{"id":1071009,"date":"2023-02-15T01:02:00","date_gmt":"2023-02-15T09:02:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&p=1071009"},"modified":"2024-09-25T03:30:22","modified_gmt":"2024-09-25T10:30:22","slug":"aaai-2023-2","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/aaai-2023-2\/","title":{"rendered":"AAAI 2023 | \u8ba9\u4eba\u5de5\u667a\u80fd\u6280\u672f\u53d8\u5f97\u6709\u7406\u8bba\u610f\u4e49\u4e0e\u8d1f\u8d23\u4efb\u7684\u65b0\u601d\u8003"},"content":{"rendered":"\n

\u7f16\u8005\u6309\uff1a\u672c\u5e74\u5ea6\u7684 AAAI \u5927\u4f1a\u5df2\u7ecf\u62c9\u5f00\u5e37\u5e55\u3002\u4e0a\u671f\u6587\u7ae0\u4ecb\u7ecd\u4e86\u5728\u5de5\u4e1a\u5e94\u7528\u9886\u57df\u5185\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u7684\u6700\u65b0\u79d1\u7814\u6210\u679c \uff0c\u672c\u671f\u6587\u7ae0\u7ee7\u7eed\u653e\u9001\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5728\u5165\u9009 AAAI 2023 \u7684\u7814\u7a76\u5de5\u4f5c\u4e2d\uff0c\u5bf9\u4eba\u5de5\u667a\u80fd\u521b\u4f5c\u3001\u4eba\u5de5\u667a\u80fd\u7406\u8bba\u3001\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u76f8\u5173\u8bdd\u9898\u7684\u6700\u65b0\u601d\u8003\u3002<\/p>\n\n\n\n

\n\n\n\n

\u4eba\u5de5\u667a\u80fd\u521b\u4f5c<\/h2>\n\n\n\n

VideoDubber\uff1a\u8bed\u97f3\u65f6\u957f\u63a7\u5236\u7684\u673a\u5668\u7ffb\u8bd1\u89c6\u9891\u914d\u97f3\u6a21\u578b<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2211.16934<\/p>\n\n\n\n

Demo\u94fe\u63a5\uff1ahttps:\/\/speechresearch.github.io\/videodubbing<\/p>\n\n\n\n

\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/microsoft\/NeuralSpeech\/tree\/master\/VideoDubber<\/p>\n\n\n\n

\u89c6\u9891\u8bd1\u5236\uff08video dubbing\uff09\u4e00\u822c\u6307\u5c06\u89c6\u9891\u4e2d\u7684\u8bed\u97f3\u7531\u539f\u59cb\u8bed\u8a00\u7ffb\u8bd1\u4e3a\u76ee\u6807\u8bed\u8a00\uff0c\u5e76\u4fdd\u8bc1\u7ffb\u8bd1\u540e\u97f3\u753b\u540c\u6b65\u3002\u901a\u5e38\uff0c\u89c6\u9891\u8bd1\u5236\u7531\u591a\u4e2a\u7ea7\u8054\u7cfb\u7edf\u7ec4\u6210\uff0c\u5305\u62ec\u8bed\u97f3\u8bc6\u522b\u3001\u673a\u5668\u7ffb\u8bd1\u548c\u8bed\u97f3\u5408\u6210\u3002\u4ee5\u5f80\u7684\u5de5\u4f5c\u901a\u5e38\u53ea\u5728\u673a\u5668\u7ffb\u8bd1\u9636\u6bb5\u63a7\u5236\u7ffb\u8bd1\u540e\u7684\u5355\u8bcd\/\u5b57\u6bcd\u7684\u6570\u91cf\uff0c\u4e0d\u8003\u8651\u4e0d\u540c\u8bed\u8a00\u4e2d\u5355\u8bcd\/\u5b57\u7b26\u53d1\u97f3\u6301\u7eed\u65f6\u95f4\u7684\u76f8\u5f02\u6027\u3002\u4e3a\u4e86\u540c\u65f6\u4fdd\u8bc1\u7ffb\u8bd1\u7ed3\u679c\u7684\u81ea\u7136\u6027\u548c\u8bed\u97f3\u540c\u6b65\u6027\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u7684 VideoDubber \u5728\u673a\u5668\u7ffb\u8bd1\u9636\u6bb5\u5f15\u5165\u4e86\u8bed\u97f3\u65f6\u957f\u63a7\u5236\uff0c\u76f4\u63a5\u4ee5\u7ffb\u8bd1\u4e2d\u6bcf\u4e2a token \u7684\u8bed\u97f3\u65f6\u957f\u5339\u914d\u76ee\u6807\u8bed\u97f3\u7684\u957f\u5ea6\uff0c\u8fdb\u800c\u51cf\u5c11\u56e0\u8bed\u97f3\u5408\u6210\u9636\u6bb5\u65f6\u957f\u7684\u8fc7\u5206\u8c03\u6574\u5bf9\u7ffb\u8bd1\u81ea\u7136\u5ea6\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n

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

\"\u56fe1\uff1aVideoDubber
\u56fe1\uff1aVideoDubber \u7684\u6574\u4f53\u67b6\u6784<\/figcaption><\/figure>\n\n\n\n

