{"id":1069890,"date":"2022-09-26T22:54:00","date_gmt":"2022-09-27T05:54:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&p=1069890"},"modified":"2024-09-25T01:43:02","modified_gmt":"2024-09-25T08:43:02","slug":"usb","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/usb\/","title":{"rendered":"USB\uff1a\u9996\u4e2a\u5c06\u89c6\u89c9\u3001\u8bed\u8a00\u548c\u97f3\u9891\u5206\u7c7b\u4efb\u52a1\u8fdb\u884c\u7edf\u4e00\u7684\u534a\u76d1\u7763\u5206\u7c7b\u5b66\u4e60\u57fa\u51c6"},"content":{"rendered":"\n

\u4f5c\u8005\uff1a\u738b\u4e00\u680b\u3001\u738b\u664b\u4e1c<\/em><\/p>\n\n\n\n

\u7f16\u8005\u6309\uff1a\u5f53\u524d\uff0c\u534a\u76d1\u7763\u5b66\u4e60\u7684\u53d1\u5c55\u5982\u706b\u5982\u837c\u3002\u4f46\u662f\u73b0\u6709\u7684\u534a\u76d1\u7763\u5b66\u4e60\u57fa\u51c6\u5927\u591a\u5c40\u9650\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u5206\u7c7b\u4efb\u52a1\uff0c\u6392\u9664\u4e86\u5bf9\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u97f3\u9891\u5904\u7406\u7b49\u5206\u7c7b\u4efb\u52a1\u7684\u4e00\u81f4\u548c\u591a\u6837\u5316\u8bc4\u4f30\u3002\u6b64\u5916\uff0c\u5927\u90e8\u5206\u534a\u76d1\u7763\u8bba\u6587\u7531\u5927\u578b\u673a\u6784\u53d1\u8868\uff0c\u5b66\u672f\u754c\u7684\u5b9e\u9a8c\u5ba4\u5f80\u5f80\u7531\u4e8e\u8ba1\u7b97\u8d44\u6e90\u7684\u9650\u5236\u800c\u5f88\u96be\u53c2\u4e0e\u5230\u63a8\u52a8\u8be5\u9886\u57df\u7684\u53d1\u5c55\u4e2d\u3002\u4e3a\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u8054\u5408\u897f\u6e56\u5927\u5b66\u3001\u4e1c\u4eac\u5de5\u4e1a\u5927\u5b66\u3001\u5361\u5185\u57fa\u6885\u9686\u5927\u5b66\u3001\u9a6c\u514b\u65af-\u666e\u6717\u514b\u7814\u7a76\u6240\u7b49\u673a\u6784\u7684\u79d1\u7814\u4eba\u5458\u63d0\u51fa\u4e86 Unified SSL Benchmark\uff08USB\uff09\uff1a\u7b2c\u4e00\u4e2a\u5c06\u89c6\u89c9\u3001\u8bed\u8a00\u548c\u97f3\u9891\u5206\u7c7b\u4efb\u52a1\u8fdb\u884c\u7edf\u4e00\u7684\u534a\u76d1\u7763\u5206\u7c7b\u5b66\u4e60\u57fa\u51c6\u3002\u8be5\u8bba\u6587\u4e0d\u4ec5\u5f15\u5165\u4e86\u66f4\u591a\u6837\u5316\u7684\u5e94\u7528\u9886\u57df\uff0c\u8fd8\u9996\u6b21\u5229\u7528\u89c6\u89c9\u9884\u8bad\u7ec3\u6a21\u578b\u5927\u5927\u7f29\u51cf\u4e86\u534a\u76d1\u7763\u7b97\u6cd5\u7684\u9a8c\u8bc1\u65f6\u95f4\uff0c\u4f7f\u5f97\u534a\u76d1\u7763\u7814\u7a76\u5bf9\u7814\u7a76\u8005\uff0c\u7279\u522b\u662f\u5c0f\u7814\u7a76\u56e2\u4f53\u66f4\u52a0\u53cb\u597d\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88ab\u56fd\u9645\u4eba\u5de5\u667a\u80fd\u9886\u57df\u9876\u7ea7\u5b66\u672f\u5927\u4f1a NeurIPS 2022 \u63a5\u6536\u3002<\/p>\n\n\n\n

