{"id":8781,"date":"2023-08-14T01:57:28","date_gmt":"2023-08-13T16:57:28","guid":{"rendered":"https:\/\/agirobots.com\/?p=8781"},"modified":"2023-10-24T10:28:37","modified_gmt":"2023-10-24T01:28:37","slug":"attention-free-transformer","status":"publish","type":"post","link":"https:\/\/developers.agirobots.com\/jp\/attention-free-transformer\/","title":{"rendered":"Attention Free Transformer\u306b\u3064\u3044\u3066\u89e3\u8aac"},"content":{"rendered":"\n

\u306f\u3058\u3081\u306b<\/h2>\n\n\n\n

\u672c\u8a18\u4e8b\u3067\u306f\u3001Attention Free Transformer[1]\u3068\u547c\u3070\u308c\u308b\u30e2\u30c7\u30eb\u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u3064\u3044\u3066\u89e3\u8aac\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u306f\u3058\u3081\u306b\u5927\u67a0\u3092\u8aac\u660e\u3057\u307e\u3059\u3002\u8fd1\u5e74\u3067\u306fTransformer[2]\u3068\u547c\u3070\u308c\u308b\u624b\u6cd5\u304c\u5927\u304d\u306a\u6210\u679c\u3092\u3042\u3052\u3066\u3044\u307e\u3059\u3002Transformer\u306e\u4e2d\u5fc3\u7684\u306a\u4ed5\u7d44\u307f\u306f\u6ce8\u610f\u6a5f\u69cb[2]\u3067\u3042\u308a\u3001\u305d\u306e\u6ce8\u610f\u6a5f\u69cb\u306f\u5185\u90e8\u306b\u5185\u7a4d\u6ce8\u610f\u3068\u547c\u3070\u308c\u308b\u4ed5\u7d44\u307f\u3092\u6301\u3063\u3066\u3044\u307e\u3059\u3002\u3053\u306e\u5185\u7a4d\u6ce8\u610f\u304cTransformer\u306e\u9ad8\u3044\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u306e\u6e90\u3067\u3042\u308b\u3068\u8003\u3048\u3089\u308c\u3066\u3044\u308b\u4e00\u65b9\u3067\u3001\u305d\u308c\u304cTransformer\u3092\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u3059\u308b\u3046\u3048\u3067\u5927\u304d\u306a\u30dc\u30c8\u30eb\u30cd\u30c3\u30af\u3068\u306a\u3063\u3066\u3044\u308b\u73fe\u72b6\u304c\u3042\u308a\u307e\u3059\u3002\u7406\u7531\u306f\u3001\u5185\u7a4d\u6ce8\u610f\u3068\u306f\u5165\u529b\u3055\u308c\u305f\u5168\u3066\u306e\u30c8\u30fc\u30af\u30f3\u306e\u7d44\u307f\u5408\u308f\u305b\u306b\u5bfe\u3057\u3066\u3001\u5185\u7a4d\u3092\u8a08\u7b97\u3059\u308b\u5168\u6ce8\u610f\u3068\u547c\u3070\u308c\u308b\u65b9\u6cd5\u3092\u63a1\u7528\u3057\u3066\u3044\u308b\u304b\u3089\u3067\u3059\u3002\u5185\u7a4d\u306e\u8a08\u7b97\u7d50\u679c\u306f\u3001\u884c\u5217\u5f62\u5f0f\u3067\u4fdd\u6301\u3055\u308c\u3001\u3053\u308c\u3092\u6ce8\u610f\u884c\u5217\u3068\u547c\u3073\u307e\u3059\u3002\u5185\u7a4d\u8a08\u7b97\u306e\u6570\u304a\u3088\u3073\u6ce8\u610f\u884c\u5217\u306e\u6210\u5206\u6570\u306f\u3001\u7cfb\u5217\u9577\u306b\u5bfe\u3057\u30662\u4e57\u306e\u30aa\u30fc\u30c0\u3067\u5897\u52a0\u3057\u307e\u3059\u3002\u3064\u307e\u308a\u3001\u9577\u3044\u5165\u529b\u7cfb\u5217\u3092\u6271\u3046\u306b\u306f\u3001\u305d\u308c\u306a\u308a\u306e\u8a08\u7b97\u91cf\u3068\u30e1\u30e2\u30ea\u5bb9\u91cf\u304c\u5fc5\u9808\u306b\u306a\u308a\u307e\u3059\u3002\u3053\u308c\u3092\u89e3\u6c7a\u3059\u308b\u65b9\u6cd5\u3068\u3057\u3066\u3001\u5185\u7a4d\u6ce8\u610f\u3092\u8fd1\u4f3c\u3059\u308b\u3053\u3068\u3066\u8a08\u7b97\u30b3\u30b9\u30c8\u3092\u8efd\u91cf\u5316\u3059\u308b\u624b\u6cd5\u3068\u3001\u5185\u7a4d\u6ce8\u610f\u3059\u3089\u4f7f\u7528\u3057\u306a\u3044\u65b0\u3057\u3044\u6ce8\u610f\u6a5f\u69cb\u3092\u5b9f\u73fe\u3059\u308b\u624b\u6cd5\u304c\u3068\u3089\u308c\u3066\u3044\u307e\u3059\u3002\u672c\u8a18\u4e8b\u3067\u6271\u3046Attention Free Transformer\u306f\u3001\u5185\u7a4d\u6ce8\u610f\u3092\u4f7f\u7528\u3057\u306a\u3044Transformer\u3067\u3059\u3002\u6ce8\u610f\u6a5f\u69cb\u304c\u7121\u3044Transformer\u3067\u3042\u308b\u3068\u3044\u3046\u8aa4\u89e3\u3092\u751f\u307f\u305d\u3046\u306a\u540d\u524d\u3092\u3057\u3066\u3044\u307e\u3059\u304c\u3001\u5185\u7a4d\u6ce8\u610f\uff08\u53b3\u5bc6\u306b\u306f\u5185\u7a4d\u6ce8\u610f\u3092\u7528\u3044\u305fMulti-Head Attention\uff09\u3092\u4f7f\u7528\u3057\u306a\u3044\u3068\u3044\u3046\u3060\u3051\u3067\u3059\u3002Attention Free Transformer\u306f\u3001\u5225\u306e\u8a18\u4e8b\u3067\u89e3\u8aac\u3059\u308bRWKV[3]\u306a\u3069\u306e\u30e2\u30c7\u30eb\u306e\u3082\u3068\u3068\u306a\u3063\u3066\u304a\u308a\u91cd\u8981\u306a\u624b\u6cd5\u3067\u3059\u3002<\/p>\n\n\n\n

