{"id":34307,"date":"2024-12-04T06:56:51","date_gmt":"2024-12-03T22:56:51","guid":{"rendered":"https:\/\/17aitech.com\/?p=34307"},"modified":"2024-12-04T06:56:51","modified_gmt":"2024-12-03T22:56:51","slug":"icml-2024-%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e5%8a%a9%e5%8a%9b%e5%9f%ba%e4%ba%8eclip%e7%9a%84%e5%88%86%e5%b8%83%e5%a4%96%e6%a3%80%e6%b5%8b%e4%bb%bb%e5%8a%a1","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34307","title":{"rendered":"ICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1"},"content":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:ICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1<\/a><\/p>\n

\u5f53\u8bad\u7ec3\u6570\u636e\u96c6\u548c\u6d4b\u8bd5\u6570\u636e\u96c6\u7684\u5206\u5e03\u76f8\u540c\u65f6\uff0c\u673a\u5668\u5b66\u4e60<\/mark>\u6a21\u578b\u53ef\u4ee5\u8868\u73b0\u51fa\u4f18\u8d8a\u7684\u6027\u80fd\u3002\u7136\u800c\u5728\u5f00\u653e\u4e16\u754c\u73af\u5883\u4e2d\uff0c\u6a21\u578b\u7ecf\u5e38\u4f1a\u9047\u5230\u5206\u5e03\u5916\uff08Out-of-Distribution, OOD\uff0c\u4e0b\u79f0\u201cOOD\u201d\uff09\u6837\u672c\uff0cOOD\u6837\u672c\u53ef\u80fd\u4f1a\u5bfc\u81f4\u6a21\u578b\u505a\u51fa\u4e0d\u53ef\u9884\u6d4b\u7684\u884c\u4e3a\uff0c\u800c\u9519\u8bef\u7684\u540e\u679c\u53ef\u80fd\u662f\u81f4\u547d\u7684\uff0c\u5c24\u5176\u662f\u5728\u81ea\u52a8\u9a7e\u9a76\u7b49\u9ad8\u98ce\u9669\u573a\u666f\u4e2d [1, 2]\u3002\u56e0\u6b64OOD\u68c0\u6d4b\u5bf9\u4e8e\u4fdd\u969c\u673a\u5668\u5b66\u4e60<\/mark>\u6a21\u578b\u5728\u5b9e\u9645\u90e8\u7f72\u4e2d\u7684\u53ef\u9760\u6027\u81f3\u5173\u91cd\u8981\u3002<\/p>\n

\u5927\u591a\u6570OOD\u68c0\u6d4b\u65b9\u6cd5 [1, 3] \u53ef\u4ee5\u57fa\u4e8e\u8bad\u7ec3\u6709\u7d20\u7684\u5206\u5e03\u5185 (In-Distribution, ID) \u5206\u7c7b\u5668\u6709\u6548\u5730\u68c0\u6d4b OOD \u6837\u672c\u3002\u7136\u800c\uff0c\u5bf9\u4e8e\u4e0d\u540c\u7684ID\u6570\u636e\u96c6\uff0c\u5b83\u4eec\u9700\u8981\u91cd\u65b0\u8bad\u7ec3\u5206\u7c7b\u5668\u624d\u80fd\u8fdb\u884cOOD\u68c0\u6d4b\u3002\u6b64\u5916\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u4ec5\u4f9d\u8d56\u4e8e\u89c6\u89c9\u6a21\u5f0f\uff0c\u800c\u5ffd\u7565\u4e86\u89c6\u89c9\u56fe\u50cf\u4e0e\u6587\u672c\u6807\u7b7e\u4e4b\u95f4\u7684\u8054\u7cfb\u3002\u968f\u7740\u5927\u89c4\u6a21\u89c6\u89c9\u8bed\u8a00\u6a21\u578b<\/mark>\uff08\ufeffVision-Manguage Models , VLMs\uff0c\u4f8b\u5982CLIP [4]\uff09\u7684\u51fa\u73b0\uff0c\u4f7f\u5f97\u96f6\u6837\u672c OOD \u68c0\u6d4b\u6210\u4e3a\u4e86\u53ef\u80fd[5]\u3002\u901a\u8fc7\u6784\u5efa\u4ec5\u5177\u6709 ID \u7c7b\u522b\u6807\u7b7e\u7684\u6587\u672c\u5206\u7c7b<\/mark>\u5668\uff0c\u80fd\u591f\u5b9e\u73b0\u5728\u65e0\u9700\u91cd\u65b0\u8bad\u7ec3\u5206\u7c7b\u5668\u7684\u60c5\u51b5\u4e0b\u8de8\u4e0d\u540c\u7684 ID \u6570\u636e\u96c6\u68c0\u6d4b OOD \u6837\u672c\u3002<\/p>\n

