Attention\u306b\u3064\u3044\u3066\u3001\u6982\u8981\u304b\u3089\u8a08\u7b97\u5f0f\u306b\u81f3\u308b\u307e\u3067\u3001\u307b\u307c\u5168\u3066\u3092\u3054\u7406\u89e3\u3044\u305f\u3060\u3051\u308b\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

Transformer<\/h2>\n\n\n\n

Transformer[1]\u3068\u306f2017\u5e74\u306b\u9ad8\u7cbe\u5ea6\u306a\u6a5f\u68b0\u7ffb\u8a33\u30e2\u30c7\u30eb\u3068\u3057\u3066\u767b\u5834\u3057\u305f\u6a5f\u68b0\u5b66\u7fd2\u30e2\u30c7\u30eb\u3067\u3059\u3002Transformer\u304c\u63d0\u6848\u3055\u308c\u305f\u8ad6\u6587\u306e\u30bf\u30a4\u30c8\u30eb\u306f\u300cAttention Is All You Need\u300d\u3068\u7279\u5fb4\u7684\u3067\u3059\u3002<\/p>\n\n\n\n

Transformer\u306f\u3001\u5f93\u6765\u306e\u6a5f\u68b0\u7ffb\u8a33\u30e2\u30c7\u30eb\u3067\u4f7f\u7528\u3055\u308c\u3066\u3044\u305fSeq2Seq\u306e\u8ab2\u984c\u3067\u3042\u308b\u3001\u5927\u57df\u7684\u306a\u7279\u5fb4\u3084\u30cb\u30e5\u30a2\u30f3\u30b9\u3092\u6349\u3048\u306b\u304f\u3044\u70b9\u3068\u3001\u5b66\u7fd2\u306e\u9ad8\u901f\u5316\u304c\u96e3\u3057\u3044\u70b9\u306b\u5bfe\u3057\u3066\u3001\u305d\u306e\u89e3\u6c7a\u7b56\u3068\u3057\u3066\u63d0\u6848\u3055\u308c\u305f\u30e2\u30c7\u30eb\u3067\u3059\u3002\u8a73\u3057\u304f\u306f\u4ee5\u4e0b\u306e\u8a18\u4e8b\u3067\u89e3\u8aac\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\n
\"\"<\/dt>

Transformer\u306b\u3064\u3044\u3066\u672c\u8cea\u3092\u5206\u304b\u308a\u3084\u3059\u304f\u89e3\u8aac\uff01<\/p>

\u7686\u3055\u3093\u3053\u3093\u306b\u3061\u306f\uff01 \u53bb\u5e74\uff082022\u5e7411\u6708\uff09\u306b\u767a\u8868\u3055\u308c\u305fChatGPT\u306e\u8a71\u984c\u304c\u5c3d\u304d\u306a\u3044\u4eca\u65e5\u3053\u306e\u9803\u3001ChatGPT\u3067\u4f7f\u308f\u308c\u3066\u3044\u308b\u91cd\u8981\u306a\u6280\u8853\u306e\uff11\u3064\u3067\u3042\u308bTransformer\u306b\u3064\u3044\u3066\u3001\u8208\u5473\u3092\u6301\u3063\u305f\u65b9\u306f\u6ca2\u5c71 ... <\/p><\/div><\/dd><\/dl><\/div><\/a><\/div>\n<\/div><\/figure>\n\n\n\n

Transformer\u306f\u5f93\u6765\u304b\u3089\u4f7f\u7528\u3055\u308c\u3066\u304d\u305fCNN\u3084RNN\u3068\u306f\u7570\u306a\u308a\u3001Attention\u6a5f\u69cb\u3068\u547c\u3070\u308c\u308b\u4ed5\u7d44\u307f\u306b\u57fa\u3065\u3044\u305f\u30a8\u30f3\u30b3\u30fc\u30c0\u30c7\u30b3\u30fc\u30c0\u30e2\u30c7\u30eb\u3067\u3059\u3002Attention\u6a5f\u69cb\u3068\u306f\u6ce8\u610f\u8868\u73fe\u3092\u5b66\u7fd2\u3059\u308b\u4ed5\u7d44\u307f\u3067\u3059\u3002LSTM\u306a\u3069\u3067\u4f7f\u7528\u3055\u308c\u3066\u3044\u308b\u30b2\u30fc\u30c8\u6a5f\u69cb\u306b\u4f3c\u3066\u304a\u308a\u3001\u5fc5\u8981\u306a\u60c5\u5831\u306f\u5f37\u8abf\u3057\u3001\u4e0d\u5fc5\u8981\u306a\u60c5\u5831\u306f\u76ee\u7acb\u305f\u306a\u304f\u3059\u308b\u4ed5\u7d44\u307f\u3067\u3059\u3002\u3088\u308a\u826f\u3044\u6ce8\u610f\u8868\u73fe\u3092\u5b66\u7fd2\u3067\u304d\u308c\u3070\u81a8\u5927\u306a\u60c5\u5831\u304b\u3089\u6ce8\u610f\u3092\u5411\u3051\u308b\u3079\u304d\u60c5\u5831\u306b\u7126\u70b9\u3092\u5f53\u3066\u3066\u51e6\u7406\u3092\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u306b\u306a\u308a\u307e\u3059\u3002\u6ce8\u610f\u306e\u82f1\u5358\u8a9e\u3067\u3042\u308bAttention\uff08\u30a2\u30c6\u30f3\u30b7\u30e7\u30f3\uff09\u304b\u3089\u304d\u3066\u3044\u307e\u3059\u3002Transformer\u3067\u306fMulti-Head Attention\u3068\u547c\u3070\u308c\u308bAttention\u6a5f\u69cb\u304c\u4f7f\u308f\u308c\u3066\u3044\u307e\u3059\u3002\u4e0b\u56f3\u306bTransformer\u306e\u69cb\u9020\u3092\u793a\u3057\u307e\u3059\u3002<\/p>\n\n\n

\n
\"\"
Transformer<\/figcaption><\/figure><\/div>\n\n\n

\u8ad6\u6587\u3067\u63d0\u6848\u3055\u308c\u305fTransformer\u306f\u3001\u4e0a\u56f3\u306e\u3088\u3046\u306a\u30a8\u30f3\u30b3\u30fc\u30c0\u30c7\u30b3\u30fc\u30c0\u69cb\u9020\u3092\u3057\u3066\u3044\u307e\u3059\u304c\u3001Transformer\u306e\u672c\u8cea\u7684\u306a\u6a5f\u80fd\u5358\u4f4d\u3067\u8003\u3048\u308b\u3068\u3001\u30a8\u30f3\u30b3\u30fc\u30c0\u90e8\u5206\u3060\u3051\u3092\u898b\u308c\u3070\u5341\u5206\u3067\u3059\u3002\u3053\u308c\u304c\u6b21\u306b\u8aac\u660e\u3059\u308bTransformer\u306e\u57fa\u672c\u5358\u4f4d\u3001Transformer Block\u3067\u3059\u3002<\/p>\n\n\n\n