\u6982\u8981\u306f\u4ee5\u4e0a\u3067\u3059\u304c\u3001\u3053\u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u3064\u3044\u3066\u8a73\u3057\u304f\u8aac\u660e\u3059\u308b\u306b\u306f\u3001\u5143\u7956Transformer\u306e\u5185\u7a4d\u6ce8\u610f\u304b\u3089\u3057\u3063\u304b\u308a\u3068\u7406\u89e3\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002\u305d\u3053\u3067\u3001\u672c\u8a18\u4e8b\u3067\u306f\u3001\u5185\u7a4d\u6ce8\u610f\u306b\u3064\u3044\u3066\u8a73\u3057\u304f\u8aac\u660e\u3057\u305f\u3046\u3048\u3067\u3001Attention Free Transformer\u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u3064\u3044\u3066\u8aac\u660e\u3057\u3066\u3044\u304d\u307e\u3059\u3002<\/p>\n\n\n\n

\u5185\u7a4d\u6ce8\u610f<\/h2>\n\n\n\n

\u307e\u305a\u306f\u3001\u5143\u7956\u306e\u6ce8\u610f\u6a5f\u69cb\u3067\u3042\u308b\u5185\u7a4d\u6ce8\u610f\u306b\u3064\u3044\u3066\u8aac\u660e\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u5185\u7a4d\u6ce8\u610f\u306f\u3001Query\u3068Key\u3001Value\u3068\u547c\u3070\u308c\u308b3\u3064\u306e\u5165\u529b\u3092\u53d7\u3051\u53d6\u308a\u307e\u3059\u3002\u305d\u3057\u3066\u3001Query\u3068Key\u306e\u5185\u7a4d\u3092\u8a08\u7b97\u3057\u3001\uff08\u30bd\u30d5\u30c8\u30de\u30c3\u30af\u30b9\u95a2\u6570\u3092\u9069\u7528\u3057\u3066\uff09\u6ce8\u610f\u884c\u5217\u3092\u751f\u6210\u3057\u307e\u3059\u3002\u3053\u306e\u6ce8\u610f\u884c\u5217\u306e\u5404\u6210\u5206\u3092\u4fc2\u6570\u3068\u3057\u3001Value\u3092\u7dda\u5f62\u7d50\u5408\u3057\u307e\u3059\u3002\u3053\u308c\u304c\u5185\u7a4d\u6ce8\u610f\u3067\u3059\u3002\u3088\u3063\u3066\u3001\u51fa\u529b\u306e\u7cfb\u5217\u9577\u306fQuery\u306e\u7cfb\u5217\u9577\u3068\u540c\u3058\u3067\u3001\u51fa\u529b\u306e\u30d9\u30af\u30c8\u30eb\u30b5\u30a4\u30ba\u306fValue\u306e\u30d9\u30af\u30c8\u30eb\u30b5\u30a4\u30ba\u3068\u540c\u3058\u3067\u3059\u3002<\/p>\n\n\n\n

