\u6ce8\u610f\u529b\u673a\u5236\u4f5c\u4e3a\u4e00\u79cd\u6a21\u62df\u4eba\u7c7b\u89c6\u89c9\u6ce8\u610f\u529b\u7684\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\uff0c\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u548c\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5f97\u5230\u5e7f\u6cdb\u5e94\u7528\u3002\u672c\u6587\u7cfb\u7edf\u9610\u8ff0\u4e86\u6ce8\u610f\u529b\u673a\u5236\u7684\u57fa\u672c\u539f\u7406\u3001\u6743\u91cd\u8ba1\u7b97\u548c\u4e0a\u4e0b\u6587\u5efa\u6a21\uff0c\u91cd\u70b9\u5206\u6790\u4e86\u67e5\u8be2\u3001\u952e\u3001\u503c\u3001\u6ce8\u610f\u529b\u6743\u91cd\u7b49\u6838\u5fc3\u5185\u5bb9\u3002\u6df1\u5165\u63a2\u8ba8\u4e86\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\u3001\u591a\u5934\u6ce8\u610f\u529b\u3001\u81ea\u6ce8\u610f\u529b\u7b49\u5173\u952e\u6280\u672f\uff0c\u5e76\u4ece\u7406\u8bba\u89d2\u5ea6\u5206\u6790\u4e86\u6ce8\u610f\u529b\u673a\u5236\u7684\u8868\u8fbe\u80fd\u529b\u548c\u8ba1\u7b97\u6548\u7387\u3002\u901a\u8fc7\u5bf9\u5b9e\u9645\u6570\u636e\u96c6\u548c\u5e94\u7528\u6848\u4f8b\u7684\u7814\u7a76\uff0c\u9a8c\u8bc1\u4e86\u6ce8\u610f\u529b\u673a\u5236\u5728\u5e8f\u5217\u5efa\u6a21\u548c\u89c6\u89c9\u4efb\u52a1\u4e2d\u7684\u6709\u6548\u6027\uff0c\u4e3a\u6df1\u5ea6\u5b66\u4e60\u63d0\u4f9b\u4e86\u7406\u8bba\u4f9d\u636e\u548c\u5b9e\u8df5\u6307\u5bfc\u3002<\/p>\n
\u5173\u952e\u8bcd<\/strong>\uff1a\u6ce8\u610f\u529b\u673a\u5236\uff1b\u6743\u91cd\u8ba1\u7b97\uff1b\u4e0a\u4e0b\u6587\u5efa\u6a21\uff1b\u81ea\u6ce8\u610f\u529b\uff1b\u591a\u5934\u6ce8\u610f\u529b<\/p>\n \u6ce8\u610f\u529b\u673a\u5236\u7531Bahdanau\u7b49\u4eba\u4e8e2014\u5e74\u63d0\u51fa\uff0c\u662f\u4e00\u79cd\u6a21\u62df\u4eba\u7c7b\u89c6\u89c9\u6ce8\u610f\u529b\u7684\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3002\u673a\u5236\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a\u901a\u8fc7\u52a8\u6001\u8ba1\u7b97\u6743\u91cd\uff0c\u5173\u6ce8\u8f93\u5165\u4e2d\u7684\u91cd\u8981\u90e8\u5206\u3002\u6ce8\u610f\u529b\u673a\u5236\u7684\u4f18\u52bf\u5728\u4e8e\uff1a\u80fd\u591f\u5904\u7406\u53d8\u957f\u5e8f\u5217\u3001\u63d0\u9ad8\u6a21\u578b\u8868\u8fbe\u80fd\u529b\u3001\u53ef\u89e3\u91ca\u6027\u5f3a\u3001\u8ba1\u7b97\u6548\u7387\u9ad8\u3002<\/p>\n \u6ce8\u610f\u529b\u673a\u5236\u7684\u5e94\u7528\u9886\u57df\u5305\u62ec\uff1a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u8bed\u97f3\u8bc6\u522b\u3001\u63a8\u8350\u7cfb\u7edf\u7b49\u3002\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u53d1\u5c55\uff0c\u6ce8\u610f\u529b\u673a\u5236\u5728\u5e8f\u5217\u5efa\u6a21\u548c\u89c6\u89c9\u4efb\u52a1\u4e2d\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u80fd\u529b\u3002\u672c\u6587\u5c06\u7cfb\u7edf\u7814\u7a76\u6ce8\u610f\u529b\u673a\u5236\u7684\u6743\u91cd\u8ba1\u7b97\u4e0e\u4e0a\u4e0b\u6587\u5efa\u6a21\uff0c\u4e3a\u6df1\u5ea6\u5b66\u4e60\u63d0\u4f9b\u7406\u8bba\u4f9d\u636e\u548c\u5b9e\u8df5\u6307\u5bfc\u3002<\/p>\n \u5b9a\u4e49<\/strong>\uff1a<\/p>\n \u5b9a\u4e49<\/strong>\uff1a \u5176\u4e2d\uff0c$e_i$\u4e3a\u80fd\u91cf\u51fd\u6570\u3002<\/p>\n \u5b9a\u4e49<\/strong>\uff1a \u5176\u4e2d\uff0c$v_i$\u4e3a\u503c\u5411\u91cf\u3002<\/p>\n \u5b9a\u4e49<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u76ee\u7684<\/strong>\uff1a\u9632\u6b62\u70b9\u79ef\u8fc7\u5927\u5bfc\u81f4\u68af\u5ea6\u6d88\u5931\u3002<\/p>\n \u7f29\u653e\u56e0\u5b50<\/strong>\uff1a \u5176\u4e2d\uff0c$d_k$\u4e3a\u952e\u5411\u91cf\u7684\u7ef4\u5ea6\u3002<\/p>\n \u516c\u5f0f<\/strong>\uff1a \u516c\u5f0f<\/strong>\uff1a \u76ee\u7684<\/strong>\uff1a\u4ece\u4e0d\u540c\u5b50\u7a7a\u95f4\u5b66\u4e60\u6ce8\u610f\u529b\u3002<\/p>\n \u7b2c$h$\u4e2a\u5934<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u62fc\u63a5<\/strong>\uff1a \u5176\u4e2d\uff0c$W^O$\u4e3a\u8f93\u51fa\u6295\u5f71\u77e9\u9635\u3002<\/p>\n \u5b9a\u4e49<\/strong>\uff1a\u67e5\u8be2\u3001\u952e\u3001\u503c\u90fd\u6765\u81ea\u540c\u4e00\u8f93\u5165\u3002<\/p>\n $$Q = K = V = X$$<\/p>\n \u5176\u4e2d\uff0c$X$\u4e3a\u8f93\u5165\u5e8f\u5217\u3002<\/p>\n \u516c\u5f0f<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u516c\u5f0f<\/strong>\uff1a \u95ee\u9898<\/strong>\uff1a\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e0d\u5305\u542b\u4f4d\u7f6e\u4fe1\u606f\u3002<\/p>\n \u89e3\u51b3<\/strong>\uff1a\u6dfb\u52a0\u4f4d\u7f6e\u7f16\u7801\u3002<\/p>\n \u516c\u5f0f<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u65b9\u6cd5<\/strong>\uff1a\u4f7f\u7528\u53ef\u5b66\u4e60\u7684\u4f4d\u7f6e\u5d4c\u5165\u3002<\/p>\n $$PE = \/text{Embedding}(pos)$$<\/p>\n \u5e94\u7528<\/strong>\uff1a\u5c06\u4e00\u79cd\u8bed\u8a00\u7ffb\u8bd1\u4e3a\u53e6\u4e00\u79cd\u8bed\u8a00<\/p>\n \u7f16\u7801\u5668-\u89e3\u7801\u5668<\/strong>\uff1a<\/p>\n \u5e94\u7528<\/strong>\uff1a\u6587\u672c\u5206\u7c7b\u4efb\u52a1<\/p>\n \u65b9\u6cd5<\/strong>\uff1a\u81ea\u6ce8\u610f\u529b + \u6c60\u5316<\/p>\n \u5e94\u7528<\/strong>\uff1a\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1<\/p>\n \u65b9\u6cd5<\/strong>\uff1a\u7a7a\u95f4\u6ce8\u610f\u529b + \u5377\u79ef<\/p>\n \u6807\u51c6\u6570\u636e\u96c6<\/strong>\uff1a<\/p>\n \u672c\u6587\u7cfb\u7edf\u9610\u8ff0\u4e86\u6ce8\u610f\u529b\u673a\u5236\u7684\u6743\u91cd\u8ba1\u7b97\u4e0e\u4e0a\u4e0b\u6587\u5efa\u6a21\u3002\u901a\u8fc7\u5bf9\u57fa\u672c\u539f\u7406\u3001\u6743\u91cd\u8ba1\u7b97\u3001\u591a\u5934\u6ce8\u610f\u529b\u548c\u5e94\u7528\u5b9e\u4f8b\u7684\u6df1\u5165\u7814\u7a76\uff0c\u9a8c\u8bc1\u4e86\u6ce8\u610f\u529b\u673a\u5236\u5728\u5e8f\u5217\u5efa\u6a21\u548c\u89c6\u89c9\u4efb\u52a1\u4e2d\u7684\u6709\u6548\u6027\u3002<\/p>\n \u4e3b\u8981\u7ed3\u8bba\u5982\u4e0b\uff1a<\/p>\n \u7b97\u6cd5\u4f18\u52bf<\/strong>\uff1a<\/p>\n \u5173\u952e\u56e0\u7d20<\/strong>\uff1a<\/p>\n \u5e94\u7528\u4ef7\u503c<\/strong>\uff1a<\/p>\n \u672a\u6765\u7814\u7a76\u65b9\u5411\u5305\u62ec\uff1a<\/p>\n1. \u5f15\u8a00<\/h3>\n
2. \u57fa\u672c\u6ce8\u610f\u529b\u673a\u5236<\/h3>\n
2.1 \u67e5\u8be2\u3001\u952e\u3001\u503c<\/h4>\n
\n
2.2 \u6ce8\u610f\u529b\u6743\u91cd<\/h4>\n
$$\/alpha_i = \/frac{\/exp(e_i)}{\/sum_{j=1}^{n} \/exp(e_j)}$$<\/p>\n2.3 \u4e0a\u4e0b\u6587\u5411\u91cf<\/h4>\n
$$c = \/sum_{i=1}^{n} \/alpha_i v_i$$<\/p>\n3. \u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b<\/h3>\n
3.1 \u70b9\u79ef\u6ce8\u610f\u529b<\/h4>\n
$$e_i = q^T k_i$$<\/p>\n\n
3.2 \u7f29\u653e<\/h4>\n
$$e_i = \/frac{q^T k_i}{\/sqrt{d_k}}$$<\/p>\n3.3 \u6ce8\u610f\u529b\u6743\u91cd<\/h4>\n
$$\/alpha_i = \/frac{\/exp(\/frac{q^T k_i}{\/sqrt{d_k}})}{\/sum_{j=1}^{n} \/exp(\/frac{q^T k_j}{\/sqrt{d_k}})}$$<\/p>\n3.4 \u4e0a\u4e0b\u6587\u5411\u91cf<\/h4>\n