\u5c3d\u7ba1\u73b0\u6709\u7684\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u65b9\u6cd5\u5c55\u73b0\u51fa\u4e86\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u6027\u80fd\uff0c\u4f46\u662f\u5b83\u4eec\u5728\u9047\u5230\u8f83\u96be\u68c0\u6d4b\u7684\u5206\u5e03\u5916\u6837\u672c\u65f6\u5e38\u5e38\u4f1a\u68c0\u6d4b\u5931\u8d25\uff0c\u6211\u4eec\u8ba4\u4e3a\u73b0\u6709\u4ec5\u4f9d\u8d56ID \u7c7b\u522b\u6807\u7b7e\u6784\u5efa\u6587\u672c\u5206\u7c7b<\/mark>\u5668\u7684\u65b9\u6cd5\u5f88\u5927\u7a0b\u5ea6\u4e0a\u9650\u5236\u4e86 CLIP \u8bc6\u522b\u6765\u81ea\u5f00\u653e\u6807\u7b7e\u7a7a\u95f4\u6837\u672c\u7684\u56fa\u6709\u80fd\u529b\u3002\u5982\u56fe1 (a) \u6240\u793a\uff0c\u4ec5\u4f9d\u8d56ID \u7c7b\u522b\u6807\u7b7e\u6784\u5efa\u6587\u672c\u5206\u7c7b<\/mark>\u5668\u7684\u65b9\u6cd5\u96be\u4ee5\u533a\u5206\u8f83\u96be\u68c0\u6d4b\u7684OOD\u6837\u672c \uff08ID\u6570\u636e\u96c6\uff1a\ufeffCUB-200-2011\uff0cOOD\u6570\u636e\u96c6\uff1aPlaces\uff09\u3002<\/p>\n

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

\u56fe1. \u7814\u7a76\u52a8\u673a\u793a\u610f\u56fe\uff1a(a) \u4ec5\u4f9d\u8d56ID \u7c7b\u522b\u6807\u7b7e\u6784\u5efa\u6587\u672c\u5206\u7c7b<\/mark>\u5668, (b) \u4f7f\u7528\u771f\u5b9eOOD\u6807\u7b7e, (c) \u4f7f\u7528LLM\u60f3\u8c61\u6f5c\u5728\u7684\u5f02\u5e38\u503c\u66b4\u9732<\/p>\n