\u4ee5\u4e0b\u306b\u3001\\(t\\)\u756a\u76ee\u51fa\u529b\u30d9\u30af\u30c8\u30eb\u306e\u8a08\u7b97\u5f0f\u3092\u793a\u3057\u307e\u3059\u3002 <\/p>\n\n\n\n

$$\\begin{eqnarray}
\\text{Attention}(\\boldsymbol{Q}, \\boldsymbol{K}, \\boldsymbol{V})_t &=& \\left(\\text{softmax}\\left(\\frac{\\boldsymbol{QK}^N}{\\sqrt{d}}\\right)\\boldsymbol{V}\\right)_t \\\\
&=& \\frac{\\sum_{n=1}^N \\text{exp}(\\frac{\\boldsymbol{q}_t\\boldsymbol{k}_n^T}{\\sqrt{d}}\\boldsymbol{v}_n)}{\\sum_{n=1}^N \\text{exp}(\\frac{\\boldsymbol{q}_t\\boldsymbol{k}_n^T}{\\sqrt{d}})} \\in \\mathbb{R}^{d}
\\end{eqnarray}$$ <\/p>\n\n\n\n

\u5f0f\u4e2d\u3001\\(\\boldsymbol{Q, K, V}\\)\u306f\u3001\u884c\u30d9\u30af\u30c8\u30eb\u3067\u8868\u3055\u308c\u305f\u30c8\u30fc\u30af\u30f3\u30d9\u30af\u30c8\u30eb\u3092\u4e26\u3079\u3066\u884c\u5217\u306b\u3057\u305f\u3082\u306e\u3067\u3001\\(\\boldsymbol{q}_t\\)\u306a\u3069\u306eQuery\u3001Key\u3001Value\u306e\u305d\u308c\u305e\u308c\u306b\u5bfe\u5fdc\u3059\u308b\u5c0f\u6587\u5b57\u306f\u30c8\u30fc\u30af\u30f3\u30d9\u30af\u30c8\u30eb\u3092\u8868\u3057\u307e\u3059\u3002\u307e\u305f\u3001\\(N\\)\u306f\u7cfb\u5217\u9577\uff08\u4e00\u822c\u7684\u306b\u306f\\(T\\)\u3092\u7528\u3044\u308b\u304c\u8ee2\u7f6e\u306e\u8a18\u53f7\u3068\u91cd\u306a\u308b\u305f\u3081\\(N\\)\u3068\u3057\u305f\uff09\u3001\\(d\\)\u306fKey\u306e\u30d9\u30af\u30c8\u30eb\u30b5\u30a4\u30ba\u3092\u8868\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u3088\u308a\u8a73\u7d30\u306a\u8a08\u7b97\u306e\u30a4\u30e1\u30fc\u30b8\u3092\u6301\u3063\u3066\u3044\u305f\u3060\u304f\u305f\u3081\u306b\u3001\u4ee5\u4e0b\u306b\u8a08\u7b97\u5f0f\u3092\u30a4\u30e9\u30b9\u30c8\u3067\u8868\u3057\u305f\u3082\u306e\u3092\u8f09\u305b\u307e\u3059\u3002<\/p>\n\n\n\n