Transformer Block<\/h2>\n\n\n\n

Transformer Block[2]\u306fTransformer\u306e\u57fa\u672c\u5358\u4f4d\u3067\u3001Multi-Head Attention\u3068Position-Wise Feed-Forward Networks\u304b\u3089\u306a\u308a\u307e\u3059\uff08\u4e0b\u56f3\uff09\u3002BERT\u3084GPT\u306a\u3069\u306eTransformer\u30d9\u30fc\u30b9\u306e\u30e2\u30c7\u30eb\u306b\u304a\u3044\u3066\u3001Transformer\u3068\u306fTransformer Block\u306e\u3053\u3068\u3092\u610f\u5473\u3057\u3066\u3044\u308b\u3053\u3068\u304c\u5927\u534a\u3067\u3059\u3002<\/p>\n\n\n

\n
\"\"
Transformer Block<\/figcaption><\/figure><\/div>\n\n\n

Transformer Block\u306f\u3001Query\u3001Key\u3001Value\u306e\uff13\u3064\u306e\u5165\u529b\u3092\u6301\u3061\u3001Multi-Head Attention\u3092\u4f7f\u3063\u3066\u6ce8\u610f\u8868\u73fe\u306b\u57fa\u3065\u304f\u8a08\u7b97\u3092\u3057\u307e\u3059\u3002Position-Wise Feed-Forward Networks\uff08\u56f3\u4e2d\u3001Feed Forward\uff09\u3067\u306f\u3001Multi-Head Attention\u306e\u51fa\u529b\u7d50\u679c\u3092\u30c8\u30fc\u30af\u30f3\u6bce\u306b\u5909\u63db\u3057\u307e\u3059\u3002\u30c8\u30fc\u30af\u30f3\u6bce\u306b\u8a08\u7b97\u3059\u308b\u305f\u3081Position-Wise\u3068\u547c\u3070\u308c\u307e\u3059\u3002\u307e\u305f\u3001\u5404\u3005\u306e\u5c64\u306b\u5e73\u884c\u3057\u305f\u30b7\u30e7\u30fc\u30c8\u30ab\u30c3\u30c8\u30d1\u30b9\u306fSkip-Connection\u3068\u547c\u3070\u308c\u3001\u52fe\u914d\u6d88\u5931\u304c\u8d77\u3053\u308a\u306b\u304f\u304f\u3057\u30e2\u30c7\u30eb\u306e\u6df1\u5c64\u5316\u3092\u53ef\u80fd\u306b\u3057\u307e\u3059\uff08\u4ee5\u4e0b\u3067\u88dc\u8db3\u3057\u3066\u3044\u307e\u3059\uff09\u3002<\/p>\n\n\n\n

Multi-Head Attention\u306f\u3001\u5165\u529b\u306e\u76f4\u524d\u306bLinear\u5c64\u3092\u6301\u3064Scaled Dot-Product Attention\uff08\uff1dSingle-Head Attention[3]\uff09\u3092\u8907\u6570\u4e26\u5217\u306b\u914d\u7f6e\u3057\u305f\u69cb\u9020\u3092\u3057\u3066\u3044\u307e\u3059\u3002Scaled Dot-Product Attention\u306f\u540d\u524d\u306e\u901a\u308a\u3001\u5185\u7a4d\u306b\u57fa\u3065\u304f\u6ce8\u610f\u8a08\u7b97\u3092\u884c\u3044\u307e\u3059\u304c\u3001\u305d\u308c\u5358\u4f53\u3067\u306f\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u305f\u305a\u3001\u5b66\u7fd2\u306b\u3088\u308a\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3067\u304d\u306a\u3044\u305f\u3081\u3001\u76f4\u524d\u306b\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u3064Linear\u5c64\u3092\u8a2d\u3051\u3066\u591a\u7a2e\u591a\u69d8\u306a\u7279\u5fb4\u90e8\u5206\u7a7a\u9593\u306b\u304a\u3051\u308b\u6ce8\u610f\u8868\u73fe\u306e\u5b66\u7fd2\u3092\u53ef\u80fd\u306b\u3057\u3066\u3044\u307e\u3059\u3002\u5b66\u7fd2\u306b\u3088\u308a\u7372\u5f97\u3055\u308c\u305f\u6ce8\u610f\u8868\u73fe\u3092\u53ef\u8996\u5316\u3059\u308b\u3068\u3001\u5404\u30d8\u30c3\u30c9\u304c\u7570\u306a\u308b\u6ce8\u610f\u8868\u73fe\u3092\u7372\u5f97\u3057\u3066\u3044\u308b\u3053\u3068\u304c\u78ba\u8a8d\u3055\u308c\u3066\u3044\u307e\u3059[1,3,4]\u3002\u30d8\u30c3\u30c9\u304c1\u3064\u306e\u5834\u5408\u306f\u3001Single-Head Attention\u306b\u306a\u308a\u307e\u3059\u304c\u3001\u8907\u6570\u4e26\u5217\u306b\u3059\u308b\u3053\u3068\u3067\u30e2\u30c7\u30eb\u306e\u8868\u73fe\u80fd\u529b\u3092\u9ad8\u304f\u3057\u3066\u3044\u308b\u306e\u3067\u3059\u3002<\/p>\n\n\n

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

Position-Wise Feed-Forward Networks\u3068\u306f\u3001Multi-Head Attention\u306e\u5f8c\u306b\u7d9a\u304f\u9806\u4f1d\u64ad\u578b\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u3059\u3002\u5148\u306b\u8aac\u660e\u3057\u305f\u901a\u308aPosition-Wise\u306b\u306f\u4f4d\u7f6e\u5358\u4f4d\u3067\u3068\u3044\u3046\u610f\u5473\u304c\u3042\u308a\u3001\u5404\u30c8\u30fc\u30af\u30f3\u6bce\u306b\u8a08\u7b97\u3092\u3057\u307e\u3059\u3002Multi-Head Attention\u306e\u4e2d\u306bLinear\u5c64\u304c\u3042\u308a\u3001\u3053\u308c\u3082\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u4e00\u7a2e\u3067\u3059\u304c\u3001\u6d3b\u6027\u5316\u95a2\u6570\u304c\u7121\u3044\u305f\u3081\u3001\u975e\u7dda\u5f62\u5909\u63db\u3092\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u305b\u3093\u3002\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306b\u3088\u308b\u5b66\u7fd2\u306b\u304a\u3044\u3066\u3001\u975e\u7dda\u5f62\u306e\u6d3b\u6027\u5316\u95a2\u6570\u3092\u4f7f\u3046\u3053\u3068\u306b\u306f\u5927\u304d\u306a\u610f\u5473\u304c\u3042\u308a\u307e\u3059\u3002\u305d\u306e\u70ba\u3001Multi-Head Attention\u306e\u76f4\u5f8c\u306bPosition-Wise Feed-Forward Networks\u304c\u3042\u308b\u306e\u3067\u3059\u3002<\/p>\n\n\n\n