\n\n\n\n

\u76d1\u7763\u5b66\u4e60\u901a\u8fc7\u6784\u5efa\u6a21\u578b\u6765\u62df\u5408\u6709\u6807\u8bb0\u6570\u636e\uff0c\u5f53\u4f7f\u7528\u76d1\u7763\u5b66\u4e60 (supervised learning)\u5bf9\u5927\u91cf\u9ad8\u8d28\u91cf\u7684\u6807\u8bb0\u6570\u636e\uff08labeled data\uff09\u8fdb\u884c\u8bad\u7ec3\u65f6\uff0c\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u4f1a\u4ea7\u751f\u6709\u7ade\u4e89\u529b\u7684\u7ed3\u679c\u3002\u4f8b\u5982\uff0c\u636e Paperswithcode \u7f51\u7ad9\u7edf\u8ba1\uff0c\u5728 ImageNet \u8fd9\u4e00\u767e\u4e07\u91cf\u7ea7\u7684\u6570\u636e\u96c6\u4e0a\uff0c\u4f20\u7edf\u7684\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\u53ef\u4ee5\u8fbe\u5230\u8d85\u8fc788%\u7684\u51c6\u786e\u7387\u3002\u7136\u800c\uff0c\u83b7\u53d6\u5927\u91cf\u6709\u6807\u7b7e\u7684\u6570\u636e\u5f80\u5f80\u8d39\u65f6\u8d39\u529b\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u7f13\u89e3\u5bf9\u6807\u6ce8\u6570\u636e\u7684\u4f9d\u8d56\uff0c\u534a\u76d1\u7763\u5b66\u4e60\uff08semi-supervised learning\/SSL\uff09\u81f4\u529b\u4e8e\u5728\u4ec5\u6709\u5c11\u91cf\u7684\u6807\u6ce8\u6570\u636e\u65f6\u5229\u7528\u5927\u91cf\u65e0\u6807\u7b7e\u6570\u636e\uff08unlabeled data\uff09\u6765\u63d0\u5347\u6a21\u578b\u7684\u6cdb\u5316\u6027\u3002\u534a\u76d1\u7763\u5b66\u4e60\u4ea6\u662f\u673a\u5668\u5b66\u4e60\u7684\u91cd\u8981\u4e3b\u9898\u4e4b\u4e00\u3002\u6df1\u5ea6\u5b66\u4e60\u4e4b\u524d\uff0c\u8fd9\u4e00\u9886\u57df\u7684\u7814\u7a76\u8005\u4eec\u63d0\u51fa\u4e86\u8bf8\u5982\u534a\u76d1\u7763\u652f\u6301\u5411\u91cf\u673a\u3001\u71b5\u6b63\u5219\u5316\u3001\u534f\u540c\u8bad\u7ec3\u7b49\u7ecf\u5178\u7b97\u6cd5\u3002<\/p>\n\n\n\n

\u6df1\u5ea6\u534a\u76d1\u7763\u5b66\u4e60<\/h2>\n\n\n\n

\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u5174\u8d77\uff0c\u6df1\u5ea6\u534a\u76d1\u7763\u5b66\u4e60\u7b97\u6cd5\u4e5f\u53d6\u5f97\u4e86\u957f\u8db3\u7684\u8fdb\u6b65\u3002\u540c\u65f6\uff0c\u5305\u62ec\u5fae\u8f6f\u3001\u8c37\u6b4c\u3001\u548c Meta \u7b49\u5728\u5185\u7684\u79d1\u6280\u516c\u53f8\u4e5f\u8ba4\u8bc6\u5230\u4e86\u534a\u76d1\u7763\u5b66\u4e60\u5728\u5b9e\u9645\u573a\u666f\u4e2d\u7684\u5de8\u5927\u6f5c\u529b\u3002\u4f8b\u5982\uff0c\u8c37\u6b4c\u5229\u7528\u566a\u58f0\u5b66\u751f\u8bad\u7ec3\uff08noisy student training\uff09\u8fd9\u4e00\u534a\u76d1\u7763\u7b97\u6cd5\u63d0\u9ad8\u4e86\u5176\u5728\u641c\u7d22\u65b9\u9762\u7684\u6027\u80fd[1]\u3002\u5f53\u524d\u6700\u5177\u4ee3\u8868\u6027\u7684\u534a\u76d1\u7763\u7b97\u6cd5\u901a\u5e38\u5bf9\u6807\u6ce8\u6570\u636e\u4f7f\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u8fdb\u884c\u8bad\u7ec3\uff0c\u5bf9\u65e0\u6807\u6ce8\u6570\u636e\u4f7f\u7528\u4e00\u81f4\u6027\u6b63\u5219\u6280\u672f\uff08consistency regularization\uff09\u9f13\u52b1\u5bf9\u8f93\u5165\u6270\u52a8\u8fdb\u884c\u4e0d\u53d8\u9884\u6d4b\u3002\u4f8b\u5982\uff0c\u8c37\u6b4c\u5728 NeurIPS 2020 \u63d0\u51fa\u7684 FixMatch[2] \u7b97\u6cd5\uff0c\u5229\u7528\u589e\u5f3a\u951a\u5b9a\uff08augmentation anchoring\uff09\u548c\u56fa\u5b9a\u9608\u503c\uff08fixed thresholding\uff09\u6280\u672f\u6765\u589e\u5f3a\u6a21\u578b\u5bf9\u4e0d\u540c\u5f3a\u5ea6\u589e\u5f3a\u6570\u636e\u7684\u6cdb\u5316\u6027\u548c\u51cf\u5c11\u566a\u58f0\u4f2a\u6807\u7b7e\uff08noisy pseudo labels\uff09\u7684\u5f71\u54cd\u3002\u5728\u8bad\u7ec3\u4e2d\uff0cFixMatch \u8fc7\u6ee4\u4e86\u4f4e\u4e8e\u7528\u6237\u6307\u5b9a\uff08user-provided \/ pre-defined\uff09\u9608\u503c\u7684\u65e0\u6807\u7b7e\u6570\u636e\u3002<\/p>\n\n\n\n