\"\" <\/p>\n\n\n\n

\u6700\u5f8c\u306e\u90e8\u5206\u306e\u307f\u306b\u6ce8\u76ee\u3057\u3066\u3082\u3089\u3046\u3068\u3001\u5185\u7a4d\u6ce8\u610f\u304c\u4f55\u3092\u8a08\u7b97\u3059\u308b\u3082\u306e\u304b\u3092\u30a4\u30e1\u30fc\u30b8\u3057\u3084\u3059\u3044\u3067\u3057\u3087\u3046\u3002\u307e\u305a\u7d76\u5bfe\u306b\u7406\u89e3\u3057\u3066\u304a\u304f\u3079\u304d\u4e8b\u306e\uff11\u3064\u76ee\u306f\u3001\u51fa\u529b\u306fValue\u306e\u91cd\u307f\u4ed8\u304d\u548c\u3067\u3042\u308b\u3053\u3068\u3067\u3059\u3002\uff12\u3064\u76ee\u306f\u3001\u305d\u306e\u7dda\u5f62\u7d50\u5408\u306b\u304a\u3051\u308b\u4fc2\u6570\u3092\u6c7a\u3081\u308b\u969b\u306b\u3001Key\u3068Value\u306e\u5185\u7a4d\u306b\u57fa\u3065\u304f\u6ce8\u610f\u884c\u5217\u3092\u8a08\u7b97\u3057\u3066\u3044\u308b\u70b9\u3067\u3059\u3002<\/p>\n\n\n\n

\u305d\u3057\u3066\u3001\u3053\u306e\u5185\u7a4d\u3092\u8a08\u7b97\u3059\u308b\u306e\u304c\u7269\u51c4\u304f\u5927\u5909\u306a\u306e\u3067\u3059\u3002\u4f8b\u3048\u3070\u30011\u4e07\u500b\u306e\u30c8\u30fc\u30af\u30f3\u5217\u3092\u5165\u529b\u3068\u3059\u308b\u5834\u5408\u3001\u6ce8\u610f\u884c\u5217\u3092\u751f\u6210\u3059\u308b\u306b\u306f\u30011\u4e07\u00d71\u4e07=1\u5104\u306e\u5185\u7a4d\u8a08\u7b97\u304c\u5fc5\u8981\u306b\u306a\u308a\u3001\u304b\u3064\u305d\u308c\u3092\u4fdd\u6301\u3059\u308b\u30e1\u30e2\u30ea\u3082\u305d\u308c\u3060\u3051\u5fc5\u8981\u306b\u306a\u308a\u307e\u3059\u30021\u3064\u306e\u6210\u5206\u3092\u4fdd\u5b58\u3059\u308b\u306b32bit\uff08=4\u30d0\u30a4\u30c8\uff09\u4f7f\u7528\u3059\u308b\u5834\u5408\u30011\u5104\u500b\u306e\u6210\u5206\u3092\u6301\u3064\u884c\u5217\u306e\u4fdd\u5b58\u306b\u5fc5\u8981\u306a\u5bb9\u91cf\u306f\u7d04400MB\u306b\u306a\u308a\u307e\u3059\u3002\u305d\u308c\u306b\u30011\u5c64\u306e\u307f\u3067\u6e08\u3080\u8a71\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u304b\u3089\u3001\u4f55\u5341\u5c64\u3082\u30b9\u30bf\u30c3\u30af\u3059\u308b\u3053\u3068\u3092\u8003\u3048\u308b\u3068\u3001\u5fc5\u8981\u306a\u8a08\u7b97\u30b3\u30b9\u30c8\u3084\u30e1\u30e2\u30ea\u30fc\u30b5\u30a4\u30ba\u306f\u6050\u308d\u3057\u3044\u307b\u3069\u81a8\u5927\u3067\u3059\u3002\u5185\u7a4d\u6ce8\u610f\u3092\u8efd\u91cf\u5316\u3057\u306a\u3051\u308c\u3070\u3001Transformer\u3092\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u3059\u308b\u306e\u304c\u96e3\u3057\u3044\u3053\u3068\u3092\u3054\u7406\u89e3\u3044\u305f\u3060\u3051\u308b\u306e\u3067\u306f\u306a\u3044\u304b\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u3061\u306a\u307f\u306b\u3001\u4eca\u56de\u7d39\u4ecb\u3059\u308b\u3001Attention Free Transformer\u306f\u3001\u5185\u7a4d\u6ce8\u610f\u3060\u3051\u3067\u306f\u306a\u304f\u3001Multi-Head Attention\u5168\u4f53\u3092\u7f6e\u304d\u63db\u3048\u308b\u624b\u6cd5\u306a\u306e\u3067\u3001Multi-Head Attention\u306b\u3064\u3044\u3066\u3082\u89e3\u8aac\u3057\u3066\u304a\u304d\u307e\u3059\u3002Multi-Head Attention\u306f\u5185\u7a4d\u6ce8\u610f\uff08Scaled Dot-Product Attention\uff09\u3092\u7528\u3044\u3066\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u8a08\u7b97\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