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

VideoDubber \u8bbe\u8ba1\u4e86\u4e24\u79cd\u4e0e\u65f6\u957f\u76f8\u5173\u7684\u4f4d\u7f6e\u7f16\u7801\uff08\u7edd\u5bf9\u65f6\u957f\u7f16\u7801\u548c\u76f8\u5bf9\u65f6\u957f\u7f16\u7801\uff09\u5728\u673a\u5668\u7ffb\u8bd1\u8fc7\u7a0b\u4e2d\u96c6\u6210\u65f6\u957f\u4fe1\u606f\uff0c\u5f15\u5bfc\u5bf9\u6bcf\u4e2a\u5355\u8bcd\u97f3\u957f\u7684\u9884\u6d4b\uff0c\u4ece\u800c\u63a7\u5236\u751f\u6210\u53e5\u5b50\u7684\u8bed\u97f3\u957f\u5ea6\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cVideoDubber\u5728\u56db\u4e2a\u8bed\u8a00\u4e2d\u7684\u89c6\u9891\u8bd1\u5236\u540c\u6b65\u6027\u4f18\u4e8e\u57fa\u7ebf\u6a21\u578b\uff0c\u800c\u4e14\u5728\u4fdd\u8bc1\u7ffb\u8bd1\u8d28\u91cf\u7684\u540c\u65f6\uff0c\u8fd8\u5b9e\u73b0\u4e86\u66f4\u597d\u7684\u8bed\u97f3\u65f6\u957f\u63a7\u5236\u3002\u6b64\u5916\uff0c\u7531\u4e8e\u771f\u5b9e\u89c6\u9891\u8bd1\u5236\u6570\u636e\u96c6\u7684\u4e0d\u8db3\uff0c\u7814\u7a76\u56e2\u961f\u8fd8\u6784\u5efa\u4e86\u4e00\u4e2a\u4ece\u7535\u5f71\u4e2d\u6536\u96c6\u7684\u771f\u5b9e\u573a\u666f\u6d4b\u8bd5\u96c6\uff0c\u5bf9\u89c6\u9891\u8bd1\u5236\u4efb\u52a1\u8fdb\u884c\u4e86\u7efc\u5408\u8bc4\u4ef7\u3002\u5b9e\u9a8c\u8bc1\u660e\uff0c\u5728\u8003\u8651\u8bed\u97f3\u7b49\u65f6\u6027\u7684\u771f\u5b9e\u6d4b\u8bd5\u96c6\u4e2d\uff0cVideoDubber \u7684\u7b49\u65f6\u6027\u63a7\u5236\u80fd\u529b\u548c\u7ffb\u8bd1\u8d28\u91cf\u66f4\u4f18\u79c0\u3002\u6b64\u5916\uff0c\u4e3b\u89c2\u8bc4\u4ef7\u8868\u660e\uff0c VideoDubber \u81ea\u52a8\u8bd1\u5236\u89c6\u9891\u7684\u6574\u4f53\u8d28\u91cf\u6709\u663e\u8457\u63d0\u5347\u3002<\/p>\n\n\n\n

\u4eba\u5de5\u667a\u80fd\u7406\u8bba<\/h2>\n\n\n\n

\u57fa\u4e8e\u56e0\u679c\u7684\u7ec4\u5408\u5728\u7ebf\u5b66\u4e60<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2206.01995<\/p>\n\n\n\n

\u672c\u6587\u63a2\u8ba8\u4e86\u57fa\u4e8e\u56e0\u679c\u7684\u7ec4\u5408\u5728\u7ebf\u5b66\u4e60\u95ee\u9898\u3002\u5728\u8fd9\u4e2a\u95ee\u9898\u4e2d\uff0c\u73a9\u5bb6\u5728\u672a\u77e5\u7684\u56e0\u679c\u6a21\u578b\u4e2d\u4ea4\u4e92\uff0c\u5e72\u9884\u53d8\u91cf\u7684\u9009\u62e9\u5c06\u5f71\u54cd\u56e0\u679c\u6a21\u578b\u4e2d\u53ef\u89c1\u53d8\u91cf\u7684\u6700\u7ec8\u7ed3\u679c\u3002\u7b97\u6cd5\u901a\u8fc7\u73a9\u5bb6\u7684\u4ea4\u4e92\u5b66\u4e60\u56e0\u679c\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5e76\u66f4\u65b0\u5e72\u9884\u7b56\u7565\u5f97\u51fa\u591a\u8f6e\u4ea4\u4e92\u7684\u7d2f\u79ef\u6094\u503c\uff08\u5373\u9009\u62e9\u7684\u5e72\u9884\u65b9\u5f0f\u4e0e\u4f7f\u5f97\u76ee\u6807\u5956\u52b1\u53d8\u91cf\u6700\u5927\u5316\u7684\u6700\u4f18\u96c6\u5408\u5728\u7d2f\u8ba1\u5956\u52b1\u4e0a\u7684\u5dee\u8ddd\uff09\u7684\u6700\u5c0f\u503c\u3002<\/p>\n\n\n\n