\u8907\u96d1\u306a\u69cb\u9020\u3092\u3057\u3066\u3044\u307e\u3059\u304c\u3001\u6839\u672c\u306fScaled Dot-Product Attention\u3067\u3042\u308a\u3001\u305d\u308c\u3055\u3048\u7406\u89e3\u3067\u304d\u3066\u3057\u307e\u3048\u3070\u3001\u3042\u3068\u306f\u8868\u73fe\u80fd\u529b\u3092\u4e0a\u3052\u308b\u305f\u3081\u306e\u5de5\u592b\u3067\u3042\u308a\u3001Multi-Head Attention\u306b\u3064\u3044\u3066\u3082\u5bb9\u6613\u306b\u7406\u89e3\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u3067\u3059\u3002\u305d\u3053\u3067\u672c\u8a18\u4e8b\u3067\u306f\u3001Scaled Dot-Product Attention\u306b\u3064\u3044\u3066\u6700\u521d\u306b\u8a73\u7d30\u3092\u8aac\u660e\u3057\u305f\u5f8c\u306b\u3001\u305d\u306e\u8868\u73fe\u80fd\u529b\u3092\u5411\u4e0a\u3055\u305b\u305fSingle-Head Attention\u3001\u3055\u3089\u306b\u8868\u73fe\u80fd\u529b\u3092\u5411\u4e0a\u3055\u305b\u305fMulti-Head Attention\u306b\u3064\u3044\u3066\u9806\u756a\u306b\u89e3\u8aac\u3092\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

<\/i><\/div>
\n

Skip-Connection\u306b\u3064\u3044\u3066\u88dc\u8db3\u3057\u307e\u3059\u3002Skip-Connection\u3068\u306f\u3001ResNet[5]\u3067\u63d0\u6848\u3055\u308c\u305f\u3001\u5c64\u3092\u30b9\u30ad\u30c3\u30d7\u3059\u308b\u60c5\u5831\u4f1d\u9054\u7d4c\u8def\u306e\u3053\u3068\u3092\u3044\u3044\u307e\u3059\u3002Skip-Connection\u3092\u8a2d\u3051\u308b\u3053\u3068\u3067\u3001\u6052\u7b49\u5199\u50cf\u3092\u5b9f\u73fe\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u3001\u307e\u305f\u3001\u5b66\u7fd2\u6642\u306e\u52fe\u914d\u6d88\u5931\u554f\u984c\u306e\u6539\u5584\u306b\u5bc4\u4e0e\u3059\u308b\u305f\u3081\u3001\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u6df1\u5c64\u5316\u3092\u5b9f\u73fe\u3057\u307e\u3057\u305f\u3002\u4e0b\u56f3\u306f\u3001\uff12\u3064\u306e\u5c64\uff08Layer 1\u3068Layer 2\uff09\u3092\u30b9\u30ad\u30c3\u30d7\u3059\u308b\u60c5\u5831\u4f1d\u9054\u7d4c\u8def\u304c\u3042\u308b\u4f8b\u3092\u8868\u3057\u3066\u3044\u307e\u3059\u3002\\(\\boldsymbol{x}\\)\u304c\u5165\u529b\u3055\u308c\u305f\u3068\u304d\u3001\uff12\u3064\u306e\u5c64\u306b\u3088\u308b\u8a08\u7b97\u7d50\u679c\u306f\\(\\mathcal{F}(\\boldsymbol{x})\\)\u3067Skip-Connection\u306b\u3088\u308a\\(\\boldsymbol{x}\\)\u304c\u305d\u306e\u307e\u307e\u4f1d\u9054\u3055\u308c\u308b\u306e\u3067\u3001\u6700\u7d42\u7684\u306a\u51fa\u529b\\(\\mathcal{H}(\\boldsymbol{x})\\)\u306f\\(\\mathcal{F}(\\boldsymbol{x})+\\boldsymbol{x}\\)\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n

\n
\"\"
Skip-Connection\u306e\u6982\u8981<\/figcaption><\/figure><\/div><\/div><\/div>\n\n\n\n

Scaled Dot-Product Attention<\/h2>\n\n\n\n

Scaled Dot-Product Attention\u3068\u306f\u3001\u65e5\u672c\u8a9e\u3067\u30b9\u30b1\u30fc\u30eb\u5316\u5185\u7a4d\u6ce8\u610f\u3068\u8a33\u3059\u3053\u3068\u304c\u53ef\u80fd\u3067\u3042\u308b\u3088\u3046\u306b\u3001\u5185\u7a4d\u3092\u5229\u7528\u3057\u305f\u30d9\u30af\u30c8\u30eb\u9593\u306e\u985e\u4f3c\u6027\u306b\u57fa\u3065\u304f\u5909\u63db\u3092\u884c\u3044\u307e\u3059\u3002\u5165\u529b\u306f\u3001\uff13\u3064\u3067Query\u3001Key\u3001Value\u3068\u547c\u3070\u308c\u307e\u3059\u3002\u3053\u308c\u3089\u306e\u30cd\u30fc\u30df\u30f3\u30b0\u306f\u30c7\u30fc\u30bf\u30d9\u30fc\u30b9\u3067\u4f7f\u308f\u308c\u3066\u3044\u308b\u30ef\u30fc\u30c9\u304b\u3089\u6765\u3066\u3044\u307e\u3059\u3002\u5185\u90e8\u3067\u884c\u308f\u308c\u3066\u3044\u308b\u64cd\u4f5c\u306f\u30c7\u30fc\u30bf\u30d9\u30fc\u30b9\u3068\u306f\u5927\u304d\u304f\u7570\u306a\u308a\u307e\u3059\u304c\u3001Query\u306b\u57fa\u3065\u3044\u3066Key\u306b\u4f55\u3089\u304b\u306e\u5909\u66f4\u3092\u65bd\u3057\u3001Value\u3092\u53d6\u308a\u51fa\u3057\u3066\u304f\u308b\u64cd\u4f5c\u3092\u3059\u308b\u305f\u3081\u3001\u306a\u3093\u3068\u306a\u304f\u4f3c\u3066\u3044\u308b\u611f\u3058\u306f\u3057\u307e\u3059\u306d\u3002<\/p>\n\n\n\n

Scaled Dot-Product Attention\u306e\u69cb\u9020\u3092\u4e0b\u56f3\u306b\u793a\u3057\u307e\u3059\u3002<\/p>\n\n\n

\n
\"\"
Scaled Dot-Product Attention\u306e\u69cb\u9020<\/figcaption><\/figure><\/div>\n\n\n

\uff13\u3064\u306e\u5165\u529b\u3092\u6301\u3061\u307e\u3059\u304c\u3001\u3053\u3053\u3067\u306f\u4ed5\u7d44\u307f\u3092\u7406\u89e3\u3057\u3084\u3059\u304f\u3059\u308b\u305f\u3081\u306bKey\u3068Value\u306f\u540c\u3058\u5165\u529b\u3068\u3057\u3066\u8aac\u660e\u3092\u3057\u307e\u3059\u3002\u4e0a\u56f3\u306f\u8907\u96d1\u306a\u69cb\u9020\u3092\u8868\u3057\u3066\u3044\u308b\u3088\u3046\u306b\u898b\u3048\u307e\u3059\u304c\u3001\u5358\u7d14\u306b\\(\\text{softmax}\\left(\\frac{\\boldsymbol{QK^T}}{\\sqrt{d_k}}\\right)\\)\u3092\u8a08\u7b97\u3057\u3066\u3044\u308b\u306b\u3059\u304e\u305a\u3001\u5b66\u7fd2\u8981\u7d20\u306f\u5168\u304f\u3042\u308a\u307e\u305b\u3093\u3002<\/p>\n\n\n\n