\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a\ufeffEnvisioning Outlier Exposure (EOE) \u7684\u5206\u5e03\u5916\u68c0\u6d4b\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u5229\u7528\u901a\u8fc7\u5229\u7528\u5927\u578b\u8bed\u8a00\u6a21\u578b<\/mark> (LLM) \u7684\u4e13\u5bb6\u77e5\u8bc6\u548c\u63a8\u7406\u80fd\u529b\u6765\u60f3\u8c61\u6f5c\u5728\u7684\u5f02\u5e38\u503c\u66b4\u9732\uff0c\u4ece\u800c\u63d0\u5347VLMs\u7684OOD\u68c0\u6d4b\u6027\u80fd (\u5982\u56fe1 (c) \u6240\u793a)\uff0c\u540c\u65f6\u65e0\u9700\u8bbf\u95ee\u4efb\u4f55\u5b9e\u9645\u7684 OOD \u6570\u636e\u3002\u6211\u4eec\u8bbe\u8ba1\u4e86 (1) \u57fa\u4e8e\u89c6\u89c9\u76f8\u4f3c\u6027\u7684 LLM \u63d0\u793a\uff0c\u4ee5\u751f\u6210\u4e13\u95e8\u7528\u4e8e OOD \u68c0\u6d4b\u7684\u6f5c\u5728\u5f02\u5e38\u503c\u7c7b\u6807\u7b7e\uff0c\u4ee5\u53ca (2) \u57fa\u4e8e\u6f5c\u5728\u5f02\u5e38\u503c\u60e9\u7f5a\u7684\u65b0\u8bc4\u5206\u51fd\u6570\uff0c\u4ee5\u6709\u6548\u533a\u5206\u96be\u4ee5\u8bc6\u522b\u7684 OOD \u6837\u672c\u3002\u5b9e\u9a8c\u8868\u660e\uff0cEOE \u5728\u4e0d\u540c\u7684 OOD \u4efb\u52a1\u4e2d\u5b9e\u73b0\u4e86\u4f18\u8d8a\u7684\u6027\u80fd\uff0c\u5e76\u4e14\u53ef\u4ee5\u6709\u6548\u5730\u6269\u5c55\u5230 ImageNet-1K \u6570\u636e\u96c6\u3002<\/p>\n

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

\u00d8\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2406.00806<\/p>\n

\u00d8\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/tmlr-group\/EOE<\/p>\n

\u63a5\u4e0b\u6765\u5c06\u7b80\u8981\u5730\u5411\u5927\u5bb6\u5206\u4eab\u6211\u4eec\u8fd1\u671f\u53d1\u8868\u5728 ICML 2024 \u4e0a\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u65b9\u5411\u7684\u7814\u7a76\u7ed3\u679c\u3002<\/p>\n

\u9884\u5907\u77e5\u8bc6<\/p>\n

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

\u65b9\u6cd5\u4ecb\u7ecd<\/p>\n

EOE\u65e8\u5728\u901a\u8fc7\u5229\u7528 LLM \u751f\u6210\u6f5c\u5728\u7684\u5f02\u5e38\u503c\u7c7b\u522b\u6807\u7b7e\u6765\u63d0\u9ad8\u96f6\u6837\u672c OOD \u68c0\u6d4b\u6027\u80fd\u3002\u7136\u800c\uff0c\u7531\u4e8e\u6a21\u578b\u90e8\u7f72\u65f6\u9047\u5230\u7684OOD\u7c7b\u522b\u662f\u672a\u77e5\u7684\uff0c\u90a3\u4e48\uff0c\u6211\u4eec\u5e94\u8be5\u5982\u4f55\u5f15\u5bfc LLM \u751f\u6210\u6240\u9700\u7684\u5f02\u5e38\u503c\u7c7b\u522b\u6807\u7b7e\uff1f\u5728\u83b7\u53d6\u5f02\u5e38\u503c\u7c7b\u522b\u6807\u7b7e\u540e\uff0c\u6211\u4eec\u5982\u4f55\u624d\u80fd\u66f4\u597d\u5730\u533a\u5206 ID \u548c OOD \u6837\u672c\uff1f\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u57fa\u4e8e\u89c6\u89c9\u76f8\u4f3c\u6027\u539f\u5219\u8bbe\u8ba1\u7684\u4e13\u95e8\u9488\u5bf9 OOD \u68c0\u6d4b\u7684 LLM \u63d0\u793a\uff0c\u5e76\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u8bc4\u5206\u51fd\u6570\u6765\u66f4\u597d\u5730\u533a\u5206ID\/OOD\u6837\u672c\u3002\u6211\u4eec\u65b9\u6cd5\u7684\u603b\u4f53\u6846\u67b6\u5982\u56fe 2\u6240\u793a\u3002<\/p>\n