\u6b64\u524d\u7684\u7406\u8bba\u7814\u7a76\u4e2d\u5bf9\u56e0\u679c\u5728\u7ebf\u5b66\u4e60\u4e2d\u7684\u7ec4\u5408\u4f18\u5316\u95ee\u9898\u7684\u63a2\u8ba8\u6709\u6240\u6b20\u7f3a\uff0c\u4ece\u800c\u5bfc\u81f4\u7406\u8bba\u6094\u503c\u4e2d\u51fa\u73b0\u5173\u4e8e\u56fe\u5927\u5c0f\u7684\u6307\u6570\u9879\u3002\u5728\u7ec4\u5408\u5728\u7ebf\u5b66\u4e60\u7684\u7814\u7a76\u4e2d\uff0c\u6bcf\u4e2a\u57fa\u7840\u9009\u62e9\u81c2\uff08base arm\uff09\u7684\u72ec\u7acb\u6027\u4e0e\u56e0\u679c\u5728\u7ebf\u5b66\u4e60\u90fd\u4e0d\u517c\u5bb9\u3002\u4e3a\u4e86\u89e3\u51b3\u4e0d\u517c\u5bb9\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u7279\u6b8a\u7684\u56e0\u679c\u6a21\u578b\u2014\u2014\u4e8c\u5143\u5e7f\u4e49\u7ebf\u6027\u6a21\u578b\uff08binary generalized linear model\uff09\uff0c\u4ec5\u6709\u6700\u591a n(n-1)\/2 \u4e2a\u53c2\u6570\uff08n \u4e3a\u53d8\u91cf\u4e2a\u6570\uff09\u3002\u5bf9\u4e8e\u4e00\u822c\u7684\u9a6c\u5c14\u79d1\u592b\u4e8c\u5143\u5e7f\u4e49\u7ebf\u6027\u6a21\u578b\uff0c\u672c\u6587\u8bbe\u8ba1\u4e86\u4e00\u79cd\u4f7f\u7528\u6781\u5927\u4f3c\u7136\u4f30\u8ba1\u7684\u5728\u7ebf\u7b97\u6cd5 BGLM-OFU\uff0c\u53ef\u4ee5\u5b9e\u73b0 O(n^{3\/2} T^{1\/2} ln T) \u7684\u671f\u671b\u7d2f\u79ef\u6094\u503c\u3002<\/p>\n\n\n\n

\u9488\u5bf9\u4e8c\u5143\u7ebf\u6027\u6a21\u578b\uff08binary linear model\uff09\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u89e3\u51b3\u65b9\u6848\uff1a\u5229\u7528 do-calculus \u7b97\u6cd5\uff0c\u5c06\u5305\u542b\u9690\u85cf\u7ed3\u70b9\u7684\u56e0\u679c\u6a21\u578b\u90e8\u5206\u8f6c\u5316\u4e3a\u9a6c\u5c14\u53ef\u592b\u56e0\u679c\u6a21\u578b\uff0c\u4ece\u800c\u90e8\u5206\u89e3\u51b3\u5305\u542b\u9690\u85cf\u8282\u70b9\u7684\u56e0\u679c\u6a21\u578b\u4e2d\u7684\u7ec4\u5408\u5728\u7ebf\u5b66\u4e60\u95ee\u9898\u3002\u6b64\u5916\uff0c\u65b0\u7684\u7b97\u6cd5 BLM-LR \u4e5f\u5728\u672c\u6587\u4e2d\u88ab\u63d0\u51fa\uff0c\u5b83\u4f7f\u7528\u7ebf\u6027\u56de\u5f52\u4ee3\u66ff\u6781\u5927\u4f3c\u7136\u4f30\u8ba1\uff0c\u8fdb\u4e00\u6b65\u6446\u8131\u5bf9\u53c2\u6570\u53d6\u503c\u7684\u5047\u8bbe\uff0c\u540c\u65f6\u53ef\u4ee5\u5b9e\u73b0 O(n^{5\/2} T^{1\/2} ln T) \u7684\u7d2f\u79ef\u671f\u671b\u6094\u503c\u3002<\/p>\n\n\n\n

\"\u56fe2\uff1a\u5728\u4e00\u4e2a\u5177\u67099\u4e2a\u8282\u70b9\u7684\u5e73\u884c\u56fe\u4e8c\u5143\u7ebf\u6027\u6a21\u578b\u4e0a\u8fdb\u884c\u6027\u80fd\u6a21\u62df\uff0cBGLM-OFU
\u56fe2\uff1a\u5728\u4e00\u4e2a\u5177\u67099\u4e2a\u8282\u70b9\u7684\u5e73\u884c\u56fe\u4e8c\u5143\u7ebf\u6027\u6a21\u578b\u4e0a\u8fdb\u884c\u6027\u80fd\u6a21\u62df\uff0cBGLM-OFU \u548c BLM-LR \u7b97\u6cd5\u7684\u8868\u73b0\u90fd\u4f18\u4e8e\u4f20\u7edf\u7684 UCB \u548c Epsilon-Greedy \u7b97\u6cd5<\/figcaption><\/figure>\n\n\n\n

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

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

\u4e3b\u52a8Token\u878d\u5408\u7b97\u5b50<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2203.06108<\/p>\n\n\n\n