\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0e\u4e1c\u4eac\u5de5\u4e1a\u5927\u5b66\u7b49\u5728 NeurIPS 2021 \u5408\u4f5c\u63d0\u51fa\u7684 FlexMatch[3] \u5219\u8003\u8651\u5230\u4e86\u4e0d\u540c\u7c7b\u4e4b\u95f4\u7684\u5b66\u4e60\u96be\u5ea6\u4e0d\u540c\uff0c\u56e0\u6b64\u63d0\u51fa\u4e86\u8bfe\u7a0b\u4f2a\u6807\u7b7e\uff08curriculum pseudo labeling\uff09\u6280\u672f\uff0c\u5bf9\u4e8e\u4e0d\u540c\u7c7b\u5e94\u8be5\u91c7\u7528\u4e0d\u540c\u7684\u9608\u503c\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u5bb9\u6613\u5b66\u4e60\u7684\u7c7b\u522b\uff0c\u6a21\u578b\u5e94\u8be5\u8bbe\u7f6e\u9ad8\u9608\u503c\u4ee5\u964d\u4f4e\u566a\u58f0\u4f2a\u6807\u7b7e\u7684\u5f71\u54cd\uff1b\u5bf9\u4e8e\u96be\u5b66\u4e60\u7684\u7c7b\u522b\uff0c\u6a21\u578b\u5e94\u8be5\u8bbe\u7f6e\u4f4e\u9608\u503c\u9f13\u52b1\u8be5\u7c7b\u7684\u62df\u5408\u3002\u6bcf\u4e2a\u7c7b\u7684\u5b66\u4e60\u96be\u5ea6\u8bc4\u4f30\u53d6\u51b3\u4e8e\u843d\u5165\u8be5\u7c7b\u4e14\u9ad8\u4e8e\u56fa\u5b9a\u503c\u7684\u672a\u6807\u8bb0\u6570\u636e\u6837\u672c\u7684\u6570\u91cf\u3002<\/p>\n\n\n\n

\u540c\u65f6\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u8fd8\u5408\u4f5c\u63d0\u51fa\u4e86\u4e00\u4e2a\u7edf\u4e00\u7684\u57fa\u4e8e Pytorch \u7684\u534a\u76d1\u7763\u65b9\u6cd5\u4ee3\u7801\u5e93 TorchSSL[4]\uff0c\u5bf9\u8be5\u9886\u57df\u7684\u6df1\u5ea6\u65b9\u6cd5\u3001\u5e38\u7528\u6570\u636e\u96c6\u548c\u57fa\u51c6\u7ed3\u679c\u8fdb\u884c\u4e86\u7edf\u4e00\u7684\u652f\u6301\u3002<\/p>\n\n\n\n

\"\u56fe1\uff1aFlexMatch
\u56fe1\uff1aFlexMatch \u7b97\u6cd5\u6d41\u7a0b<\/em><\/figcaption><\/figure>\n\n\n\n

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

\u5f53\u524d\u534a\u76d1\u7763\u5b66\u4e60\u4ee3\u7801\u5e93\u5b58\u5728\u7684\u95ee\u9898\u4e0e\u6311\u6218<\/h2>\n\n\n\n

\u5c3d\u7ba1\u534a\u76d1\u7763\u5b66\u4e60\u7684\u53d1\u5c55\u5982\u706b\u5982\u837c\uff0c\u4f46\u662f\uff0c\u7814\u7a76\u5458\u4eec\u6ce8\u610f\u5230\u76ee\u524d\u5927\u90e8\u5206\u534a\u76d1\u7763\u65b9\u5411\u7684\u8bba\u6587\u53ea\u5173\u6ce8\u8ba1\u7b97\u673a\u89c6\u89c9 (CV) \u5206\u7c7b\u4efb\u52a1\uff0c\u5bf9\u4e8e\u5176\u4ed6\u9886\u57df\uff0c\u4f8b\u5982\u81ea\u7136\u8bed\u8a00\u5904\u7406 (NLP)\u3001\u97f3\u9891\u5904\u7406 (audio)\uff0c\u7814\u7a76\u8005\u65e0\u6cd5\u5f97\u77e5\u8fd9\u4e9b\u5728 CV \u4efb\u52a1\u4e0a\u6709\u6548\u7684\u7b97\u6cd5\u5230\u4e86\u4e0d\u540c\u9886\u57df\u662f\u5426\u4f9d\u7136\u6709\u6548\u3002\u53e6\u5916\uff0c\u5927\u90e8\u5206\u534a\u76d1\u7763\u76f8\u5173\u7684\u8bba\u6587\u90fd\u662f\u7531\u5927\u578b\u673a\u6784\u53d1\u8868\uff0c\u5b66\u672f\u754c\u7684\u5b9e\u9a8c\u5ba4\u5f80\u5f80\u7531\u4e8e\u8ba1\u7b97\u8d44\u6e90\u7684\u9650\u5236\u800c\u5f88\u96be\u53c2\u4e0e\u5230\u63a8\u52a8\u8be5\u9886\u57df\u7684\u53d1\u5c55\u4e2d\u3002\u603b\u7684\u6765\u8bf4\uff0c\u534a\u76d1\u7763\u5b66\u4e60\u57fa\u51c6\u76ee\u524d\u5b58\u5728\u4ee5\u4e0b\u4e24\u4e2a\u95ee\u9898\uff1a<\/p>\n\n\n\n