\u5177\u4f53\u4f8b\u3092\u898b\u305f\u65b9\u304c\u7406\u89e3\u304c\u9032\u3080\u3068\u601d\u3046\u306e\u3067\u3001\u5177\u4f53\u4f8b\u3092\u898b\u3066\u307f\u307e\u3057\u3087\u3046\u3002Query\u306b\u306f\u4efb\u610f\u306e\u5358\u8a9e\u3092\u8868\u3059\u30d9\u30af\u30c8\u30eb\u3092\u3001Key\u3068Value\u306b\u306f\u6587\u7ae0\u3092\u5358\u8a9e\u3054\u3068\u306b\u30d9\u30af\u30c8\u30eb\u5316\u3057\u3066\u4e26\u3079\u305f\u884c\u5217\u3092\u5165\u529b\u3059\u308b\u3053\u3068\u3092\u8003\u3048\u307e\u3059\uff08\u3061\u306a\u307f\u306b\u5358\u8a9e\u306e\u30d9\u30af\u30c8\u30eb\u5316\u3067\u306f\u57cb\u3081\u8fbc\u307f\u5c64\u3092\u7528\u3044\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\uff09\u3002\u4f8b\u3048\u3070\u3001\u300cI have a pen.\u300d\u3068\u3044\u3046\u6587\u7ae0\u3092\u5358\u8a9e\u3054\u3068\u306b\u30c8\u30fc\u30af\u30f3\u5316\u3057\u30d9\u30af\u30c8\u30eb\u306b\u5909\u63db\u3057\u305f\u3068\u3057\u307e\u3059\uff08\u4e0b\u56f3\uff09\u3002\u8272\u306e\u9055\u3044\u306f\u30d9\u30af\u30c8\u30eb\u306e\u9055\u3044\u3092\u8868\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n

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

\u3053\u3053\u3067\u3001Query\uff08\\(\\boldsymbol{Q}\\)\uff09\u3068\u3057\u3066\u300chave\u300d\u3092\u8868\u3059\u30d9\u30af\u30c8\u30eb\u3092\u3001Key\uff08\\(\\boldsymbol{K}\\)\uff09\u3068Value\uff08\\(\\boldsymbol{V}\\)\uff09\u306b\u306f\u5404\u30d9\u30af\u30c8\u30eb\u3092\u9806\u756a\u306b\u4e26\u3079\u305f\u884c\u5217\u3092\u4e0e\u3048\u308b\u3053\u3068\u306b\u3057\u307e\u3059\u3002<\/p>\n\n\n

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

\u3053\u306e\u3068\u304d\u3001\u30bd\u30d5\u30c8\u30de\u30c3\u30af\u30b9\u95a2\u6570\u306e\u5165\u529b\u3067\u3042\u308b\u3001\\(\\frac{\\boldsymbol{QK^T}}{\\sqrt{d_k}}\\)\u3067\u306f\u3001Query\u306e\u30d9\u30af\u30c8\u30eb\u3068\u5165\u529b\u6587\u7ae0\u306e\u5168\u30d9\u30af\u30c8\u30eb\u306b\u5bfe\u3057\u3066\u5185\u7a4d\u3092\u8a08\u7b97\u3057\u3001Key\u306e\u30d9\u30af\u30c8\u30eb\u306e\u6b21\u5143\u6570\\(d_k\\)\u306e\u30eb\u30fc\u30c8\u3067\u30b9\u30b1\u30fc\u30eb\u5316\u3057\u305f\u30d9\u30af\u30c8\u30eb\u3092\u8a08\u7b97\u3057\u307e\u3059\u3002\u3059\u306a\u308f\u3061\u3001\u300cI have a pen.\u300d\u3092\u69cb\u6210\u3059\u308b\u30c8\u30fc\u30af\u30f3\u306b\u5bfe\u3057\u3066\u3001\u300chave\u300d\u3092\u8868\u3059\u30d9\u30af\u30c8\u30eb\u306b\u3069\u306e\u7a0b\u5ea6\u985e\u4f3c\u3057\u3066\u3044\u308b\u304b\u3092\u8a08\u7b97\u3059\u308b\u306e\u3067\u3059\u3002\u30b9\u30b1\u30fc\u30eb\u5316\u306b\u3064\u3044\u3066\u306f\u30bd\u30d5\u30c8\u30de\u30c3\u30af\u30b9\u95a2\u6570\u3068\u306e\u517c\u306d\u5408\u3044\u3067\u5fc5\u8981\u306b\u306a\u308a\u307e\u3059\u3002\u5185\u7a4d\u306e\u8a08\u7b97\u7d50\u679c\u306e\u307e\u307e\u3067\u306f\u6271\u3044\u306b\u304f\u3044\u305f\u3081\u3001\u30bd\u30d5\u30c8\u30de\u30c3\u30af\u30b9\u95a2\u6570\u3092\u9069\u7528\u3057\u3001\u985e\u4f3c\u3057\u3066\u3044\u308b\u5358\u8a9e\u307b\u30691\u306b\u8fd1\u304f\u3001\u76f8\u9055\u3057\u3066\u3044\u308b\u5358\u8a9e\u307b\u30690\u306b\u8fd1\u304f\u3001\u307e\u305f\u5168\u4f53\u306e\u548c\u304c\uff11\u306b\u306a\u308b\u3088\u3046\u306a\u5909\u63db\u3092\u65bd\u3057\u307e\u3059\u3002\u3053\u306e\u7d50\u679c\u3092\u7528\u3044\u3066\u3001Value\u3068\u5185\u7a4d\u3092\u8a08\u7b97\u3057\u307e\u3059\u3002Value\u3068\u306e\u5185\u7a4d\u3068\u306f\u6587\u7ae0\u3092\u69cb\u6210\u3059\u308b\u5404\u30d9\u30af\u30c8\u30eb\u306eQuery\u30d9\u30af\u30c8\u30eb\u3068\u306e\u985e\u4f3c\u5ea6\u306b\u57fa\u3065\u304f\u7dda\u5f62\u7d50\u5408\u306b\u306a\u308a\u307e\u3059\u3002\u4e00\u8a00\u3067\u307e\u3068\u3081\u308b\u3068\u3001Scaled Dot-Product Attention\u3067\u306fQuery\u30d9\u30af\u30c8\u30eb\u3068Key\u306e\u5404\u30d9\u30af\u30c8\u30eb\u306e\u985e\u4f3c\u6027\u306b\u57fa\u3065\u3044\u3066Value\u306e\u5404\u30d9\u30af\u30c8\u30eb\u306e\u7dda\u5f62\u7d50\u5408\u3092\u8a08\u7b97\u3057\u307e\u3059\u3002\u3053\u306e\u8a08\u7b97\u306e\u6d41\u308c\u3092\u4e0b\u56f3\u306b\u793a\u3057\u307e\u3059\u3002<\/p>\n\n\n

\n
\"\"
Scaled Dot-Product Attention\u306e\u51e6\u7406<\/figcaption><\/figure><\/div>\n\n\n