\u5982\u4f55\u8fdb\u884c\u6709\u6548\u7684 token \u4fe1\u606f\u7684\u878d\u5408\uff08token-mixing\uff09\u662f\u5f53\u524d\u8ba1\u7b97\u673a\u89c6\u89c9\u57fa\u7840\u7f51\u7edc\u8bbe\u8ba1\u7684\u91cd\u70b9\u3002\u7efc\u5408\u5206\u6790\u73b0\u6709\u7684 CNN\u3001Transformer \u548c MLP \u7f51\u7edc\u4e2d\u4e0d\u540c token \u878d\u5408\u65b9\u5f0f\u7684\u4f18\u7f3a\u70b9\u540e\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u521b\u65b0\u6027\u5730\u63d0\u51fa\u4e86\u4e00\u79cd\u4e3b\u52a8\u4e14\u9ad8\u6548\u7684 token \u878d\u5408\u65b9\u5f0f\uff0c\u4f5c\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u4e2d\u901a\u7528\u7684\u57fa\u7840\u7b97\u5b50\u3002<\/p>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u5bf9\u4e8e\u7a7a\u95f4 token \u8fdb\u884c\u7ec6\u7c92\u5ea6\uff08channel-wise\uff09\u4e3b\u52a8\u878d\u5408\u7684\u7b97\u5b50\uff08Active Token Mixer\uff0c\u7b80\u79f0\u4e3a ATM\uff09\u3002ATM \u5c06\u7a7a\u95f4\u4fe1\u606f\u878d\u5408\u7684\u8fc7\u7a0b\u5206\u89e3\u4e3a\u4e09\u4e2a\u5e76\u884c\u7684\u5206\u652f\uff0c\u5373\u6c34\u5e73\u65b9\u5411\u7684\u878d\u5408\u3001\u7ad6\u76f4\u65b9\u5411\u7684\u878d\u5408\u548c\u539f\u7279\u5f81\u7684\u5904\u7406\u3002\u5728\u6c34\u5e73\u65b9\u5411\u548c\u7ad6\u76f4\u65b9\u5411\u4e0a\uff0cATM \u80fd\u591f\u6839\u636e\u4e0d\u540c\u4f4d\u7f6e\u7684\u8bed\u4e49\u4fe1\u606f\u7814\u5224\u5bf9\u5e94\u7a7a\u95f4\u8303\u56f4\u5185\u4e0d\u540c\u4f4d\u7f6e\u7684\u4fe1\u606f\u878d\u5408\u65b9\u5f0f\uff0c\u4ece\u800c\u5b9e\u73b0\u7a7a\u95f4 token \u7684\u81ea\u9002\u5e94\u878d\u5408\u3002\u57fa\u4e8e ATM\uff0c\u7814\u7a76\u5458\u4eec\u642d\u5efa\u4e86\u4e00\u79cd\u9ad8\u6548\u7684\u57fa\u7840\u89c6\u89c9\u7f51\u7edc ATMNet\u3002\u5927\u91cf\u7684\u5b9e\u9a8c\u8bc1\u660e\uff0c\u4e0d\u540c\u6a21\u578b\u5927\u5c0f\u7684 ATMNet \u5728\u5206\u7c7b\u4efb\u52a1\uff08ImageNet-1k\uff09\u3001\u7269\u4f53\u68c0\u6d4b\u4efb\u52a1\uff08COCO\uff09\u548c\u8bed\u4e49\u5206\u5272\u4efb\u52a1\uff08ADE20K\uff09\u4e0a\u5747\u80fd\u53d6\u5f97 SOTA \u6027\u80fd\u3002<\/p>\n\n\n\n

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

\"\u56fe3\uff1a\u5728\u6c34\u5e73\u65b9\u5411\u4e0a\u57fa\u4e8e\u4e3b\u52a8
\u56fe3\uff1a\u5728\u6c34\u5e73\u65b9\u5411\u4e0a\u57fa\u4e8e\u4e3b\u52a8 token \u878d\u5408\u7684\u8fc7\u7a0b\u793a\u610f\u56fe\uff08a\uff09\u548c\u5b8c\u6574\u7684 ATM \u6a21\u5757\u793a\u610f\u56fe\uff08b\uff09<\/figcaption><\/figure>\n\n\n\n

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

\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd<\/h2>\n\n\n\n

Prototypical Fine-tuning: \u81ea\u9002\u5e94\u6570\u636e\u96c6\u5927\u5c0f\u7684\u7a33\u5065\u5fae\u8c03\u65b9\u6cd5<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2211.13638<\/p>\n\n\n\n

\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u5404\u79cd NLP \u4efb\u52a1\u4e0a\u90fd\u5df2\u53d6\u5f97\u4e86\u5de8\u5927\u6210\u529f\u3002\u9884\u8bad\u7ec3\u6a21\u578b\u5f3a\u5927\u7684\u5224\u522b\u80fd\u529b\u53ef\u4ee5\u5f52\u56e0\u4e8e\u4e00\u4e2a\u8f83\u5f31\u7684\u5f52\u7eb3\u504f\u5dee\uff0c\u4f7f\u5f97\u6a21\u578b\u7684\u8868\u8fbe\u6027\u53d7\u8f83\u5c11\u7ea6\u675f\u3002\u4f46\u8fd9\u540c\u65f6\u4e5f\u9020\u6210\u4e86\u8fc7\u62df\u5408\u548c\u5c40\u90e8\u6700\u4f18\u7684\u95ee\u9898\uff0c\u5728\u8bad\u7ec3\u6837\u672c\u6570\u91cf\u8f83\u5c11\u65f6\u5c24\u4e3a\u660e\u663e\u3002\u8bf8\u5982 prototypical learning \u7b49\u975e\u53c2\u6570\u6a21\u578b\u8f83\u4e3a\u76f4\u63a5\u5efa\u6a21\u6570\u636e\u7684\u7c7b\u5185\u4e0e\u7c7b\u95f4\u5173\u7cfb\u5f15\u5165\u4e86\u8f83\u5f3a\u7684\u5f52\u7eb3\u504f\u5dee\u3002<\/p>\n\n\n\n