\uff081\uff09\u591a\u6837\u6027\u4e0d\u8db3\u3002\u73b0\u6709\u7684\u534a\u76d1\u7763\u5b66\u4e60\u57fa\u51c6\u5927\u591a\u5c40\u9650\u4e8e CV \u5206\u7c7b\u4efb\u52a1\uff08\u5373 CIFAR-10\/100\uff0cSVHN\uff0cSTL-10 \u548c ImageNet \u5206\u7c7b\uff09\uff0c\u6392\u9664\u4e86\u5bf9 NLP\u3001audio \u7b49\u5206\u7c7b\u4efb\u52a1\u7684\u4e00\u81f4\u548c\u591a\u6837\u5316\u8bc4\u4f30\uff0c\u800c\u5728 NLP \u548c audio \u4e2d\u7f3a\u4e4f\u8db3\u591f\u7684\u6807\u8bb0\u6570\u636e\u4e5f\u662f\u4e00\u4e2a\u666e\u904d\u95ee\u9898\u3002<\/p>\n\n\n\n

\uff082\uff09\u8017\u65f6\u4e14\u5bf9\u5b66\u672f\u754c\u4e0d\u53cb\u597d\u3002\u73b0\u6709\u7684\u534a\u76d1\u7763\u5b66\u4e60\u57fa\u51c6\uff08\u5982 TorchSSL\uff09\u901a\u5e38\u662f\u8017\u65f6\u4e14\u4e0d\u73af\u4fdd\u7684\uff0c\u56e0\u4e3a\u5b83\u5f80\u5f80\u9700\u8981\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u3002\u5177\u4f53\u800c\u8a00\uff0c\u4f7f\u7528TorchSSL \u8bc4\u4f30 FixMatch[1]\u5927\u7ea6\u9700\u8981300\u4e2a GPU \u65e5\u3002\u5982\u6b64\u9ad8\u7684\u8bad\u7ec3\u6210\u672c\u4f7f\u5f97\u8bb8\u591a\u7814\u7a76\u5b9e\u9a8c\u5ba4\uff08\u5c24\u5176\u662f\u5b66\u672f\u754c\u7684\u5b9e\u9a8c\u5ba4\u6216\u5c0f\u7814\u7a76\u56e2\u4f53\uff09\u65e0\u6cd5\u8d1f\u62c5\u5f97\u8d77 SSL \u7684\u76f8\u5173\u7814\u7a76\uff0c\u4ece\u800c\u963b\u788d\u4e86 SSL \u7684\u8fdb\u5c55\u3002<\/p>\n\n\n\n

USB\uff1a\u4efb\u52a1\u591a\u6837\u5316\u4e14\u5bf9\u7814\u7a76\u8005\u66f4\u53cb\u597d\u7684\u65b0\u57fa\u51c6\u5e93<\/h2>\n\n\n\n

\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u8054\u5408\u897f\u6e56\u5927\u5b66\u3001\u4e1c\u4eac\u5de5\u4e1a\u5927\u5b66\u3001\u5361\u5185\u57fa\u6885\u9686\u5927\u5b66\u3001\u9a6c\u514b\u65af-\u666e\u6717\u514b\u7814\u7a76\u6240\u7b49\u673a\u6784\u7684\u79d1\u7814\u4eba\u5458\u63d0\u51fa\u4e86 Unified SSL Benchmark\uff08USB\uff09\uff0c\u8fd9\u662f\u7b2c\u4e00\u4e2a\u5c06\u89c6\u89c9\u3001\u8bed\u8a00\u548c\u97f3\u9891\u5206\u7c7b\u4efb\u52a1\u8fdb\u884c\u7edf\u4e00\u7684\u534a\u76d1\u7763\u5206\u7c7b\u5b66\u4e60\u57fa\u51c6\u3002\u76f8\u6bd4\u4e8e\u4e4b\u524d\u7684\u534a\u76d1\u7763\u5b66\u4e60\u57fa\u51c6\uff08\u5982 TorchSSL\uff09\u53ea\u5173\u6ce8\u5c11\u91cf\u89c6\u89c9\u4efb\u52a1\uff0c\u8be5\u57fa\u51c6\u4e0d\u4ec5\u5f15\u5165\u4e86\u66f4\u591a\u6837\u5316\u7684\u5e94\u7528\u9886\u57df\uff0c\u8fd8\u9996\u6b21\u5229\u7528\u89c6\u89c9\u9884\u8bad\u7ec3\u6a21\u578b\uff08pretrained vision Transformer\uff09\u5927\u5927\u7f29\u51cf\u4e86\u534a\u76d1\u7763\u7b97\u6cd5\u7684\u9a8c\u8bc1\u65f6\u95f4\uff08\u4ece7000 GPU \u65f6\u7f29\u51cf\u81f3900 GPU \u65f6\uff09\uff0c\u4ece\u800c\u4f7f\u5f97\u534a\u76d1\u7763\u7814\u7a76\u5bf9\u7814\u7a76\u8005\u3001\u7279\u522b\u662f\u5c0f\u7814\u7a76\u56e2\u4f53\u66f4\u4e3a\u53cb\u597d\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88ab\u56fd\u9645\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u9876\u7ea7\u5b66\u672f\u5927\u4f1a NeurIPS 2022 \u63a5\u6536\u3002\uff08\u70b9\u51fb\u9605\u8bfb\u539f\u6587\uff0c\u4e86\u89e3\u8bba\u6587\u7ec6\u8282\uff09<\/p>\n\n\n\n