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

\u56fe2. EOE\u603b\u4f53\u6846\u67b6\u56fe<\/p>\n

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

Fine-grained OOD \u68c0\u6d4b\u4e5f\u79f0\u4e3a\u5f00\u653e\u96c6\u8bc6\u522b\uff0c\u5728Fine-grained OOD \u68c0\u6d4b\u4e2d\uff0cID \u548c OOD \u6837\u672c\u90fd\u5c5e\u4e8e\u540c\u4e00\u4e3b\u8981\u7c7b\u522b(\u4f8b\u5982\u201c\u9e1f\u201d\u7c7b\uff09\uff0c\u5e76\u4e14\u5b50\u7c7b\u4e4b\u95f4\u5b58\u5728\u5185\u5728\u7684\u89c6\u89c9\u76f8\u4f3c\u6027 (\u4f8b\u5982\u201c\u9ebb\u96c0\u201d\u548c\u201c\u71d5\u5b50\u201d\uff09\u3002\u56e0\u6b64\uff0c\u6307\u793a LLM \u76f4\u63a5\u63d0\u4f9b\u540c\u4e00\u4e3b\u8981\u7c7b\u522b\u5185\u7684\u4e0d\u540c\u5b50\u7c7b\u66f4\u4e3a\u5408\u9002\u3002<\/p>\n

\u4ee5\u4e0a\u4e09\u79cdOOD \u68c0\u6d4b\u7684 LLM \u63d0\u793a\u5982\u56fe3\u6240\u793a<\/p>\n

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

\u56fe3. \u57fa\u4e8e\u89c6\u89c9\u76f8\u4f3c\u6027\u539f\u5219\u8bbe\u8ba1\u7684\u4e09\u7c7bLLM\u63d0\u793a<\/p>\n

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

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

\u56fe4. EOE\u4f2a\u4ee3\u7801<\/mark><\/p>\n

\u6211\u4eec\u7684\u65b9\u6cd5\u4f18\u70b9\u603b\u7ed3\u5982\u4e0b\uff1a<\/p>\n

EOE \u4e0d\u4f9d\u8d56\u4e8e\u672a\u77e5 OOD \u6570\u636e\u7684\u5148\u9a8c\u77e5\u8bc6<\/mark>\uff0c\u56e0\u6b64\u7279\u522b\u9002\u5408\u5f00\u653e\u4e16\u754c\u573a\u666f\u3002<\/p>\n

\u96f6\u6837\u672c\uff1a\u540c\u4e00\u4e2a\u9884\u8bad\u7ec3\u6a21\u578b\u53ef\u6709\u6548\u5730\u5e94\u7528\u4e8e\u5404\u79cd\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684 ID \u6570\u636e\u96c6\uff0c\u800c\u65e0\u9700\u5bf9\u6bcf\u4e2a\u7279\u5b9a ID \u6570\u636e\u96c6\u8fdb\u884c\u5355\u72ec\u8bad\u7ec3\u3002EOE \u4ec5\u901a\u8fc7\u4e86\u89e3 ID \u7c7b\u6807\u7b7e\u5373\u53ef\u5b9e\u73b0\u5353\u8d8a\u7684 OOD \u68c0\u6d4b\u6027\u80fd\u3002<\/p>\n

\u53ef\u6269\u5c55\u6027\u548c\u901a\u7528\u6027\uff1a\u4e0e\u540c\u6837\u751f\u6210\u6f5c\u5728OOD \u7c7b\u6807\u7b7e\u7684\u73b0\u6709\u96f6\u6837\u672c OOD \u68c0\u6d4b\u65b9\u6cd5 [6]\u76f8\u6bd4\uff0cEOE \u53ef\u4ee5\u8f7b\u677e\u5e94\u7528\u4e8e ImageNet-1K \u7b49\u5927\u89c4\u6a21\u6570\u636e\u96c6\u3002\u6b64\u5916\uff0cEOE \u5728\u4e0d\u540c\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u901a\u7528\u6027\uff0c\u5305\u62ecFar, Near\u548cFine-grainedOOD \u68c0\u6d4b\u3002<\/p>\n