\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u7efc\u5408\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u6a21\u578b\u4e0e\u975e\u53c2\u6570\u6a21\u578b\u7684\u4f18\u70b9\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u81ea\u7531\u4e60\u5f97\u4e00\u4e2a\u5f52\u7eb3\u504f\u5dee\uff0c\u4ece\u800c\u9002\u5e94\u4e0d\u540c\u590d\u6742\u5ea6\u7684\u6570\u636e\u96c6\u3002\u672c\u6587\u63d0\u51fa\u4e86 Prototypical Fine-tuning (PFit)\u3002PFit \u5c06\u6bcf\u4e2a\u7c7b\u4e2d\u7684\u6570\u636e\u8868\u793a\u4e3a\u4e00\u7ec4 prototype\uff0c\u6bcf\u4e2a prototype \u5747\u5efa\u6a21\u4e3a\u6df7\u5408\u5206\u91cf\u3002\u540c\u65f6\uff0cPFit \u57fa\u4e8e\u6a21\u578b\u6027\u80fd\u548c\u6570\u636e\u5206\u5e03\u7684\u590d\u6742\u7a0b\u5ea6\u81ea\u52a8\u5b66\u4e60 prototype \u4e2a\u6570\u3002PFit \u786e\u4fdd\u4e86\u5047\u8bbe\u7a7a\u95f4\u8db3\u591f\u5145\u5206\uff0c\u4f7f\u5f97\u5728\u4fdd\u8bc1\u6a21\u578b\u6709\u8db3\u591f\u6cdb\u5316\u80fd\u529b\u7684\u540c\u65f6\uff0c\u66f4\u52a0\u8f7b\u6613\u5730\u5b66\u5230\u4efb\u52a1\u7684\u53ef\u884c\u89e3\u3002<\/p>\n\n\n\n

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

\"\u56fe4\uff1aPrototypical
\u56fe4\uff1aPrototypical Fine-tuning \u6a21\u578b\u56fe<\/figcaption><\/figure>\n\n\n\n

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

PFit \u7531\u56db\u90e8\u5206\u6784\u6210\uff1a<\/p>\n\n\n\n

\u25cf Mixture Prototype Initialization \uff1a\u5229\u7528\u9884\u8bad\u7ec3\u6a21\u578b\u8868\u5f81\u4e2d\u7684\u8bed\u4e49\u7279\u5f81\u6765\u521d\u59cb\u5316 prototype \u4ee5\u52a0\u5feb\u6536\u655b\uff1b<\/p>\n\n\n\n

\u25cf Infinite Mixture Prototype Creation\uff1a\u6269\u5c55 Infinite Mixture Prototype (IMP) \u4ee5\u7075\u6d3b\u5730\u6355\u6349\u6570\u636e\u5206\u5e03\uff0c\u5e76\u5728\u7528\u6570\u636e\u9a71\u52a8\u65b9\u5f0f\u7684\u540c\u65f6\u5fae\u8c03 prototypes \u4e0e\u9884\u8bad\u7ec3\u6a21\u578b\uff1b<\/p>\n\n\n\n

\u25cf Adaptive Prototype Simplification\uff1a\u4e3a\u63d0\u9ad8\u6cdb\u5316\u80fd\u529b\u548c\u6548\u7387\uff0c\u5229\u7528\u7b80\u5355\u7684\u5f52\u7eb3\u504f\u5dee\u5e76\u7ef4\u62a4\u4e00\u7ec4\u80fd\u591f\u5145\u5206\u8868\u793a\u6570\u636e\u5206\u5e03\u7684 prototypes\uff1b<\/p>\n\n\n\n

\u25cf Dynamic Diversity Regularization\uff1a\u52a8\u6001\u5730\u589e\u5f3a prototypes \u7684\u591a\u6837\u6027\uff0c\u4ee5\u63d0\u9ad8\u6a21\u578b\u8868\u8fbe\u529b\uff1b<\/p>\n\n\n\n

\u5728\u591a\u4e2a\u6570\u636e\u96c6\u4e0a\u7684\u5b9e\u9a8c\u8868\u660e\uff0cPFit \u5728\u4f4e\u8d44\u6e90\u573a\u666f\u4e0b\u53ef\u4ee5\u663e\u8457\u63d0\u5347\u6027\u80fd\uff0c\u4e14\u5728\u9ad8\u8d44\u6e90\u573a\u666f\u4e2d\u53ef\u4ee5\u8fbe\u5230\u4e0e\u539f\u6a21\u578b\u76f8\u5f53\u751a\u81f3\u66f4\u4f73\u7684\u6548\u679c\u3002<\/p>\n\n\n\n

\u57fa\u4e8e\u76f8\u4f3c\u5ea6\u5206\u5e03\u7684\u6210\u5458\u63a8\u65ad<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2211.15918<\/p>\n\n\n\n