\u6587\u7ae0\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2208.07204.pdf
\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/microsoft\/Semi-supervised-learning<\/p>\n\n\n\n

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

USB \u63d0\u4f9b\u7684\u89e3\u51b3\u65b9\u6848<\/h2>\n\n\n\n

\u90a3\u4e48\uff0cUSB \u5982\u4f55\u4e00\u6b21\u6027\u89e3\u51b3\u5f53\u524d\u534a\u76d1\u7763\u57fa\u51c6\u6240\u5b58\u5728\u7684\u95ee\u9898\u5462\uff1f\u7814\u7a76\u5458\u4eec\u4e3b\u8981\u8fdb\u884c\u4e86\u5982\u4e0b\u6539\u8fdb\uff1a<\/p>\n\n\n\n

\uff081\uff09\u4e3a\u589e\u5f3a\u4efb\u52a1\u591a\u6837\u6027\uff0cUSB \u5f15\u5165\u4e865\u4e2a CV \u6570\u636e\u96c6\uff0c5\u4e2a NLP \u6570\u636e\u96c6\u548c5\u4e2a audio \u6570\u636e\u96c6\uff0c\u5e76\u63d0\u4f9b\u4e86\u4e00\u4e2a\u591a\u6837\u5316\u4e14\u5177\u6709\u6311\u6218\u6027\u7684\u57fa\u51c6\uff0c\u4ece\u800c\u80fd\u591f\u5bf9\u6765\u81ea\u4e0d\u540c\u9886\u57df\u7684\u591a\u4e2a\u4efb\u52a1\u8fdb\u884c\u4e00\u81f4\u7684\u8bc4\u4f30\u3002\u88681\u63d0\u4f9b\u4e86 USB \u4e0e TorchSSL \u7684\u4efb\u52a1\u548c\u8bad\u7ec3\u65f6\u95f4\u7b49\u65b9\u9762\u7684\u8be6\u7ec6\u5bf9\u6bd4\u3002<\/p>\n\n\n\n

\"\u88681\uff1aUSB
\u88681\uff1aUSB \u4e0e TorchSSL \u6846\u67b6\u7684\u4efb\u52a1\u548c\u8bad\u7ec3\u65f6\u95f4\u5bf9\u6bd4<\/em><\/figcaption><\/figure>\n\n\n\n

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

\uff082\uff09\u4e3a\u4e86\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\uff0c\u7814\u7a76\u5458\u4eec\u5c06\u9884\u8bad\u7ec3\u7684 vision Transformer \u5f15\u5165 SSL\uff0c\u800c\u4e0d\u662f\u4ece\u5934\u8bad\u7ec3 ResNets\u3002\u5177\u4f53\u800c\u8a00\uff0c\u7814\u7a76\u5458\u4eec\u53d1\u73b0\u5728\u4e0d\u5f71\u54cd\u6027\u80fd\u7684\u60c5\u51b5\u4e0b\u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b\u53ef\u4ee5\u5927\u5927\u51cf\u5c11\u8bad\u7ec3\u8fed\u4ee3\u6b21\u6570\uff08\u4f8b\u5982\uff0c\u5c06 CV \u4efb\u52a1\u7684\u8bad\u7ec3\u8fed\u4ee3\u6b21\u6570\u4ece100\u4e07\u6b65\u51cf\u5c11\u523020\u4e07\u6b65\uff09\u3002<\/p>\n\n\n\n

\uff083\uff09\u4e3a\u4e86\u5bf9\u7814\u7a76\u4eba\u5458\u66f4\u52a0\u53cb\u597d\uff0c\u7814\u7a76\u5458\u4eec\u5f00\u6e90\u5b9e\u73b0\u4e8614\u79cd SSL \u7b97\u6cd5\u5e76\u5f00\u6e90\u4e86\u4e00\u4e2a\u6a21\u5757\u5316\u4ee3\u7801\u5e93\u548c\u76f8\u5173\u7684\u914d\u7f6e\u6587\u4ef6\u4ee5\u4f9b\u7814\u7a76\u8005\u8f7b\u677e\u518d\u73b0 USB \u62a5\u544a\u4e2d\u7684\u7ed3\u679c\u3002\u4e3a\u4e86\u5feb\u901f\u4e0a\u624b\uff0cUSB \u8fd8\u63d0\u4f9b\u4e86\u8be6\u7ec6\u7684\u6587\u6863\u548c\u6559\u7a0b\u3002\u6b64\u5916\uff0cUSB \u8fd8\u63d0\u4f9b\u4e86 pip \u5305\u4ee5\u4f9b\u4f7f\u7528\u8005\u76f4\u63a5\u8c03\u7528 SSL \u7b97\u6cd5\u3002\u7814\u7a76\u5458\u4eec\u627f\u8bfa\u672a\u6765\u4f1a\u5728 USB \u4e2d\u4e0d\u65ad\u52a0\u5165\u65b0\u7684\u7b97\u6cd5\uff08\u4f8b\u5982\u4e0d\u5e73\u8861\u534a\u76d1\u7763\u7b97\u6cd5\u7b49\uff09\u548c\u66f4\u591a\u66f4\u5177\u6311\u6218\u6027\u7684\u6570\u636e\u96c6\u3002\u88682\u5c55\u793a\u4e86 USB \u4e2d\u5df2\u652f\u6301\u7684\u7b97\u6cd5\u548c\u6a21\u5757\u3002<\/p>\n\n\n\n