\u5b9e\u9a8c\u7ed3\u679c<\/p>\n

\u6211\u4eec\u5728\u4e0d\u540cOOD\u4efb\u52a1\u7684\u591a\u4e2a\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\u3002\u88681\u5c55\u793a\u4e86\u5728ImageNet-1K \u4e0a\u8fdb\u884cFar OOD \u68c0\u6d4b\u5b9e\u9a8c\u7ed3\u679c\uff0c\u5176\u4e2d\uff0cGround Truth\u8868\u793a\u4f7f\u7528\u771f\u5b9eOOD\u6807\u7b7e\u65f6\u7684\u6027\u80fd\uff0c\u5728\u5b9e\u9645\u90e8\u7f72\u4e2d\u662f\u4e0d\u53ef\u83b7\u53d6\u7684\u3002\u7ed3\u679c\u8868\u660e\uff0cEOE \u4e0e\u5fae\u8c03\u65b9\u6cd5\u76f8\u5f53\uff0c\u5e76\u8d85\u8d8a\u4e86 MCM\u00a0[5]\u3002<\/p>\n

\u88681. Far OOD\u5b9e\u9a8c\u7ed3\u679c<\/p>\n

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

\u6211\u4eec\u4e5f\u62a5\u544a\u4e86\u5728Near OOD\u548cFine-grained OOD\u4efb\u52a1\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u5982\u88682\u548c\u88683\u6240\u793a\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u5747\u5b9e\u73b0\u4e86\u6700\u4f73\u7684\u68c0\u6d4b\u6027\u80fd\u3002<\/span><\/p>\n

\u88682.\u00a0Near OOD\u5b9e\u9a8c\u7ed3\u679c<\/p>\n

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

\u88683. Fine-grained OOD\u5b9e\u9a8c\u7ed3\u679c<\/p>\n

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

\u6211\u4eec\u5bf9EOE\u7684\u5404\u4e2a\u6a21\u578b\u8fdb\u884c\u4e86\u6d88\u878d\u5b9e\u9a8c\uff0c\u5305\u62ec\u4e0d\u540c\u7684\u5f97\u5206\u51fd\u6570\uff0cLLM\u63d0\u793a\uff1a\u4e0d\u540c\u7684LLM\u548c\u4e0d\u540c\u957f\u5ea6\u7684\u6f5c\u5728OOD \u7c7b\u6807\u7b7e\u6570\u91cf\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u6211\u4eec\u8bbe\u8ba1\u7684\u5f97\u5206\u51fd\u6570\u548c\u57fa\u4e8e\u89c6\u89c9\u76f8\u4f3c\u6027\u539f\u5219\u8bbe\u8ba1\u7684LLM prompt\u83b7\u5f97\u4e86\u6700\u4f18\u7684\u6027\u80fd\uff0c\u5e76\u4e14\u6211\u4eec\u7684\u65b9\u6cd5\u5728\u4e0d\u540c\u7684LLM\u548c\u4e0d\u540c\u957f\u5ea6\u7684\u6f5c\u5728OOD \u7c7b\u6807\u7b7e\u6570\u91cf\u4e0a\u90fd\u83b7\u5f97\u4e86\u4f18\u5f02\u7684\u8868\u73b0\u3002\u540c\u65f6\uff0c\u6211\u4eec\u4e5f\u5bf9\u89c6\u89c9\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u7ed3\u6784\u8fdb\u884c\u4e86\u6d88\u878d\u5b9e\u9a8c\uff0c\u8be6\u7ec6\u7684\u5b9e\u9a8c\u7ed3\u679c\u8bf7\u5927\u5bb6\u53c2\u8003\u539f\u6587\u3002<\/p>\n

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

\u56fe5.\u00a0\u6d88\u878d\u5b9e\u9a8c\u00a0\u2013\u4e0d\u540c\u7684\u5f97\u5206\u51fd\u6570\u3001LLM\u63d0\u793a\u548c\u4e0d\u540c\u7684LLM<\/p>\n

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

\u56fe5. \u6d88\u878d\u5b9e\u9a8c \u2013\u751f\u6210\u6f5c\u5728OOD \u7c7b\u6807\u7b7e\u6570\u91cf<\/p>\n