Scaled Dot-Product Attention\u304c\u3069\u306e\u3088\u3046\u306a\u8a08\u7b97\u51e6\u7406\u3092\u884c\u3046\u304b\u7406\u89e3\u3067\u304d\u305f\u3068\u3053\u308d\u3067\u3001\u305d\u306e\u50cd\u304d\u306b\u3064\u3044\u3066\u8aac\u660e\u3057\u307e\u3059\u3002\u4e0b\u56f3\u306b\u5148\u307b\u3069\u306e\u4f8b\u306b\u304a\u3051\u308bScaled Dot-Product Attention\u306e\u50cd\u304d\u3092\u793a\u3057\u307e\u3059\u3002Key\u3068Value\u306b\u306f\u300cI have a pen.\u300d\u3092\u8868\u3059\u884c\u5217\u3092\u3001Query\u306b\u306f\u300chave\u300d\u3092\u8868\u3059\u30d9\u30af\u30c8\u30eb\u3092\u4e0e\u3048\u308b\u3068\u3001\u300chave\u300d\u306e\u30d9\u30af\u30c8\u30eb\u3068\u306e\u985e\u4f3c\u6027\u306b\u57fa\u3065\u3044\u3066\u6587\u7ae0\u5168\u4f53\u306e\u5358\u8a9e\u30d9\u30af\u30c8\u30eb\u3092\u7528\u3044\u3066\u7dda\u5f62\u7d50\u5408\u3092\u8a08\u7b97\u3057\u30d9\u30af\u30c8\u30eb\u3092\u51fa\u529b\u3057\u307e\u3059\u3002\u51fa\u529b\u3055\u308c\u308b\u30d9\u30af\u30c8\u30eb\u306e\u30b5\u30a4\u30ba\u306fQuery\u306e\u30d9\u30af\u30c8\u30eb\u3068\u540c\u3058\u3067\u3059\u3002\u305d\u306e\u70ba\u3001\u51fa\u529b\u30d9\u30af\u30c8\u30eb\u3092\u5358\u8a9e\u30d9\u30af\u30c8\u30eb\u3068\u3057\u3066\u8003\u3048\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002\u3053\u308c\u3092\u518d\u5e30\u7684\u306b\u884c\u3048\u3070\u3001\u6587\u7ae0\u3092\u751f\u6210\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u3067\u3042\u308b\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u306d\u3002<\/p>\n\n\n

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

Query\u306b\u7528\u3044\u308b\u30d9\u30af\u30c8\u30eb\u306f\u3001Key\u3084Value\u306b\u542b\u307e\u308c\u308b\u30d9\u30af\u30c8\u30eb\u3067\u3042\u308b\u5fc5\u8981\u306f\u3042\u308a\u307e\u305b\u3093\uff08\u4e0b\u56f3\uff09\u3002<\/p>\n\n\n

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

\u3053\u306e\u3088\u3046\u306b\u3001Key\u3084Value\u306f\u540c\u3058\u3067\u3001Query\u306f\u7570\u306a\u308b\u5165\u529b\u3067\u3042\u308bAttention\u3092SourceTarget Attention\u3068\u547c\u3073\u307e\u3059\u3002\u4e00\u65b9\u3067\u3001Query\u3082Key\u3084Value\u3068\u540c\u3058\u306b\u3059\u308b\u3053\u3068\u3082\u3067\u304d\u3001\u305d\u306e\u3088\u3046\u306aAttention\u3092Self-Attention\u3068\u547c\u3073\u307e\u3059\u3002<\/p>\n\n\n

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

Transformer\u3067\u306fSelf-Attention\u304c\u591a\u304f\u4f7f\u308f\u308c\u3066\u304a\u308a\u3001\u91cd\u8981\u306a\u610f\u5473\u3092\u6301\u3061\u307e\u3059\u3002\u4e0b\u56f3\u306f\u3001Scaled Dot-Product Attention\u3092Self-Attention\u3068\u3057\u3066\u4f7f\u7528\u3057\u305f\u5834\u5408\u306e\u4f8b\u3092\u8868\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n

\n
\"\"
Scaled Dot-Product Attention\u306b\u3088\u308b\u6ce8\u610f\u8868\u73fe<\/figcaption><\/figure><\/div>\n\n\n

\u4ee5\u4e0a\u304c\u3001Scaled Dot-Product Attention\u306e\u8aac\u660e\u306b\u306a\u308a\u307e\u3059\u3002\u3053\u306e\u3088\u3046\u306b\u5165\u529b\u3055\u308c\u305f\u30d9\u30af\u30c8\u30eb\u306e\u985e\u4f3c\u6027\u306b\u57fa\u3065\u3044\u3066\u8a08\u7b97\u3059\u308b\u306e\u3067\u3059\u304c\u3001Scaled Dot-Product Attention\u306b\u306f\u5927\u304d\u306a\u6b20\u70b9\u304c\u3042\u308a\u307e\u3059\u3002\u305d\u308c\u306f\u3001\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u305f\u306a\u3044\u3068\u3044\u3046\u70b9\u3067\u3059\u3002\u3059\u306a\u308f\u3061\u3001\u591a\u7a2e\u591a\u69d8\u306a\u7279\u5fb4\u90e8\u5206\u7a7a\u9593\u306b\u304a\u3051\u308b\u6ce8\u610f\u8868\u73fe\u3092\u5b66\u7fd2\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u305a\u3001\u57cb\u3081\u8fbc\u307f\u5c64\u306b\u3088\u308a\u5909\u63db\u3055\u308c\u305f\u5206\u6563\u8868\u73fe\u306b\u304a\u3051\u308b\u30d9\u30af\u30c8\u30eb\u306e\u985e\u4f3c\u6027\u306b\u57fa\u3065\u304f\u30ef\u30f3\u30d1\u30bf\u30fc\u30f3\u306a\u5909\u63db\u3057\u304b\u3067\u304d\u306a\u3044\u306e\u3067\u3059\u3002\u305d\u3053\u3067\u3001\u591a\u7a2e\u591a\u69d8\u306a\u7279\u5fb4\u90e8\u5206\u7a7a\u9593\u306b\u304a\u3051\u308b\u6ce8\u610f\u8868\u73fe\u3092\u5b66\u7fd2\u53ef\u80fd\u306b\u3059\u308b\u305f\u3081\u306b\u3001Query\u3001Key\u3001Value\u306e\u5165\u529b\u306e\u76f4\u524d\u306b\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u3082\u3064Linear\u5c64\u3092\u7528\u610f\u3057\u3001\u591a\u7a2e\u591a\u69d8\u306a\u7279\u5fb4\u90e8\u5206\u7a7a\u9593\u306e\u6ce8\u610f\u8868\u73fe\u3092\u5b66\u7fd2\u3067\u304d\u308b\u3088\u3046\u306b\u3057\u305f\u306e\u304c\u6b21\u306b\u8aac\u660e\u3059\u308bSingle-Head Attention\u3067\u3059\u3002<\/p>\n\n\n\n