\"\u88682\uff1aUSB
\u88682\uff1aUSB \u4e2d\u5df2\u652f\u6301\u7684\u7b97\u6cd5\u548c\u6a21\u5757<\/em><\/figcaption><\/figure>\n\n\n\n

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

\u534a\u76d1\u7763\u5b66\u4e60\u901a\u8fc7\u5229\u7528\u5927\u91cf\u65e0\u6807\u7b7e\u6570\u636e\u6765\u8bad\u7ec3\u66f4\u7cbe\u786e\u3001\u66f4\u9c81\u68d2\u7684\u6a21\u578b\uff0c\u5728\u672a\u6765\u6709\u7740\u91cd\u8981\u7684\u7814\u7a76\u548c\u5e94\u7528\u4ef7\u503c\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u671f\u5f85\u901a\u8fc7 USB \u8fd9\u4e00\u5de5\u4f5c\uff0c\u80fd\u591f\u4e88\u529b\u5b66\u672f\u754c\u548c\u5de5\u4e1a\u754c\u5728\u534a\u76d1\u7763\u5b66\u4e60\u9886\u57df\u53d6\u5f97\u66f4\u5927\u7684\u8fdb\u5c55\u3002<\/p>\n\n\n\n

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

[1] https:\/\/ai.googleblog.com\/2021\/07\/from-vision-to-language-semi-supervised.html<\/p>\n\n\n\n

[2] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596\u2013608, 2020.<\/p>\n\n\n\n

[3] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34, 2021.<\/p>\n\n\n\n

[4] TorchSSL: https:\/\/github.com\/TorchSSL\/TorchSSL<\/p>\n","protected":false},"excerpt":{"rendered":"