\u6211\u4eec\u5bf9EOE\u7684\u6709\u6548\u6027\u8fdb\u884c\u4e86\u5206\u6790\uff0c\u5b9e\u9645\u4e0a\uff0c\u751f\u6210\u7684\u5f02\u5e38\u7c7b\u6807\u7b7e\u4e0d\u592a\u53ef\u80fd\u6709\u5f88\u9ad8\u7684\u6982\u7387\u547d\u4e2d\u771f\u5b9e\u503c OOD \u7c7b\u3002\u8fd9\u662f\u56e0\u4e3a\u5728\u6a21\u578b\u7684\u5b9e\u9645\u90e8\u7f72\u4e2d\u9047\u5230\u7684 OOD \u6570\u636e\u662f\u591a\u79cd\u591a\u6837\uff0c\u4e14\u4e0d\u53ef\u9884\u6d4b\u7684\u3002\u7136\u800c\uff0c\u901a\u8fc7\u89c6\u89c9\u76f8\u4f3c\u6027\u89c4\u5219\u5f15\u5bfc\uff0c\u5373\u4f7f\u6ca1\u6709\u547d\u4e2d\u771f\u5b9e\u7684OOD\u7c7b\u7684\u60c5\u51b5\u4e0b\uff0cEOE\u751f\u6210\u7684\u6f5c\u5728\u5f02\u5e38\u7c7b\u6807\u7b7e\u4ecd\u7136\u53ef\u4ee5\u63d0\u9ad8\u6a21\u578b\u5728 OOD \u68c0\u6d4b\u4e2d\u7684\u8868\u73b0\u3002<\/p>\n

\u4e3a\u4e86\u8bf4\u660e\u4ee5\u4e0a\u8bba\u70b9\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u901a\u8fc7 T-SNE\u4ece\u6807\u7b7e\u5339\u914d\u5206\u6570\u7684 softmax \u8f93\u51fa\u4e2d\u5f97\u51fa\u7684\u53ef\u89c6\u5316\u6548\u679c\u3002\u6211\u4eec\u7684 EOE \u4e0e\u5bf9\u6bd4\u65b9\u6cd5 MCM \u4e4b\u95f4\u7684\u53ef\u89c6\u5316\u7ed3\u679c\u5982\u56fe6\u6240\u793a\u3002\u57fa\u4e8e ImageNet-10 \u7684 ID \u7c7b\u6807\u7b7e\uff0cLLM \u751f\u6210\u57fa\u4e8e\u89c6\u89c9\u76f8\u4f3c\u6027\u89c4\u5219\u7684\u6f5c\u5728\u5f02\u5e38\u6807\u7b7e\u201c\u6f5c\u8247\u201d\u3002\u5f53\u9047\u5230 OOD \u7c7b \u201c\u84b8\u6c7d\u673a\u8f66\u201d\uff08ImageNet-20 \u4e2d\u7684\u7c7b\uff09\u65f6\uff0c\u201c\u84b8\u6c7d\u673a\u8f66\u201d \u5728 \u548c \u4e2d\u4e0e\u201c\u6f5c\u8247\u201d\u7684\u76f8\u4f3c\u5ea6\u6700\u9ad8\u3002\u56e0\u6b64\uff0cEOE \u4f1a\u5c06\u5176\u805a\u7c7b<\/mark>\u4e3a \u201c\u6f5c\u8247\u201d\uff0c\u4ece\u800c\u5c06\u5176\u68c0\u6d4b\u4e3a OOD \u7c7b\u3002\u7136\u800c\uff0c\u5982\u679c\u6ca1\u6709\u6f5c\u5728\u7684\u5f02\u5e38\u7c7b\u6807\u7b7e\uff0c\u6211\u4eec\u53ef\u4ee5\u53d1\u73b0 MCM \u503e\u5411\u4e8e\u5c06\u6240\u6709 OOD \u7c7b\u6807\u7b7e\u805a\u7c7b<\/mark>\u5728\u4e00\u8d77\u3002\u8fd9\u53ef\u80fd\u5bfc\u81f4\u5c06\u96be\u4ee5\u8bc6\u522b\u7684 OOD \u6837\u672c\u8bc6\u522b\u4e3a ID \u7c7b\u3002\u603b\u4e4b\uff0c\u5728\u6211\u4eec\u7684 EOE \u6846\u67b6\u4e2d\uff0c1) \u5c5e\u4e8e\u540c\u4e00\u7c7b\u7684 OOD \u6837\u672c\u503e\u5411\u4e8e\u88ab\u805a\u96c6\u5728\u4e00\u8d77\uff0c2) \u540c\u4e00\u7ec4\u7684\u6837\u672c\u88ab\u5f52\u7c7b\u5230\u4e0e\u5b83\u4eec\u5728\u89c6\u89c9\u4e0a\u76f8\u4f3c\u7684\u8bbe\u60f3\u7684\u5f02\u5e38\u503c\u7c7b\u4e2d\uff08\u201c\u84b8\u6c7d\u673a\u8f66\u201d vs<\/em>\u201c\u6f5c\u8247\u201d\uff09\u3002\u8fd9\u4e9b\u89c2\u5bdf\u7ed3\u679c\u8868\u660e\uff0c\u6211\u4eec\u7684 EOE \u53ef\u4ee5\u5728\u4e0d\u89e6\u53ca\u5b9e\u9645 OOD \u7c7b\u522b\u7684\u60c5\u51b5\u4e0b\u589e\u5f3a OOD \u68c0\u6d4b\uff0c\u5e76\u4e14\u5728\u8bed\u4e49\u4e0a\u4e5f\u66f4\u6613\u4e8e\u89e3\u91ca\u3002\u6211\u4eec\u5e0c\u671b\u8fd9\u9879\u5de5\u4f5c\u80fd\u4e3a OOD \u68c0\u6d4b\u9886\u57df\u7684\u672a\u6765\u7814\u7a76\u63d0\u4f9b\u4e00\u4e2a\u65b0\u7684\u601d\u8def\u3002<\/p>\n