<\/i><\/div>
\n

\u203bScaled Dot-Product Attention\u306e\u56f3\u4e2d\u3001Mask\u90e8\u5206\u306b\u3064\u3044\u3066\u8aac\u660e\u3092\u3057\u307e\u3059\u3002Generator\u306e\u5b66\u7fd2\u3067\u306f\u3001\u6587\u7ae0\u306e\u306f\u3058\u3081\u3060\u3051\u3092\u898b\u305b\u3066\u3001\u305d\u306e\u5f8c\u3092\u4e88\u6e2c\u3059\u308b\u3088\u3046\u306b\u5b66\u7fd2\u3092\u884c\u3044\u307e\u3059\u3002Key\u306b\u306f\u6587\u7ae0\u5168\u4f53\u304c\u683c\u7d0d\u3055\u308c\u3066\u3044\u307e\u3059\u306e\u3067\u3001\u305d\u306e\u307e\u307e\u51e6\u7406\u3057\u3066\u3057\u307e\u3046\u3068\u3001\u4e88\u6e2c\u3057\u305f\u3044\u672a\u6765\u306e\u60c5\u5831\u304c\u5165\u529b\u306b\u542b\u307e\u308c\u308b\u3053\u3068\u306b\u306a\u308a\u307e\u3059\u3002\u305d\u3053\u3067\u3001\u5185\u7a4d\u3092\u3068\u3063\u305f\u5f8c\u306b\u30de\u30b9\u30af\u51e6\u7406\u3092\u3057\u3066\u672a\u6765\u306e\u60c5\u5831\u3092\u898b\u308c\u306a\u3044\u3088\u3046\u306b\u3057\u3066\u3044\u308b\u306e\u3067\u3059\u3002\u30de\u30b9\u30af\u3057\u3066\u3057\u307e\u3048\u3070\u3001\u30de\u30b9\u30af\u3055\u308c\u30d9\u30af\u30c8\u30eb\u306f\u30bd\u30d5\u30c8\u30de\u30c3\u30af\u30b9\u306b\u3088\u308b\u7dda\u5f62\u7d50\u5408\u3067\u4f7f\u308f\u308c\u306a\u304f\u306a\u308a\u307e\u3059\u3002<\/p>\n<\/div><\/div>\n\n\n\n

Single-Head Attention<\/h2>\n\n\n\n

Single-Head Attention\u3068\u306f\u3001\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u305f\u306a\u3044Scaled Dot-Product Attention\u306e\u8868\u73fe\u80fd\u529b\u3092\u5e83\u3052\u308b\u305f\u3081\u306b\u3001\u5404\u5165\u529b\u306e\u76f4\u524d\u306b\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u3064Linear\u5c64\u3092\u8ffd\u52a0\u3057\u305f\u3082\u306e\u3067\u3059\u3002\u3053\u308c\u306b\u3088\u308a\u3001\u5165\u529b\u3055\u308c\u308b\u30d9\u30af\u30c8\u30eb\u306e\u7279\u5fb4\u7a7a\u9593\u306b\u4f9d\u5b58\u3057\u306a\u3044\u6ce8\u610f\u8868\u73fe\u3092\u5b66\u7fd2\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n

\n
\"\"
Single-Head Attention<\/figcaption><\/figure><\/div>\n\n\n

\u3059\u306a\u308f\u3061\u3001\u5148\u307b\u3069\u306eScaled Dot-Product Attention\u3067\u306f\u3001\u57cb\u3081\u8fbc\u307f\u5c64\u306b\u3088\u308a\u5909\u63db\u3055\u308c\u305f\u5206\u6563\u8868\u73fe\u306b\u304a\u3051\u308b\u6ce8\u610f\u8868\u73fe\u3057\u304b\u5b66\u7fd2\u3067\u304d\u307e\u305b\u3093\u3067\u3057\u305f\u304c\u3001Linear\u5c64\u304c\u8ffd\u52a0\u3055\u308c\u305f\u3053\u3068\u3067\u5fc5\u305a\u3057\u3082\u5358\u8a9e\u30d9\u30af\u30c8\u30eb\u540c\u58eb\u304c\u8fd1\u304f\u306a\u304f\u3066\u3082\u3001Linear\u5c64\u3067\u8fd1\u304f\u306a\u308b\u3088\u3046\u306b\u5909\u63db\u3059\u308b\u3088\u3046\u306b\u91cd\u307f\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u8abf\u7bc0\u3059\u308c\u3070\u3088\u3044\u3053\u3068\u304c\u308f\u304b\u308a\u307e\u3059\u3002\u3053\u308c\u304c\u3001Single-Head Attention\u3067\u3059\u3002\u3057\u304b\u3057\u3001Single-Head Attention\u306b\u3082\u8ab2\u984c\u304c\u3042\u308a\u307e\u3059\u3002\u79c1\u305f\u3061\u304c\u4f7f\u7528\u3059\u308b\u8a00\u8a9e\u306b\u306f\u3001\u540c\u3058\u5358\u8a9e\u3067\u3082\u610f\u5473\u3084\u6587\u6cd5\u306a\u3069\u304c\u7570\u306a\u308b\u3082\u306e\u304c\u591a\u6570\u3042\u308a\u307e\u3059\u304c\u3001Single-Head Attention\u3067\u306f\u3001\u305d\u308c\u3089\u3092\u5e73\u5747\u5316\u3057\u3066\u6271\u3063\u3066\u3057\u307e\u3044\u3001\u8907\u96d1\u306a\u6587\u7ae0\u69cb\u9020\u306b\u5bfe\u5fdc\u3067\u304d\u306a\u3044\u306e\u3067\u3059\u3002\u305d\u3053\u3067\u3001\u8003\u6848\u3055\u308c\u305f\u306e\u304c\u6b21\u306b\u8aac\u660e\u3059\u308bMulti-Head Attention\u3067\u3059\u3002<\/p>\n\n\n\n

Multi-Head Attention<\/h2>\n\n\n\n

Multi-Head Attention\u3068\u306f\u3001Single-Head Attention\u3092\u591a\u6570\u4e26\u5217\u306b\u914d\u7f6e\u3059\u308b\u3053\u3068\u3067\u3001\u3055\u307e\u3056\u307e\u306a\u6ce8\u610f\u8868\u73fe\u306e\u5b66\u7fd2\u3092\u53ef\u80fd\u306b\u3057\u305fAttention\u6a5f\u69cb\u3067\u3059\u3002<\/p>\n\n\n\n

\u539f\u8ad6\u6587\u306b\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u8a18\u8ff0\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

\n

Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions.<\/p>\n\u53c2\u8003\u6587\u732e[1]\u306e5\u30da\u30fc\u30b8\u306e\u5192\u982d<\/cite><\/blockquote>\n\n\n\n

\u3068\u8a18\u8f09\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u3059\u306a\u308f\u3061\u3001Single-Head Attention\u3067\u306f\u591a\u7a2e\u591a\u69d8\u306a\u610f\u5473\u3084\u6587\u6cd5\u3092\u3082\u3064\u5358\u8a9e\u306b\u5bfe\u3057\u3066\u3082\u5358\u4e00\u306e\u6ce8\u610f\u8868\u73fe\u306b\u5e73\u5747\u5316\u3055\u308c\u3066\u3057\u307e\u3044\u307e\u3059\u304c\u3001Single-Head Attention\u3092\u591a\u6570\u4e26\u5217\u306b\u914d\u7f6e\u3057\u3066Multi-Head\u306b\u3059\u308c\u3070\u3001\u8907\u6570\u306e\u7279\u5fb4\u90e8\u5206\u7a7a\u9593\u306b\u304a\u3051\u308b\u6ce8\u610f\u8868\u73fe\u306e\u7372\u5f97\u3092\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n