\u672c\u6587\u63ed\u793a\u4e86\u4e00\u79cd\u56fe\u50cf\u68c0\u7d22\u548c\u8bc6\u522b\u7c7b\u6a21\u578b\u4e2d\u57fa\u4e8e\u6210\u5458\u63a8\u65ad\u7684\u65b0\u578b\u5b89\u5168\u98ce\u9669\u3002\u4ee5\u5f80\u7684\u6210\u5458\u63a8\u65ad\u653b\u51fb\u76f8\u5173\u7814\u7a76\u53ea\u5173\u6ce8\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\uff0c\u5ffd\u89c6\u4e86\u68c0\u7d22\u548c\u7ec6\u7c92\u5ea6\u8bc6\u522b\u4efb\u52a1\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u540c\u6837\u5b58\u5728\u88ab\u653b\u51fb\u7684\u98ce\u9669\u3002\u7136\u800c\uff0c\u7531\u4e8e\u56fe\u50cf\u68c0\u7d22\u548c\u7ec6\u7c92\u5ea6\u8bc6\u522b\u7b97\u6cd5\u4e0e\u56fe\u50cf\u5206\u7c7b\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u8303\u5f0f\u5dee\u5f02\u8f83\u5927\uff0c\u539f\u6709\u65b9\u6cd5\u5e76\u4e0d\u9002\u7528\u3002<\/p>\n\n\n\n

\u56e0\u6b64\uff0c\u4e3a\u4e86\u80fd\u591f\u66f4\u597d\u5730\u8bc4\u4f30\u56fe\u50cf\u68c0\u7d22\u548c\u7ec6\u7c92\u5ea6\u8bc6\u522b\u6a21\u578b\u5728\u6210\u5458\u63a8\u65ad\u653b\u51fb\u4e0a\u7684\u5b89\u5168\u98ce\u9669\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u65b9\u6cd5\u3002\u7531\u4e8e\u56fe\u50cf\u68c0\u7d22\u4e3b\u8981\u5173\u6ce8\u6837\u672c\u95f4\u7684\u5173\u8054\u5173\u7cfb\uff0c\u6240\u4ee5\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u6837\u672c\u95f4\u76f8\u4f3c\u5ea6\u7684\u5206\u5e03\u6765\u8fdb\u884c\u6210\u5458\u63a8\u65ad\u7684\u65b0\u653b\u51fb\u65b9\u6cd5\u3002\u5b9e\u9a8c\u8bc1\u660e\uff0c\u65b0\u65b9\u6cd5\u5bf9\u4e8e\u56fe\u50cf\u68c0\u7d22\u548c\u4ee5\u5f80\u7684\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u90fd\u80fd\u5f62\u6210\u6709\u6548\u7684\u653b\u51fb\u5a01\u80c1\u3002<\/p>\n\n\n\n

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

\"\u56fe5\uff1a\u57fa\u4e8e\u6837\u672c\u95f4\u76f8\u4f3c\u5ea6\u7684\u5206\u5e03\u6765\u8fdb\u884c\u6210\u5458\u63a8\u65ad\u7684\u65b0\u653b\u51fb\u65b9\u6cd5\"
\u56fe5\uff1a\u57fa\u4e8e\u6837\u672c\u95f4\u76f8\u4f3c\u5ea6\u7684\u5206\u5e03\u6765\u8fdb\u884c\u6210\u5458\u63a8\u65ad\u7684\u65b0\u653b\u51fb\u65b9\u6cd5<\/figcaption><\/figure>\n\n\n\n