$$\\begin{eqnarray}
\\text{MultiHead Attention}(\\boldsymbol{Q}, \\boldsymbol{K}, \\boldsymbol{V}) &=& \\text{Concat}(head_1, head_2, \\cdots, head_h)\\boldsymbol{W_O}\\\\
\\text{where}\\ head_i &=& \\text{ScaledDotProductAttention}(\\boldsymbol{QW^Q}_i, \\boldsymbol{KW^K}_i, \\boldsymbol{VW^V}_i)
\\end{eqnarray}$$<\/p>\n\n\n\n

\u7279\u306b\u96e3\u3057\u3044\u3053\u3068\u306f\u884c\u3063\u3066\u304a\u3089\u305a\u3001\u5185\u7a4d\u6ce8\u610f\u306eQuery, Key, Value\u306e\u5404\u5165\u529b\u306e\u76f4\u524d\u306b\u91cd\u307f\u3092\\(\\boldsymbol{W^Q}, \\boldsymbol{W^K}, \\boldsymbol{W^V}\\)\u3068\u3059\u308b\u7dda\u5f62\u5c64\u3092\u8ffd\u52a0\u3057\u305fSingle-Head Attention\u304c\u3001\u4e26\u5217\u306b\u4e26\u3093\u3067\u3044\u3066\u3001\u4efb\u610f\u306e\u30d8\u30c3\u30c9\u306e\u51fa\u529b\u3092\\(\\rm{head}_i\\)\u3068\u3059\u308b\u3068\u304d\u3001\u305d\u308c\u3089\u3092\u5408\u4f53\u3057\u3066\u3001\u91cd\u307f\u3092\\(\\boldsymbol{W_O}\\)\u3068\u3059\u308b\u7dda\u5f62\u5c64\u306b\u9069\u7528\u3057\u3066\u3044\u308b\u3060\u3051\u3067\u3059\u3002\u8981\u3059\u308b\u306b\u3001\u591a\u7a2e\u591a\u69d8\u306a\u7279\u5fb4\u3092\u5b66\u7fd2\u3067\u304d\u308b\u3088\u3046\u306b\u4e26\u5217\u5316\u3057\u305f\u306e\u3067\u3059\u3002\u56f3\u306b\u63cf\u304f\u3068\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u611f\u3058\u3067\u3059\u3002<\/p>\n\n\n

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

Multi-Head Attention\u306b\u3064\u3044\u3066\u8a73\u3057\u304f\u77e5\u308a\u305f\u3044\u65b9\u306f\u4ee5\u4e0b\u306e\u8a18\u4e8b\u3092\u53c2\u8003\u306b\u3057\u3066\u4e0b\u3055\u3044\u3002<\/p>\n\n\n\n

\n
\"\"<\/dt>

\u3010Transformer\u306e\u57fa\u790e\u3011Multi-Head Attention\u306e\u4ed5\u7d44\u307f<\/p>

\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 \u672c\u8a18\u4e8b\u306e\u69cb\u6210\u306f\u3001\u306f\u3058\u3081\u306bTransformer\u304a\u3088\u3073Transformer ... <\/p><\/div><\/dd><\/dl><\/div><\/a><\/div>\n<\/div><\/figure>\n\n\n\n

Attention Free Transformer<\/h2>\n\n\n\n

Attention Free Transformer\uff08\u4ee5\u964d\u3001AFT\uff09\u306e\u8ad6\u6587\u3067\u306f\u3001AFT\u306b\u3064\u3044\u3066Transformer\u306eMulti-Head Attention\u306e\u7f6e\u304d\u63db\u3048\u624b\u6cd5\u3068\u3057\u3066\u63d0\u6848\u3057\u3066\u3044\u307e\u3059\u3002\u3064\u307e\u308a\u3001\u4e0b\u56f3\u306e\u5de6\u5074\u3092\u53f3\u5074\u3067\u7f6e\u304d\u63db\u3048\u308b\u306e\u3067\u3059\u3002\u540d\u524d\u306bTransformer\u3068\u3064\u3044\u3066\u3044\u308b\u306e\u3067\u3001\u3064\u3044\u3001Transformer\u5168\u4f53\u306e\u3053\u3068\u3092\u6307\u3057\u3066\u3044\u308b\u306e\u304b\u3068\u601d\u3063\u3066\u3057\u307e\u3044\u307e\u3059\u304c\u3001\u8ad6\u6587\u306b\u8a18\u8f09\u306e\u5b9a\u7fa9\u304b\u3089\u306f\u3001\u3042\u304f\u307e\u3067\u3082Multi-Head Attention\u306e\u7f6e\u304d\u63db\u3048\u624b\u6cd5\u306b\u3059\u304e\u307e\u305b\u3093\u3002<\/p>\n\n\n\n