\u4f5c\u8005\uff1a\u738b\u4e00\u680b\u3001\u738b\u664b\u4e1c \u7f16\u8005\u6309\uff1a\u5f53\u524d\uff0c\u534a\u76d1\u7763\u5b66\u4e60\u7684\u53d1\u5c55\u5982\u706b\u5982\u837c\u3002\u4f46\u662f\u73b0\u6709\u7684\u534a\u76d1\u7763\u5b66\u4e60\u57fa\u51c6\u5927\u591a\u5c40\u9650\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u5206\u7c7b\u4efb\u52a1\uff0c\u6392\u9664\u4e86\u5bf9\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u97f3\u9891\u5904\u7406\u7b49\u5206\u7c7b\u4efb\u52a1\u7684\u4e00\u81f4\u548c\u591a\u6837\u5316\u8bc4\u4f30\u3002\u6b64\u5916\uff0c\u5927\u90e8\u5206\u534a\u76d1\u7763\u8bba\u6587\u7531\u5927\u578b\u673a\u6784\u53d1\u8868\uff0c\u5b66\u672f\u754c\u7684\u5b9e\u9a8c\u5ba4\u5f80\u5f80\u7531\u4e8e\u8ba1\u7b97\u8d44\u6e90\u7684\u9650\u5236\u800c\u5f88\u96be\u53c2\u4e0e\u5230\u63a8\u52a8\u8be5\u9886\u57df\u7684\u53d1\u5c55\u4e2d\u3002\u4e3a\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u8054\u5408\u897f\u6e56\u5927\u5b66\u3001\u4e1c\u4eac\u5de5\u4e1a\u5927\u5b66\u3001\u5361\u5185\u57fa\u6885\u9686\u5927\u5b66\u3001\u9a6c\u514b\u65af-\u666e\u6717\u514b\u7814\u7a76\u6240\u7b49\u673a\u6784\u7684\u79d1\u7814\u4eba\u5458\u63d0\u51fa\u4e86 Unified SSL Benchmark\uff08USB\uff09\uff1a\u7b2c\u4e00\u4e2a\u5c06\u89c6\u89c9\u3001\u8bed\u8a00\u548c\u97f3\u9891\u5206\u7c7b\u4efb\u52a1\u8fdb\u884c\u7edf\u4e00\u7684\u534a\u76d1\u7763\u5206\u7c7b\u5b66\u4e60\u57fa\u51c6\u3002\u8be5\u8bba\u6587\u4e0d\u4ec5\u5f15\u5165\u4e86\u66f4\u591a\u6837\u5316\u7684\u5e94\u7528\u9886\u57df\uff0c\u8fd8\u9996\u6b21\u5229\u7528\u89c6\u89c9\u9884\u8bad\u7ec3\u6a21\u578b\u5927\u5927\u7f29\u51cf\u4e86\u534a\u76d1\u7763\u7b97\u6cd5\u7684\u9a8c\u8bc1\u65f6\u95f4\uff0c\u4f7f\u5f97\u534a\u76d1\u7763\u7814\u7a76\u5bf9\u7814\u7a76\u8005\uff0c\u7279\u522b\u662f\u5c0f\u7814\u7a76\u56e2\u4f53\u66f4\u52a0\u53cb\u597d\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88ab\u56fd\u9645\u4eba\u5de5\u667a\u80fd\u9886\u57df\u9876\u7ea7\u5b66\u672f\u5927\u4f1a NeurIPS 2022 \u63a5\u6536\u3002 \u76d1\u7763\u5b66\u4e60\u901a\u8fc7\u6784\u5efa\u6a21\u578b\u6765\u62df\u5408\u6709\u6807\u8bb0\u6570\u636e\uff0c\u5f53\u4f7f\u7528\u76d1\u7763\u5b66\u4e60 (supervised learning)\u5bf9\u5927\u91cf\u9ad8\u8d28\u91cf\u7684\u6807\u8bb0\u6570\u636e\uff08labeled data\uff09\u8fdb\u884c\u8bad\u7ec3\u65f6\uff0c\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u4f1a\u4ea7\u751f\u6709\u7ade\u4e89\u529b\u7684\u7ed3\u679c\u3002\u4f8b\u5982\uff0c\u636e Paperswithcode \u7f51\u7ad9\u7edf\u8ba1\uff0c\u5728 ImageNet \u8fd9\u4e00\u767e\u4e07\u91cf\u7ea7\u7684\u6570\u636e\u96c6\u4e0a\uff0c\u4f20\u7edf\u7684\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\u53ef\u4ee5\u8fbe\u5230\u8d85\u8fc788%\u7684\u51c6\u786e\u7387\u3002\u7136\u800c\uff0c\u83b7\u53d6\u5927\u91cf\u6709\u6807\u7b7e\u7684\u6570\u636e\u5f80\u5f80\u8d39\u65f6\u8d39\u529b\u3002 \u4e3a\u4e86\u7f13\u89e3\u5bf9\u6807\u6ce8\u6570\u636e\u7684\u4f9d\u8d56\uff0c\u534a\u76d1\u7763\u5b66\u4e60\uff08semi-supervised learning\/SSL\uff09\u81f4\u529b\u4e8e\u5728\u4ec5\u6709\u5c11\u91cf\u7684\u6807\u6ce8\u6570\u636e\u65f6\u5229\u7528\u5927\u91cf\u65e0\u6807\u7b7e\u6570\u636e\uff08unlabeled data\uff09\u6765\u63d0\u5347\u6a21\u578b\u7684\u6cdb\u5316\u6027\u3002\u534a\u76d1\u7763\u5b66\u4e60\u4ea6\u662f\u673a\u5668\u5b66\u4e60\u7684\u91cd\u8981\u4e3b\u9898\u4e4b\u4e00\u3002\u6df1\u5ea6\u5b66\u4e60\u4e4b\u524d\uff0c\u8fd9\u4e00\u9886\u57df\u7684\u7814\u7a76\u8005\u4eec\u63d0\u51fa\u4e86\u8bf8\u5982\u534a\u76d1\u7763\u652f\u6301\u5411\u91cf\u673a\u3001\u71b5\u6b63\u5219\u5316\u3001\u534f\u540c\u8bad\u7ec3\u7b49\u7ecf\u5178\u7b97\u6cd5\u3002 \u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u5174\u8d77\uff0c\u6df1\u5ea6\u534a\u76d1\u7763\u5b66\u4e60\u7b97\u6cd5\u4e5f\u53d6\u5f97\u4e86\u957f\u8db3\u7684\u8fdb\u6b65\u3002\u540c\u65f6\uff0c\u5305\u62ec\u5fae\u8f6f\u3001\u8c37\u6b4c\u3001\u548c Meta \u7b49\u5728\u5185\u7684\u79d1\u6280\u516c\u53f8\u4e5f\u8ba4\u8bc6\u5230\u4e86\u534a\u76d1\u7763\u5b66\u4e60\u5728\u5b9e\u9645\u573a\u666f\u4e2d\u7684\u5de8\u5927\u6f5c\u529b\u3002\u4f8b\u5982\uff0c\u8c37\u6b4c\u5229\u7528\u566a\u58f0\u5b66\u751f\u8bad\u7ec3\uff08noisy