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

\u56fe6. \u53ef\u89c6\u5316\u7ed3\u679c<\/p>\n

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

\ufeff[1] Hendrycks, D. and Gimpel, K. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017.\u00a0<\/p>\n

[2]\u00a0\ufeffYang, J., Zhou, K., Li, Y., and Liu, Z. Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334, 2021.<\/p>\n

[3]\u00a0\ufeffLiu, W., Wang, X., Owens, J., and Li, Y. Energy-based out-of-distribution detection. In NeurIPS, 2020.<\/p>\n

[4]\u00a0\ufeffRadford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In ICML, 2021.<\/p>\n

[5]\u00a0\ufeffMing, Y., Cai, Z., Gu, J., Sun, Y., Li, W., and Li, Y. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022.<\/p>\n

[6] \ufeffEsmaeilpour, S., Liu, B., Robertson, E., and Shu, L. Zeroshot out-of-distribution detection based on the pre-trained model clip. In AAAI, 2022.<\/p>\n

\u8bfe\u9898\u7ec4\u4ecb\u7ecd<\/strong><\/p>\n

\u9999\u6e2f\u6d78\u4f1a\u5927\u5b66\u53ef\u4fe1\u673a\u5668\u5b66\u4e60<\/mark>\u548c\u63a8\u7406\u8bfe\u9898\u7ec4 (TMLR Group) \u7531\u591a\u540d\u9752\u5e74\u6559\u6388\u3001\u535a\u58eb\u540e\u7814\u7a76\u5458\u3001\u535a\u58eb\u751f\u3001\u8bbf\u95ee\u535a\u58eb\u751f\u548c\u7814\u7a76\u52a9\u7406\u5171\u540c\u7ec4\u6210\uff0c\u8bfe\u9898\u7ec4\u96b6\u5c5e\u4e8e\u7406\u5b66\u9662\u8ba1\u7b97\u673a\u7cfb\u3002\u8bfe\u9898\u7ec4\u4e13\u653b\u53ef\u4fe1\u8868\u5f81\u5b66\u4e60<\/mark>\u3001\u57fa\u4e8e\u56e0\u679c\u63a8\u7406<\/mark>\u7684\u53ef\u4fe1\u5b66\u4e60\u3001\u53ef\u4fe1\u57fa\u7840\u6a21\u578b\u7b49\u76f8\u5173\u7684\u7b97\u6cd5\uff0c\u7406\u8bba\u548c\u7cfb\u7edf\u8bbe\u8ba1\u4ee5\u53ca\u5728\u81ea\u7136\u79d1\u5b66\u4e0a\u7684\u5e94\u7528\uff0c\u5177\u4f53\u7814\u7a76\u65b9\u5411\u548c\u76f8\u5173\u6210\u679c\u8be6\u89c1\u672c\u7ec4Github (https:\/\/github.com\/tmlr-group)\u3002\u8bfe\u9898\u7ec4\u7531\u653f\u5e9c\u79d1\u7814\u57fa\u91d1\u4ee5\u53ca\u5de5\u4e1a\u754c\u79d1\u7814\u57fa\u91d1\u8d44\u52a9\uff0c\u5982\u9999\u6e2f\u7814\u7a76\u8d44\u52a9\u5c40\u6770\u51fa\u9752\u5e74\u5b66\u8005\u8ba1\u5212\uff0c\u56fd\u5bb6\u81ea\u7136\u79d1\u5b66\u57fa\u91d1\u9762\u4e0a\u9879\u76ee\u548c\u9752\u5e74\u9879\u76ee\uff0c\u4ee5\u53ca\u5fae\u8f6f\u3001\u82f1\u4f1f\u8fbe\u3001\u767e\u5ea6\u3001\u963f\u91cc\u3001\u817e\u8baf<\/mark>\u7b49\u4f01\u4e1a\u7684\u79d1\u7814\u57fa\u91d1\u3002\u9752\u5e74\u6559\u6388\u548c\u8d44\u6df1\u7814\u7a76\u5458\u624b\u628a\u624b\u5e26\uff0cGPU\u8ba1\u7b97\u8d44\u6e90\u5145\u8db3\uff0c\u957f\u671f\u62db\u6536\u591a\u540d\u535a\u58eb\u540e\u7814\u7a76\u5458\u3001\u535a\u58eb\u751f\u3001\u7814\u7a76\u52a9\u7406\u548c\u7814\u7a76\u5b9e\u4e60\u751f\u3002\u6b64\u5916\uff0c\u672c\u7ec4\u4e5f\u6b22\u8fce\u81ea\u8d39\u7684\u8bbf\u95ee\u535a\u58eb\u540e\u7814\u7a76\u5458\u3001\u535a\u58eb\u751f\u548c\u7814\u7a76\u52a9\u7406\u7533\u8bf7\uff0c\u8bbf\u95ee\u81f3\u5c113-6\u4e2a\u6708\uff0c\u652f\u6301\u8fdc\u7a0b\u8bbf\u95ee\u3002\u6709\u5174\u8da3\u7684\u540c\u5b66\u8bf7\u53d1\u9001\u4e2a\u4eba\u7b80\u5386\u548c\u521d\u6b65\u7814\u7a76\u8ba1\u5212\u5230\u90ae\u7bb1 (bhanml@comp.hkbu.edu.hk)\u3002<\/p>\n