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

\u3053\u3053\u3067\u306f\u3001\u4e00\u65e6\u3001Self-Attention\u578b\u306eAttention Free Transformer\u3092\u8003\u3048\u3066\u3044\u304f\u3053\u3068\u306b\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u5165\u529b\u3092\\(\\boldsymbol{X}\\)\u3001\u305d\u308c\u3089\u3092\u5909\u63db\u3059\u308bQuery, Key, Value\u306e\u7dda\u5f62\u5c64\u306e\u91cd\u307f\u3092\\(\\boldsymbol{W^Q}, \\boldsymbol{W^K}, \\boldsymbol{W^V}\\)\u3068\u3059\u308b\u3068\u304d\u3001\\(\\boldsymbol{X}\\)\u306f\u3001\\( \\boldsymbol{Q} = \\boldsymbol{XW^Q}\\)\u3001\\( \\boldsymbol{K} = \\boldsymbol{XW^K}\\)\u3001\\(\\boldsymbol{V} = \\boldsymbol{XW^V}\\)\u306b\u5909\u63db\u3055\u308c\u307e\u3059\u3002\u305d\u306e\u3046\u3048\u3067\u3001\u4ee5\u4e0b\u306e\u95a2\u6570\u3092\u9069\u7528\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

$$\\begin{eqnarray}
Y(\\boldsymbol{Q}, \\boldsymbol{K}, \\boldsymbol{V})_t = \\sigma(\\boldsymbol{q}_t)\\odot\\frac{\\sum_{i=1}^T \\text{exp}(w_{t, i} + \\boldsymbol{k}_i)\\odot\\boldsymbol{v}_i}{\\sum_{i=1}^T \\text{exp}(w_{t,i} + \\boldsymbol{k}_i)}\\in \\mathbb{R}^{d}
\\end{eqnarray}$$<\/p>\n\n\n\n

\u3053\u306e\u5f0f\u306f\u3001\u6642\u523b\\(t\\)\u306b\u304a\u3051\u308b\u8a08\u7b97\u5f0f\u3092\u8868\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u3053\u3067\u7279\u5fb4\u7684\u306a\u306e\u306f\u3001\u65b0\u3057\u304f\u30d1\u30e9\u30e1\u30fc\u30bf\\(w_{t, i}\\)\u3092\u5c0e\u5165\u3057\u3066\u3044\u308b\u3053\u3068\u3067\u3059\u3002\u3053\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u306f\u3001\\( \\boldsymbol{w}\\in\\mathbb{R}^{T\\times T}\\)\u306b\u304a\u3044\u3066\u3001\u6642\u523b\\(t\\)\u306e\\(i\\)\u756a\u76ee\u306e\u6210\u5206\u3092\u53d6\u308a\u51fa\u3057\u3066\u304d\u305f\u3082\u306e\u3067\u3059\u3002\u5f93\u6765\u306f\u3001Query\u3068Key\u306e\u5185\u7a4d\u3092\u8a08\u7b97\u3059\u308b\u3053\u3068\u3067\\(T\\times T\\)\u306e\u6ce8\u610f\u884c\u5217\u3092\u8a08\u7b97\u3057\u3066\u3044\u307e\u3057\u305f\u304c\u3001\u5185\u7a4d\u8a08\u7b97\u81ea\u4f53\u3092\u3084\u3081\u3066\u3057\u307e\u3046\u4ee3\u308f\u308a\u306b\u3001\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f\u306b\u985e\u4f3c\u3057\u305f\u6a5f\u69cb\u3092\u5c0e\u5165\u3059\u308b\u3068\u3044\u3046\u8003\u3048\u65b9\u3067\u3059\u3002\u305d\u3057\u3066\u3001Query\u306f\u3001LSTM\u3084GRU\u306a\u3069\u3067\u898b\u3089\u308c\u308b\u3088\u3046\u306a\u30b2\u30fc\u30c8\u3068\u3057\u3066\u5229\u7528\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u6539\u3081\u3066\u3001AFT\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u8a08\u7b97\u5f0f\u3068\u3057\u3066\u5b9a\u7fa9\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