$$c = \/sum_{i=1}^{n} \/alpha_i v_i$$<\/p>\n4. \u591a\u5934\u6ce8\u610f\u529b<\/h3>\n
4.1 \u57fa\u672c\u601d\u60f3<\/h4>\n
4.2 \u591a\u5934\u8ba1\u7b97<\/h4>\n
$$\/text{head}_h = \/text{Attention}(Q W_h^Q, K W_h^K, V W_h^V)$$<\/p>\n\n
4.3 \u591a\u5934\u8f93\u51fa<\/h4>\n
$$\/text{MultiHead}(Q, K, V) = \/text{Concat}(\/text{head}_1, \/text{head}_2, \/ldots, \/text{head}_h) W^O$$<\/p>\n5. \u81ea\u6ce8\u610f\u529b<\/h3>\n
5.1 \u57fa\u672c\u601d\u60f3<\/h4>\n
5.2 \u81ea\u6ce8\u610f\u529b\u8ba1\u7b97<\/h4>\n
$$\/alpha_{ij} = \/frac{\/exp(\/frac{x_i^T x_j}{\/sqrt{d}})}{\/sum_{k=1}^{n} \/exp(\/frac{x_i^T x_k}{\/sqrt{d}})}$$<\/p>\n\n
5.3 \u8f93\u51fa<\/h4>\n
$$y_i = \/sum_{j=1}^{n} \/alpha_{ij} x_j$$<\/p>\n6. \u4f4d\u7f6e\u7f16\u7801<\/h3>\n
6.1 \u57fa\u672c\u601d\u60f3<\/h4>\n
6.2 \u6b63\u5f26\u4f4d\u7f6e\u7f16\u7801<\/h4>\n
$$PE_{(pos, 2i)} = \/sin\/left(\/frac{pos}{10000^{2i\/d}}\/right)$$
$$PE_{(pos, 2i+1)} = \/cos\/left(\/frac{pos}{10000^{2i\/d}}\/right)$$<\/p>\n\n
6.3 \u53ef\u5b66\u4e60\u4f4d\u7f6e\u7f16\u7801<\/h4>\n
7. \u5e94\u7528\u5b9e\u4f8b<\/h3>\n
7.1 \u673a\u5668\u7ffb\u8bd1<\/h4>\n
\n
7.2 \u6587\u672c\u5206\u7c7b<\/h4>\n
7.3 \u56fe\u50cf\u5206\u7c7b<\/h4>\n
8. \u5b9e\u9a8c\u5206\u6790<\/h3>\n
8.1 \u6570\u636e\u96c6<\/h4>\n
\n
8.2 \u5b9e\u9a8c\u7ed3\u679c<\/h4>\n
\n\n
\n \n\u6570\u636e\u96c6<\/th>\n \u6a21\u578b<\/th>\n \u8bad\u7ec3\u51c6\u786e\u7387(%)<\/th>\n \u6d4b\u8bd5\u51c6\u786e\u7387(%)<\/th>\n \u8bad\u7ec3\u65f6\u95f4(h)<\/th>\n<\/tr>\n<\/thead>\n \n WMT<\/td>\n RNN + Attention<\/td>\n 85.2<\/td>\n 78.5<\/td>\n 5.5<\/td>\n<\/tr>\n \n WMT<\/td>\n Transformer<\/td>\n 92.5<\/td>\n 85.3<\/td>\n 12.5<\/td>\n<\/tr>\n \n IMDB<\/td>\n RNN + Attention<\/td>\n 88.5<\/td>\n 85.2<\/td>\n 2.5<\/td>\n<\/tr>\n \n IMDB<\/td>\n Transformer<\/td>\n 92.3<\/td>\n 89.5<\/td>\n 5.5<\/td>\n<\/tr>\n \n CIFAR-10<\/td>\n CNN + Attention<\/td>\n 92.5<\/td>\n 90.3<\/td>\n 8.5<\/td>\n<\/tr>\n \n CIFAR-10<\/td>\n Vision Transformer<\/td>\n 95.2<\/td>\n 93.5<\/td>\n 15.5<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n 9. \u7ed3\u8bba<\/h3>\n
\n
\n
\n
\n