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:ICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:ICML 2024| […]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[55,48],"class_list":["post-34307","post","type-post","status-publish","format-standard","hentry","category-news","tag-55","tag-48"],"yoast_head":"\nICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=34307\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34307\",\"url\":\"https:\/\/17aitech.com\/?p=34307\",\"name\":\"ICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34307#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34307#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8371b01a2c7b69302eb838e932f80588.png\",\"datePublished\":\"2024-12-03T22:56:51+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34307#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34307\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34307#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8371b01a2c7b69302eb838e932f80588.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8371b01a2c7b69302eb838e932f80588.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34307#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"ICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"ICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=34307","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34307","url":"https:\/\/17aitech.com\/?p=34307","name":"ICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34307#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34307#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8371b01a2c7b69302eb838e932f80588.png","datePublished":"2024-12-03T22:56:51+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34307#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34307"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34307#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8371b01a2c7b69302eb838e932f80588.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8371b01a2c7b69302eb838e932f80588.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34307#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"ICML 2024| \u5927\u8bed\u8a00\u6a21\u578b\u52a9\u529b\u57fa\u4e8eCLIP\u7684\u5206\u5e03\u5916\u68c0\u6d4b\u4efb\u52a1"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34307","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=34307"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34307\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34307"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34307"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34307"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}