student training\uff09\u8fd9\u4e00\u534a\u76d1\u7763\u7b97\u6cd5\u63d0\u9ad8\u4e86\u5176\u5728\u641c\u7d22\u65b9\u9762\u7684\u6027\u80fd[1]\u3002\u5f53\u524d\u6700\u5177\u4ee3\u8868\u6027\u7684\u534a\u76d1\u7763\u7b97\u6cd5\u901a\u5e38\u5bf9\u6807\u6ce8\u6570\u636e\u4f7f\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u8fdb\u884c\u8bad\u7ec3\uff0c\u5bf9\u65e0\u6807\u6ce8\u6570\u636e\u4f7f\u7528\u4e00\u81f4\u6027\u6b63\u5219\u6280\u672f\uff08consistency regularization\uff09\u9f13\u52b1\u5bf9\u8f93\u5165\u6270\u52a8\u8fdb\u884c\u4e0d\u53d8\u9884\u6d4b\u3002\u4f8b\u5982\uff0c\u8c37\u6b4c\u5728 NeurIPS 2020 \u63d0\u51fa\u7684 FixMatch[2] \u7b97\u6cd5\uff0c\u5229\u7528\u589e\u5f3a\u951a\u5b9a\uff08augmentation anchoring\uff09\u548c\u56fa\u5b9a\u9608\u503c\uff08fixed thresholding\uff09\u6280\u672f\u6765\u589e\u5f3a\u6a21\u578b\u5bf9\u4e0d\u540c\u5f3a\u5ea6\u589e\u5f3a\u6570\u636e\u7684\u6cdb\u5316\u6027\u548c\u51cf\u5c11\u566a\u58f0\u4f2a\u6807\u7b7e\uff08noisy pseudo labels\uff09\u7684\u5f71\u54cd\u3002\u5728\u8bad\u7ec3\u4e2d\uff0cFixMatch \u8fc7\u6ee4\u4e86\u4f4e\u4e8e\u7528\u6237\u6307\u5b9a\uff08user-provided \/ pre-defined\uff09\u9608\u503c\u7684\u65e0\u6807\u7b7e\u6570\u636e\u3002 \u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0e\u4e1c\u4eac\u5de5\u4e1a\u5927\u5b66\u7b49\u5728 NeurIPS 2021 \u5408\u4f5c\u63d0\u51fa\u7684 FlexMatch[3] \u5219\u8003\u8651\u5230\u4e86\u4e0d\u540c\u7c7b\u4e4b\u95f4\u7684\u5b66\u4e60\u96be\u5ea6\u4e0d\u540c\uff0c\u56e0\u6b64\u63d0\u51fa\u4e86\u8bfe\u7a0b\u4f2a\u6807\u7b7e\uff08curriculum pseudo labeling\uff09\u6280\u672f\uff0c\u5bf9\u4e8e\u4e0d\u540c\u7c7b\u5e94\u8be5\u91c7\u7528\u4e0d\u540c\u7684\u9608\u503c\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u5bb9\u6613\u5b66\u4e60\u7684\u7c7b\u522b\uff0c\u6a21\u578b\u5e94\u8be5\u8bbe\u7f6e\u9ad8\u9608\u503c\u4ee5\u964d\u4f4e\u566a\u58f0\u4f2a\u6807\u7b7e\u7684\u5f71\u54cd\uff1b\u5bf9\u4e8e\u96be\u5b66\u4e60\u7684\u7c7b\u522b\uff0c\u6a21\u578b\u5e94\u8be5\u8bbe\u7f6e\u4f4e\u9608\u503c\u9f13\u52b1\u8be5\u7c7b\u7684\u62df\u5408\u3002\u6bcf\u4e2a\u7c7b\u7684\u5b66\u4e60\u96be\u5ea6\u8bc4\u4f30\u53d6\u51b3\u4e8e\u843d\u5165\u8be5\u7c7b\u4e14\u9ad8\u4e8e\u56fa\u5b9a\u503c\u7684\u672a\u6807\u8bb0\u6570\u636e\u6837\u672c\u7684\u6570\u91cf\u3002 \u540c\u65f6\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u8fd8\u5408\u4f5c\u63d0\u51fa\u4e86\u4e00\u4e2a\u7edf\u4e00\u7684\u57fa\u4e8e Pytorch \u7684\u534a\u76d1\u7763\u65b9\u6cd5\u4ee3\u7801\u5e93 TorchSSL[4]\uff0c\u5bf9\u8be5\u9886\u57df\u7684\u6df1\u5ea6\u65b9\u6cd5\u3001\u5e38\u7528\u6570\u636e\u96c6\u548c\u57fa\u51c6\u7ed3\u679c\u8fdb\u884c\u4e86\u7edf\u4e00\u7684\u652f\u6301\u3002 \u5c3d\u7ba1\u534a\u76d1\u7763\u5b66\u4e60\u7684\u53d1\u5c55\u5982\u706b\u5982\u837c\uff0c\u4f46\u662f\uff0c\u7814\u7a76\u5458\u4eec\u6ce8\u610f\u5230\u76ee\u524d\u5927\u90e8\u5206\u534a\u76d1\u7763\u65b9\u5411\u7684\u8bba\u6587\u53ea\u5173\u6ce8\u8ba1\u7b97\u673a\u89c6\u89c9 (CV) \u5206\u7c7b\u4efb\u52a1\uff0c\u5bf9\u4e8e\u5176\u4ed6\u9886\u57df\uff0c\u4f8b\u5982\u81ea\u7136\u8bed\u8a00\u5904\u7406 (NLP)\u3001\u97f3\u9891\u5904\u7406 (audio)\uff0c\u7814\u7a76\u8005\u65e0\u6cd5\u5f97\u77e5\u8fd9\u4e9b\u5728 CV […]<\/p>\n","protected":false},"author":34512,"featured_media":1086642,"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-1069890","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\/1069890","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":4,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1069890\/revisions"}],"predecessor-version":[{"id":1086645,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1069890\/revisions\/1086645"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media\/1086642"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1069890"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1069890"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1069890"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1069890"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}