Multi-Head Attention\u306e\u69cb\u9020\u3092\u4e0b\u56f3\u306b\u793a\u3057\u307e\u3059\u3002<\/p>\n\n\n

\n
\"\"
Multi-Head Attention<\/figcaption><\/figure><\/div>\n\n\n

Multi-Head Attention\u306e\u4e2d\u306b\u3042\u308b\u500b\u3005\u306eSingle-Head Attention\u306f\u3001\u4e0b\u56f3\u306e\u3088\u3046\u306b\u305d\u308c\u305e\u308c\u7570\u306a\u308b\u7279\u5fb4\u3092\u5b66\u7fd2\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd[3,4]\u3067\u3059\u3002<\/p>\n\n\n

\n
\"\"
\u6587\u732e[3]\u3088\u308a\u5f15\u7528: Multi-Head Attention\u306b\u3088\u308b\u8907\u6570\u306e\u6ce8\u610f\u8868\u73fe\u306e\u7372\u5f97<\/figcaption><\/figure><\/div>\n\n\n
\"\"
[4]\u3088\u308a\u5f15\u7528<\/figcaption><\/figure>\n\n\n\n

\u3053\u306e\u3088\u3046\u306b\u3001\u6587\u7ae0\u5168\u4f53\u306b\u308f\u305f\u3063\u3066\u8907\u6570\u306e\u6ce8\u610f\u8868\u73fe\u306e\u7372\u5f97\u304c\u53ef\u80fd\u306aAttention\u6a5f\u69cb\u304c\u3001Multi-Head Attention\u3067\u3059\u3002\u3055\u3066\u3001Multi-Head Attention\u306e\u5229\u70b9\u3092\u3054\u7406\u89e3\u3044\u305f\u3060\u3044\u305f\u3068\u3053\u308d\u3067\u3001\u591a\u304f\u306e\u65b9\u306f\u5e7e\u3064\u306eSingle-Head Attention\u3092\u4e26\u5217\u306b\u3059\u308c\u3070\u3088\u3044\u304b\u3068\u3044\u3046\u7591\u554f\u3092\u6301\u3064\u3067\u3057\u3087\u3046\u3002\u591a\u3059\u304e\u308b\u3068\u5197\u9577\u306b\u306a\u308a\u3001\u5c11\u306a\u3059\u304e\u308b\u3068\u30e2\u30c7\u30eb\u306e\u8868\u73fe\u80fd\u529b\u304c\u4e0b\u304c\u3063\u3066\u3057\u307e\u3044\u307e\u3059\u3002Transformer\u306e\u539f\u8ad6\u6587[1]\u3067\u306f8\u3064\u306eSingle-Head Attention\u304c\u4e26\u5217\u306b\u7528\u3044\u3089\u308c\u3066\u3044\u307e\u3059\u3002\uff18\u3064\u3068\u805e\u3044\u3066\u5c11\u306a\u3044\u3068\u611f\u3058\u305f\u65b9\u304c\u591a\u3044\u306e\u3067\u306f\u306a\u3044\u3067\u3057\u3087\u3046\u304b\uff1f<\/p>\n\n\n\n

\u3053\u308c\u3067\u3082\u5341\u5206\u306a\u7cbe\u5ea6\u304c\u51fa\u308b\u7406\u7531\u306b\u3064\u3044\u3066\u79c1\u306a\u308a\u306e\u89e3\u91c8\u3092\u8ff0\u3079\u307e\u3059\u3002\u305d\u308c\u306f\u3001\u5404Single-Head Attention\u306e\u51fa\u529b\u3092Concat\u3057\u305f\u3068\u306b\u3001Linear\u5c64\u304c\u7528\u610f\u3055\u308c\u3066\u3044\u308b\u304b\u3089\u3060\u3068\u8003\u3048\u307e\u3059\u3002\u79c1\u305f\u3061\u306e\u4f4f\u3080\u4e16\u754c\u306f\uff13\u6b21\u5143\u3067\u3001\u3068\u3066\u3082\u5e83\u3044\u4e16\u754c\u304c\u5e83\u304c\u3063\u3066\u3044\u307e\u3059\u304c\u3001\u305d\u308c\u3089\u306fxyz\u306e\u5358\u4f4d\u30d9\u30af\u30c8\u30eb\u306e\u7dda\u5f62\u7d50\u5408\u3067\u8868\u3059\u3053\u3068\u304c\u3067\u304d\u3066\u3057\u307e\u3044\u307e\u3059\u3002\u305d\u308c\u3068\u540c\u69d8\u306b\u3001\u672c\u8cea\u7684\u306a\u6ce8\u610f\u8868\u73fe\u306f\uff18\u3064\u3057\u304b\u7372\u5f97\u3067\u304d\u306a\u304f\u3066\u3082\u3001\u305d\u306e\u5f8c\u306eConcat\u3068Linear\u5c64\u306b\u3088\u308a\u7d50\u5408\uff06\u7dda\u5f62\u5909\u63db\u3055\u308c\u308b\u306a\u304b\u3067\u3001\u65b0\u305f\u306a\u6ce8\u610f\u8868\u73fe\u306b\u5909\u63db\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u3060\u304b\u3089\u3067\u306f\u306a\u3044\u3067\u3057\u3087\u3046\u304b\uff1f<\/p>\n\n\n\n

\u5c11\u3057\u79c1\u306a\u308a\u306e\u89e3\u91c8\u3092\u6df7\u305c\u3066\u3057\u307e\u3044\u307e\u3057\u305f\u304c\u3001\u4ee5\u4e0a\u3092\u8e0f\u307e\u3048\u305f\u3068\u304d\u306eMulti-Head Attention\u306e\u8a08\u7b97\u5f0f\u306f\u4ee5\u4e0b\u306b\u793a\u3059\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

$$
\\begin{eqnarray} \\rm{MultiHead Attention}(\\boldsymbol{Q}, \\boldsymbol{K}, \\boldsymbol{V}) &=& \\rm{Concat}(head_1, head_2, \\cdots, head_h)\\boldsymbol{W}_o\\\\ \\rm{where}\\ head_i &=& \\rm{ScaledDotProductAttention}(\\boldsymbol{QW}^Q_i, \\boldsymbol{KW}^K_i, \\boldsymbol{VW}^V_i) \\end{eqnarray}
$$<\/p>\n\n\n\n