$$\\begin{eqnarray}
\\text{AFT}(\\boldsymbol{X})_t &=& \\sigma(\\boldsymbol{q}_t)\\odot\\frac{\\sum_{i=1}^T \\text{exp}(w_{t, i} + \\boldsymbol{k}_i)\\odot\\boldsymbol{v}_i}{\\sum_{i=1}^T \\text{exp}(w_{t,i} + \\boldsymbol{k}_i)}\\in \\mathbb{R}^{d}\\\\
&& \\text{where }\\boldsymbol{q}_t = (\\boldsymbol{XW^Q})_t, \\boldsymbol{k}_t = (\\boldsymbol{XW^K})_t, \\boldsymbol{v}_t = (\\boldsymbol{XW^V})_t
\\end{eqnarray}$$<\/p>\n\n\n\n

\u3053\u306e\u8a08\u7b97\u3092\u56f3\u3067\u8868\u3059\u3068\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

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

\u56f3\u306b\u3057\u3066\u307f\u305f\u3082\u306e\u306e\u3001\u306a\u305c\u3053\u306e\u3088\u3046\u306a\u8a08\u7b97\u5f0f\u304c\u601d\u3044\u3064\u304f\u306e\u304b\u3001\u3068\u3044\u3046\u7591\u554f\u304c\u79c1\u306b\u306f\u6b8b\u3063\u3066\u3044\u307e\u3059\u304c\u3001\u5185\u7a4d\u8a08\u7b97\u3092\u7121\u304f\u3057\u305f\u4ee3\u308f\u308a\u306b\u5c11\u3057\u8907\u96d1\u306a\u4ed5\u7d44\u307f\u3092\u5c0e\u5165\u3057\u305f\u3093\u3060\u308d\u3046\u306a\u3001\u3068\u3044\u3046\u89e3\u91c8\u3067\u7559\u3081\u3066\u304a\u308a\u307e\u3059...<\/p>\n\n\n\n

AFT\u306e\u7a2e\u985e<\/h2>\n\n\n\n

AFT\u306b\u306f\u5e7e\u3064\u304b\u306e\u7a2e\u985e\u304c\u8003\u3048\u3089\u308c\u3066\u3044\u307e\u3059\u3002\u4e0a\u3067\u7d39\u4ecb\u3057\u305f\u3082\u306e\u304c\u57fa\u672c\u7684\u306a\u5f0f\u3068\u306a\u308a\u3001\u305d\u308c\u3092AFT-full\u3068\u547c\u3073\u307e\u3059\u3002\u3053\u3053\u3067\u3001\\(w_{t,i}\\)\u306b\u304a\u3044\u3066\u3001\u7a93\u3092\u6301\u305f\u305b\u3001\u305d\u306e\u7bc4\u56f2\u5916\u3067\u306f0\u306b\u3059\u308b\u3053\u3068\u3067\u3001\u5c40\u6240\u6027\u3092\u6301\u305f\u305b\u305f\u3082\u306e\u304c\u3042\u308a\u3001\u305d\u308c\u3092AFT-local\u3001\u305d\u3082\u305d\u3082\u91cd\u307f\u81ea\u4f53\u3092\u7121\u304f\u3057\u3066\u30b7\u30f3\u30d7\u30eb\u306b\u3057\u305fAFT-simple\u3001\u7573\u307f\u8fbc\u307f\u3092\u8ffd\u52a0\u3057\u305f\u3001AFT-conv\u306a\u3069\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

\u8208\u5473\u304c\u3042\u308a\u307e\u3057\u305f\u3089\u3001\u8ad6\u6587\u3092\u898b\u3066\u307f\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n\n

\u3055\u3044\u3054\u306b<\/h2>\n\n\n\n

\u5185\u5bb9\u306f\u4ee5\u4e0a\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

\u672c\u8a18\u4e8b\u3092\u901a\u3058\u3066\u3001AFT\u304c\u3069\u306e\u3088\u3046\u306a\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3092\u7528\u3044\u3066\u5f93\u6765\u306e\u5185\u7a4d\u6ce8\u610f\u3092\u4ee3\u66ff\u3057\u3066\u3044\u308b\u306e\u304b\u3092\u3054\u7406\u89e3\u3044\u305f\u3060\u3051\u307e\u3057\u305f\u3067\u3057\u3087\u3046\u304b\uff1f<\/p>\n\n\n\n