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

\u7f16\u8005\u6309\uff1a\u672c\u5e74\u5ea6\u7684 AAAI \u5927\u4f1a\u5df2\u7ecf\u62c9\u5f00\u5e37\u5e55\u3002\u4e0a\u671f\u6587\u7ae0\u4ecb\u7ecd\u4e86\u5728\u5de5\u4e1a\u5e94\u7528\u9886\u57df\u5185\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u7684\u6700\u65b0\u79d1\u7814\u6210\u679c \uff0c\u672c\u671f\u6587\u7ae0\u7ee7\u7eed\u653e\u9001\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5728\u5165\u9009 AAAI 2023 \u7684\u7814\u7a76\u5de5\u4f5c\u4e2d\uff0c\u5bf9\u4eba\u5de5\u667a\u80fd\u521b\u4f5c\u3001\u4eba\u5de5\u667a\u80fd\u7406\u8bba\u3001\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u76f8\u5173\u8bdd\u9898\u7684\u6700\u65b0\u601d\u8003\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2211.16934 Demo\u94fe\u63a5\uff1ahttps:\/\/speechresearch.github.io\/videodubbing \u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/microsoft\/NeuralSpeech\/tree\/master\/VideoDubber \u89c6\u9891\u8bd1\u5236\uff08video dubbing\uff09\u4e00\u822c\u6307\u5c06\u89c6\u9891\u4e2d\u7684\u8bed\u97f3\u7531\u539f\u59cb\u8bed\u8a00\u7ffb\u8bd1\u4e3a\u76ee\u6807\u8bed\u8a00\uff0c\u5e76\u4fdd\u8bc1\u7ffb\u8bd1\u540e\u97f3\u753b\u540c\u6b65\u3002\u901a\u5e38\uff0c\u89c6\u9891\u8bd1\u5236\u7531\u591a\u4e2a\u7ea7\u8054\u7cfb\u7edf\u7ec4\u6210\uff0c\u5305\u62ec\u8bed\u97f3\u8bc6\u522b\u3001\u673a\u5668\u7ffb\u8bd1\u548c\u8bed\u97f3\u5408\u6210\u3002\u4ee5\u5f80\u7684\u5de5\u4f5c\u901a\u5e38\u53ea\u5728\u673a\u5668\u7ffb\u8bd1\u9636\u6bb5\u63a7\u5236\u7ffb\u8bd1\u540e\u7684\u5355\u8bcd\/\u5b57\u6bcd\u7684\u6570\u91cf\uff0c\u4e0d\u8003\u8651\u4e0d\u540c\u8bed\u8a00\u4e2d\u5355\u8bcd\/\u5b57\u7b26\u53d1\u97f3\u6301\u7eed\u65f6\u95f4\u7684\u76f8\u5f02\u6027\u3002\u4e3a\u4e86\u540c\u65f6\u4fdd\u8bc1\u7ffb\u8bd1\u7ed3\u679c\u7684\u81ea\u7136\u6027\u548c\u8bed\u97f3\u540c\u6b65\u6027\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u7684 VideoDubber \u5728\u673a\u5668\u7ffb\u8bd1\u9636\u6bb5\u5f15\u5165\u4e86\u8bed\u97f3\u65f6\u957f\u63a7\u5236\uff0c\u76f4\u63a5\u4ee5\u7ffb\u8bd1\u4e2d\u6bcf\u4e2a token \u7684\u8bed\u97f3\u65f6\u957f\u5339\u914d\u76ee\u6807\u8bed\u97f3\u7684\u957f\u5ea6\uff0c\u8fdb\u800c\u51cf\u5c11\u56e0\u8bed\u97f3\u5408\u6210\u9636\u6bb5\u65f6\u957f\u7684\u8fc7\u5206\u8c03\u6574\u5bf9\u7ffb\u8bd1\u81ea\u7136\u5ea6\u7684\u5f71\u54cd\u3002 VideoDubber \u8bbe\u8ba1\u4e86\u4e24\u79cd\u4e0e\u65f6\u957f\u76f8\u5173\u7684\u4f4d\u7f6e\u7f16\u7801\uff08\u7edd\u5bf9\u65f6\u957f\u7f16\u7801\u548c\u76f8\u5bf9\u65f6\u957f\u7f16\u7801\uff09\u5728\u673a\u5668\u7ffb\u8bd1\u8fc7\u7a0b\u4e2d\u96c6\u6210\u65f6\u957f\u4fe1\u606f\uff0c\u5f15\u5bfc\u5bf9\u6bcf\u4e2a\u5355\u8bcd\u97f3\u957f\u7684\u9884\u6d4b\uff0c\u4ece\u800c\u63a7\u5236\u751f\u6210\u53e5\u5b50\u7684\u8bed\u97f3\u957f\u5ea6\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cVideoDubber\u5728\u56db\u4e2a\u8bed\u8a00\u4e2d\u7684\u89c6\u9891\u8bd1\u5236\u540c\u6b65\u6027\u4f18\u4e8e\u57fa\u7ebf\u6a21\u578b\uff0c\u800c\u4e14\u5728\u4fdd\u8bc1\u7ffb\u8bd1\u8d28\u91cf\u7684\u540c\u65f6\uff0c\u8fd8\u5b9e\u73b0\u4e86\u66f4\u597d\u7684\u8bed\u97f3\u65f6\u957f\u63a7\u5236\u3002\u6b64\u5916\uff0c\u7531\u4e8e\u771f\u5b9e\u89c6\u9891\u8bd1\u5236\u6570\u636e\u96c6\u7684\u4e0d\u8db3\uff0c\u7814\u7a76\u56e2\u961f\u8fd8\u6784\u5efa\u4e86\u4e00\u4e2a\u4ece\u7535\u5f71\u4e2d\u6536\u96c6\u7684\u771f\u5b9e\u573a\u666f\u6d4b\u8bd5\u96c6\uff0c\u5bf9\u89c6\u9891\u8bd1\u5236\u4efb\u52a1\u8fdb\u884c\u4e86\u7efc\u5408\u8bc4\u4ef7\u3002\u5b9e\u9a8c\u8bc1\u660e\uff0c\u5728\u8003\u8651\u8bed\u97f3\u7b49\u65f6\u6027\u7684\u771f\u5b9e\u6d4b\u8bd5\u96c6\u4e2d\uff0cVideoDubber \u7684\u7b49\u65f6\u6027\u63a7\u5236\u80fd\u529b\u548c\u7ffb\u8bd1\u8d28\u91cf\u66f4\u4f18\u79c0\u3002\u6b64\u5916\uff0c\u4e3b\u89c2\u8bc4\u4ef7\u8868\u660e\uff0c VideoDubber \u81ea\u52a8\u8bd1\u5236\u89c6\u9891\u7684\u6574\u4f53\u8d28\u91cf\u6709\u663e\u8457\u63d0\u5347\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2206.01995 \u672c\u6587\u63a2\u8ba8\u4e86\u57fa\u4e8e\u56e0\u679c\u7684\u7ec4\u5408\u5728\u7ebf\u5b66\u4e60\u95ee\u9898\u3002\u5728\u8fd9\u4e2a\u95ee\u9898\u4e2d\uff0c\u73a9\u5bb6\u5728\u672a\u77e5\u7684\u56e0\u679c\u6a21\u578b\u4e2d\u4ea4\u4e92\uff0c\u5e72\u9884\u53d8\u91cf\u7684\u9009\u62e9\u5c06\u5f71\u54cd\u56e0\u679c\u6a21\u578b\u4e2d\u53ef\u89c1\u53d8\u91cf\u7684\u6700\u7ec8\u7ed3\u679c\u3002\u7b97\u6cd5\u901a\u8fc7\u73a9\u5bb6\u7684\u4ea4\u4e92\u5b66\u4e60\u56e0\u679c\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5e76\u66f4\u65b0\u5e72\u9884\u7b56\u7565\u5f97\u51fa\u591a\u8f6e\u4ea4\u4e92\u7684\u7d2f\u79ef\u6094\u503c\uff08\u5373\u9009\u62e9\u7684\u5e72\u9884\u65b9\u5f0f\u4e0e\u4f7f\u5f97\u76ee\u6807\u5956\u52b1\u53d8\u91cf\u6700\u5927\u5316\u7684\u6700\u4f18\u96c6\u5408\u5728\u7d2f\u8ba1\u5956\u52b1\u4e0a\u7684\u5dee\u8ddd\uff09\u7684\u6700\u5c0f\u503c\u3002 \u6b64\u524d\u7684\u7406\u8bba\u7814\u7a76\u4e2d\u5bf9\u56e0\u679c\u5728\u7ebf\u5b66\u4e60\u4e2d\u7684\u7ec4\u5408\u4f18\u5316\u95ee\u9898\u7684\u63a2\u8ba8\u6709\u6240\u6b20\u7f3a\uff0c\u4ece\u800c\u5bfc\u81f4\u7406\u8bba\u6094\u503c\u4e2d\u51fa\u73b0\u5173\u4e8e\u56fe\u5927\u5c0f\u7684\u6307\u6570\u9879\u3002\u5728\u7ec4\u5408\u5728\u7ebf\u5b66\u4e60\u7684\u7814\u7a76\u4e2d\uff0c\u6bcf\u4e2a\u57fa\u7840\u9009\u62e9\u81c2\uff08base arm\uff09\u7684\u72ec\u7acb\u6027\u4e0e\u56e0\u679c\u5728\u7ebf\u5b66\u4e60\u90fd\u4e0d\u517c\u5bb9\u3002\u4e3a\u4e86\u89e3\u51b3\u4e0d\u517c\u5bb9\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u7279\u6b8a\u7684\u56e0\u679c\u6a21\u578b\u2014\u2014\u4e8c\u5143\u5e7f\u4e49\u7ebf\u6027\u6a21\u578b\uff08binary generalized linear model\uff09\uff0c\u4ec5\u6709\u6700\u591a n(n-1)\/2 \u4e2a\u53c2\u6570\uff08n \u4e3a\u53d8\u91cf\u4e2a\u6570\uff09\u3002\u5bf9\u4e8e\u4e00\u822c\u7684\u9a6c\u5c14\u79d1\u592b\u4e8c\u5143\u5e7f\u4e49\u7ebf\u6027\u6a21\u578b\uff0c\u672c\u6587\u8bbe\u8ba1\u4e86\u4e00\u79cd\u4f7f\u7528\u6781\u5927\u4f3c\u7136\u4f30\u8ba1\u7684\u5728\u7ebf\u7b97\u6cd5 BGLM-OFU\uff0c\u53ef\u4ee5\u5b9e\u73b0 O(n^{3\/2} T^{1\/2} ln T) \u7684\u671f\u671b\u7d2f\u79ef\u6094\u503c\u3002 \u9488\u5bf9\u4e8c\u5143\u7ebf\u6027\u6a21\u578b\uff08binary linear model\uff09\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u89e3\u51b3\u65b9\u6848\uff1a\u5229\u7528 do-calculus \u7b97\u6cd5\uff0c\u5c06\u5305\u542b\u9690\u85cf\u7ed3\u70b9\u7684\u56e0\u679c\u6a21\u578b\u90e8\u5206\u8f6c\u5316\u4e3a\u9a6c\u5c14\u53ef\u592b\u56e0\u679c\u6a21\u578b\uff0c\u4ece\u800c\u90e8\u5206\u89e3\u51b3\u5305\u542b\u9690\u85cf\u8282\u70b9\u7684\u56e0\u679c\u6a21\u578b\u4e2d\u7684\u7ec4\u5408\u5728\u7ebf\u5b66\u4e60\u95ee\u9898\u3002\u6b64\u5916\uff0c\u65b0\u7684\u7b97\u6cd5 BLM-LR \u4e5f\u5728\u672c\u6587\u4e2d\u88ab\u63d0\u51fa\uff0c\u5b83\u4f7f\u7528\u7ebf\u6027\u56de\u5f52\u4ee3\u66ff\u6781\u5927\u4f3c\u7136\u4f30\u8ba1\uff0c\u8fdb\u4e00\u6b65\u6446\u8131\u5bf9\u53c2\u6570\u53d6\u503c\u7684\u5047\u8bbe\uff0c\u540c\u65f6\u53ef\u4ee5\u5b9e\u73b0 O(n^{5\/2} T^{1\/2} ln T) \u7684\u7d2f\u79ef\u671f\u671b\u6094\u503c\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2203.06108 \u5982\u4f55\u8fdb\u884c\u6709\u6548\u7684 token \u4fe1\u606f\u7684\u878d\u5408\uff08token-mixing\uff09\u662f\u5f53\u524d\u8ba1\u7b97\u673a\u89c6\u89c9\u57fa\u7840\u7f51\u7edc\u8bbe\u8ba1\u7684\u91cd\u70b9\u3002\u7efc\u5408\u5206\u6790\u73b0\u6709\u7684 CNN\u3001Transformer \u548c […]<\/p>\n","protected":false},"author":34512,"featured_media":1071159,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1071009","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1071009","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":2,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1071009\/revisions"}],"predecessor-version":[{"id":1071195,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1071009\/revisions\/1071195"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media\/1071159"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1071009"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1071009"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1071009"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1071009"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}