Transformer\u306e\u8ad6\u6587\u3067\u306f\u3001\u30d8\u30c3\u30c9\u306e\u5897\u52a0\u306b\u4f34\u3046\u8a08\u7b97\u91cf\u306e\u5897\u52a0\u3092\u9632\u3050\u305f\u3081\u306b\u3001Scaled Dot-Product Attention\u306e\u76f4\u524d\u306eLinear\u5c64\u3067\u5165\u529b\u30d9\u30af\u30c8\u30eb\u306e\u30b5\u30a4\u30ba\u3092\u30d8\u30c3\u30c9\u6570\\(h\\)\u3067\u9664\u7b97\u3057\u305f\u30b5\u30a4\u30ba\u306b\u5909\u63db\u3057\u3001Scaled Dot-Product Attention\u306e\u5165\u529b\u3068\u3057\u3066\u3044\u307e\u3059\u3002\u305d\u3057\u3066\u3001\\(h\\)\u500b\u306eSingle-Head Attention\u306e\u51fa\u529b\u3092\u7d50\u5408\u3057\u3066\u5165\u529b\u30d9\u30af\u30c8\u30eb\u3068\u540c\u3058\u30b5\u30a4\u30ba\u306b\u3057\u3066\u304b\u3089\u3001\u51fa\u529b\u76f4\u524d\u306eLinear\u5c64\u3067\u5909\u63db\u3092\u65bd\u3057\u3066\u6700\u7d42\u7684\u306a\u51fa\u529b\u3092\u8a08\u7b97\u3057\u3066\u3044\u307e\u3059\u3002\u4e0b\u56f3\u306f\u3001Self-Attention\u3068\u3057\u3066\u4f7f\u7528\u3059\u308b\u3068\u304d\u306eMulti-Head Attention\u5185\u90e8\u306b\u304a\u3051\u308b\u884c\u5217\u30b5\u30a4\u30ba\u306e\u5909\u5316\u3092\u8868\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n

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

\u307e\u3068\u3081<\/h2>\n\n\n\n

\u4ee5\u4e0a\u306e\u5185\u5bb9\u3092\u307e\u3068\u3081\u307e\u3059\u3002Transformer\u306e\u672c\u8cea\u7684\u306a\u6a5f\u80fd\u5358\u4f4d\u306fTransformer Block\u3068\u3088\u3070\u308c\u3001Multi-Head Attention\u3068Position-Wise Feed-Forward Networks\u306b\u5206\u3051\u3089\u308c\u307e\u3059\u3002Multi-Head Attention\u306f\u3001Scaled Dot-Product Attention\u3068Linear\u5c64\u304b\u3089\u306a\u308bSingle-Head Attention\u3092\u4e26\u5217\u5316\u3057\u305f\u3082\u306e\u3067\u3059\u3002<\/p>\n\n\n\n

Scaled Dot-Product Attention\u306f\u5185\u7a4d\u306b\u3088\u308bQuery\u30d9\u30af\u30c8\u30eb\u3068Key\u884c\u5217\u306e\u5404\u30d9\u30af\u30c8\u30eb\u9593\u306e\u985e\u4f3c\u6027\u306b\u57fa\u3065\u3044\u3066\u3001Value\u884c\u5217\u306e\u5404\u30d9\u30af\u30c8\u30eb\u306e\u7dda\u5f62\u7d50\u5408\u3092\u8a08\u7b97\u3057\u307e\u3059\u3002\u3057\u304b\u3057\u3001\u5b66\u7fd2\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u305f\u305a\u5b66\u7fd2\u306b\u3088\u308a\u7372\u5f97\u3055\u308c\u308b\u6ce8\u610f\u8868\u73fe\u3092\u5909\u3048\u308b\u3053\u3068\u304c\u3067\u304d\u306a\u3044\u305f\u3081\u3001\u5165\u529b\u90e8\u5206\u306b\u7dda\u5f62\u5c64\u3092\u8ffd\u52a0\u3057\u305f\u3001Single-Head Attention\u304c\u8003\u3048\u3089\u308c\u307e\u3057\u305f\u3002\u3057\u304b\u3057\u3001Single-Head Attention\u3067\u306f\u3001\u591a\u6570\u306e\u6ce8\u610f\u8868\u73fe\u304c\u5168\u3066\u5e73\u5747\u5316\u3055\u308c\u3066\u3057\u307e\u3046\u305f\u3081\u3001\u8907\u6570\u4e26\u5217\u306b\u3057\u3001\u305d\u308c\u3089\u306e\u51fa\u529b\u3092\u5408\u4f53\u3055\u305b\u3066Linear\u5c64\u306b\u3088\u308b\u5909\u63db\u3092\u884c\u3046\u3088\u3046\u306b\u3057\u305fMulti-Head Attention\u304c\u63d0\u6848\u3055\u308c\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

Multi-Head Attention\u306f\u591a\u6570\u306e\u6ce8\u610f\u8868\u73fe\u3092\u7372\u5f97\u53ef\u80fd\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u4ee5\u4e0a\u3067\u672c\u8a18\u4e8b\u306e\u5185\u5bb9\u3092\u7d42\u308f\u308a\u306b\u3057\u307e\u3059\u3002\u6700\u5f8c\u307e\u3067\u304a\u8aad\u307f\u3044\u305f\u3060\u304d\u3042\u308a\u304c\u3068\u3046\u3054\u3056\u3044\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

\u53c2\u8003\u6587\u732e<\/h2>\n\n\n\n

[1] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin, \"Attention is all you need,\" in Proc. NeurIPS<\/em>, 2017.
[2] Wenyng Duan, Liu Jiang, Ning Wang, and Hong Rao, \"Pre-Trained Bidirectional Temporal Representation for Crowd Flows Prediction in Regular Region,\" IEEE Access<\/em>, 2019.
[3] Hyeongu Yun, Taegwan Kang, and Kyomin Jung, \"Analyzing and Controlling Inter-Head Diversity in Multi-Head Attention,\" Appl. Sci<\/em>, 2021.
[4] Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning, \"What Does BERT Look At? An Analysis of BERT's Attention,\" in Proc. ACL<\/em>, 2019.
[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, \"Deep Residual Learning for Image Recognition,\" in Proc. CVPR<\/em>, 2016.<\/p>\n","protected":false},"excerpt":{"rendered":"

\u7686\u3055\u3093\u3053\u3093\u306b\u3061\u306f\uff01 \u53bb\u5e74\uff082022\u5e7411\u6708\uff09\u306b\u767a\u8868\u3055\u308c\u305fChatGPT\u306e\u8a71\u984c\u304c\u5c3d\u304d\u306a\u3044\u4eca\u65e5\u3053\u306e\u9803\u3001ChatGPT\u3067\u4f7f\u308f\u308c\u3066\u3044\u308b\u91cd\u8981\u306a\u6280\u8853\u306e\uff11\u3064\u3067\u3042\u308bTransformer\u306b\u3064\u3044\u3066\u3001\u8208\u5473\u3092\u6301\u3063\u305f\u65b9\u306f\u6ca2\u5c71 … <\/p>\n","protected":false},"author":1,"featured_media":7621,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_vk_print_noindex":"","sitemap_hide":"","_veu_custom_css":"","veu_display_promotion_alert":"common","vkexunit_cta_each_option":"","_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[170,171,106,174,206],"tags":[],"class_list":["post-7546","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog","category-learn","category-ml","category-dl","category-206"],"veu_head_title_object":{"title":"","add_site_title":""},"jetpack_featured_media_url":"https:\/\/developers.agirobots.com\/jp\/wp-content\/uploads\/2023\/02\/0E0D472B-78CA-45C4-AB6F-39BBD367B8AC.jpeg","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/7546","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/comments?post=7546"}],"version-history":[{"count":47,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/7546\/revisions"}],"predecessor-version":[{"id":10562,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/7546\/revisions\/10562"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/media\/7621"}],"wp:attachment":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/media?parent=7546"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/categories?post=7546"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/tags?post=7546"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}