AFT\u306e\u5185\u7a4d\u6ce8\u610f\u81ea\u4f53\u3092\u7f6e\u304d\u63db\u3048\u3066\u3057\u307e\u3046\u3068\u3044\u3046\u30a2\u30d7\u30ed\u30fc\u30c1\u306f\u3001\u3072\u3087\u3063\u3068\u3057\u305f\u3089Transformer\u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u3064\u3044\u3066\u3042\u308b\u7a0b\u5ea6\u7406\u89e3\u3057\u3066\u3044\u306a\u3044\u3068\u8e93\u3044\u3066\u3057\u307e\u3046\u96e3\u3057\u3044\u65b9\u6cd5\u304b\u3082\u3057\u308c\u307e\u305b\u3093\uff08\u79c1\u81ea\u8eab\u3082\u307e\u3060\u5b8c\u5168\u306b\u7406\u89e3\u3057\u305f\u3068\u306f\u601d\u3063\u3066\u3044\u307e\u305b\u3093\uff09\u3002AFT\u306f\u5225\u306e\u8a18\u4e8b\u3067\u7d39\u4ecb\u3059\u308bRWKV\u306e\u30d9\u30fc\u30b9\u3068\u306a\u308b\u624b\u6cd5\u3068\u306a\u3063\u3066\u3044\u308b\u3088\u3046\u306b\u3001\u5b9f\u306f\u91cd\u8981\u306a\u4ed5\u7d44\u307f\u3067\u3059\u3002\u3068\u306f\u3044\u3048\u3001\u91cd\u8981\u306a\u306e\u306fAFT\u306e\u7d30\u304b\u3044\u8a08\u7b97\u65b9\u6cd5\u3068\u3044\u3046\u3088\u308a\u3001Query\u306b\u3088\u308b\u30b2\u30fc\u30c6\u30a3\u30f3\u30b0\u3068WKV\uff08Weighted Key Value\uff09\u306b\u3088\u308b\u65b0\u3057\u3044\u6ce8\u610f\u6a5f\u69cb\u306e\u8003\u3048\u65b9\u3067\u3042\u308b\u3088\u3046\u306b\u611f\u3058\u307e\u3059\u3002\u3067\u3059\u306e\u3067\u3001\u6700\u4f4e\u3067\u3082\u305d\u306e\u70b9\u3092\u3054\u7406\u89e3\u3044\u305f\u3060\u3051\u308c\u3070\u5341\u5206\u304b\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u3082\u3057\u8208\u5473\u304c\u3042\u308a\u307e\u3057\u305f\u3089\u3001RWKV\u306e\u8a18\u4e8b\u3082\u8aad\u3093\u3067\u3044\u305f\u3060\u3051\u308b\u3068\u5e78\u3044\u3067\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] Shuangfei Zhai, Walter Talbott, Nitish Srivastava, Chen Huang, Hanlin Goh, Ruixiang Zhang, and Josh Susskind, \"An Attention Free Transformer,\" arXiv, 2021.
[2] 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, 2017.
[3] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, Xuzheng He, Haowen Hou, Przemyslaw Kazienko, Jan Kocon, Jiaming Kong, Bartlomiej Koptyra, Hayden Lau, Krishna Sri Ipsit Mantri, Ferdinand Mom, Atsushi Saito, Xiangru Tang, Bolun Wang, Johan S. Wind, Stansilaw Wozniak, Ruichong Zhang, Zhenyuan Zhang, Qihang Zhao, Peng Zhou, Jian Zhu, and Rui-Jie Zhu, \"RWKV: Reinventing RNNs for the Transformer Era,\" arXiv, 2023.<\/p>\n","protected":false},"excerpt":{"rendered":"

\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 \u672c\u8a18\u4e8b\u306e\u69cb\u6210\u306f\u3001\u306f\u3058\u3081\u306bTransformer\u304a\u3088\u3073Transformer … <\/p>\n","protected":false},"author":1,"featured_media":8928,"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-8781","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\/08\/IMG_2694.png","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/8781","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=8781"}],"version-history":[{"count":46,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/8781\/revisions"}],"predecessor-version":[{"id":9391,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/posts\/8781\/revisions\/9391"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/media\/8928"}],"wp:attachment":[{"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/media?parent=8781"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/categories?post=8781"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/developers.agirobots.com\/jp\/wp-json\/wp\/v2\/tags?post=8781"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}