{"id":870,"date":"2022-10-16T20:24:25","date_gmt":"2022-10-16T12:24:25","guid":{"rendered":"http:\/\/www.tamanegi.xyz\/?p=870"},"modified":"2023-08-29T19:50:05","modified_gmt":"2023-08-29T11:50:05","slug":"comex%e9%98%85%e8%af%bb%e6%8a%a5%e5%91%8a-2","status":"publish","type":"post","link":"http:\/\/tamanegi.xyz\/?p=870","title":{"rendered":"ComEx\u9605\u8bfb\u62a5\u544a"},"content":{"rendered":"<h1>ComEx<\/h1>\n<h2>\u6458\u8981<\/h2>\n<p><strong>\u65b0\u7c7b\u53d1\u73b0\u7684\u81ea\u52a8\u5316\u8bc6\u522b\uff0c\u9ad8\u6548\u7684GNCD\u65b9\u5f0f<\/strong><\/p>\n<\/p>\n<h2>Introduction<\/h2>\n<ol>\n<li>DL\u9700\u8981\u5927\u91cf\u6570\u636e\u96c6\uff0c\u5c1d\u8bd5\u4f7f\u7528NCD\u65b9\u5f0f\u89e3\u51b3SSL\u548cZSL\u7684\u5c40\u9650\u6027\u3002<\/li>\n<li>\u5f15\u7528\u7684\u8bba\u6587<strong>13<\/strong>\u63d0\u51fa\u4e86\u4e00\u4e2a\u65b9\u6cd5\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u4f46\u662f\u53ea\u80fd\u5728\u5355\u4e2a\u6570\u636e\u96c6\u4e0a\u53d6\u5f97\u8f83\u597d\u7684\u6210\u529f\uff0c\u65e0\u6cd5\u5b9e\u9645\u5e94\u7528\u3002<\/li>\n<li>\u672c\u6587\u5c1d\u8bd5\u5728GNCD\u4e0a\u63d0\u51fa\u7edf\u4e00\u6a21\u578b\u5e76\u53d6\u5f97\u8f83\u597d\u6570\u636e\u3002<\/li>\n<li>\u672c\u6a21\u578b\u4f7f\u7528<strong>Compositional Experts<\/strong>\u89e3\u51b3\u95ee\u9898\uff0c\u901a\u8fc7\u5728\u4f2a\u6807\u7b7e[^1]\u4e2d\u5f15\u5165 \u5c40\u90e8\u4e00\u81f4\u6027\u6765\u63d0\u5347\u65b0\u7c7b\u8bc6\u522b\u7684\u6027\u80fd\u3002<strong>\u663e\u8457\u63d0\u5347\u4e86<\/strong><\/li>\n<\/ol>\n<p>[^1]:\u6682\u65f6\u4e0d\u786e\u5b9a\u4f2a\u6807\u7b7e\uff08<em>pseudo labels<\/em>\uff09\u7684\u542b\u4e49\u3002<\/p>\n<h2>Related Work<\/h2>\n<h3>NCD<\/h3>\n<ul>\n<li>\u5229\u7528\u57fa\u7840\u6570\u636e\u96c6\u4e0a\u5b66\u4e60\u9884\u6d4b\u80fd\u529b\u6765\u4f30\u8ba1\u5e76\u5206\u7c7b\u65b0\u6570\u636e\u96c6[^20,21]<\/li>\n<li>\u5728\u57fa\u7840\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u5168\u76d1\u7763\u8bad\u7ec3\uff0c\u5728\u65b0\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u5fae\u8c03(<em>fine-tuning stage<\/em>)[^16]<\/li>\n<\/ul>\n<h3>Unsupervised Clustering<\/h3>\n<ul>\n<li>\u90bb\u57df\u805a\u5408\uff1a<em>data points within a neighborhood feature space likely share a same semantic label<\/em><\/li>\n<li>\u53ef\u4ee5\u907f\u514d<em>cluster collapse<\/em>\u7c07\u574d\u7f29(?)<\/li>\n<\/ul>\n<h2>Approach<\/h2>\n<h3>Definition<\/h3>\n<ul>\n<li>\n<p>\u5b9a\u4e49\uff1a\u57fa\u7840\u6570\u636e\u96c6 $latex D^b = {(x_i,y_i)}^{N^b}_{i=1}$,\u672a\u6807\u6ce8\u6570\u636e\u96c6 $latex C^b$\u3002\u8bbe\u7f6e\u6570\u636e\u96c6$latex D^n = {X_j}^{N^b+N^n}_{j={N^b+1}}$\uff0c\u5176\u4e2d\u5305\u542b\u6765\u81ea\u6570\u636e\u96c6$latex C^b$ \u7684\u6570\u636e$latex x_i$\u3002<\/p>\n<\/li>\n<li>\u76ee\u6807\u662f\u4ece\u56fe\u50cf\u7a7a\u95f4\u4e2d\u5b66\u4e60 $latex X={X_i}^{N^b+N^n}_{i=1}$\u5230\u6620\u5c04$latex Y={l}^{C^b+C^n}_{l=1}$\u4e2d<\/li>\n<li>\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u76ee\u6807\u90fd\u63d0\u53d6\u5176\u89c6\u89c9\u7279\u5f81\uff0c\u5e76\u5c06\u5176\u8f93\u5165\u6279\u5904\u7406\u548c\u7c7b\u5904\u7406\u4e2d\uff0c\u4f7f\u7528\u4ea4\u53c9\u71b5\u635f\u5931\uff08<em>the cross-entropy loss<\/em>[^3]\uff09\u8fdb\u884c\u8bad\u7ec3\u3002<\/li>\n<li>\u5bf9\u4e8e\u6765\u81ea\u57fa\u7840\u6570\u636e\u96c6\u7684\u6570\u636e\uff0c\u4f7f\u7528<em>ground-truth label<\/em>\u4f5c\u4e3a\u8bad\u7ec3\u76ee\u6807\uff1b\u5bf9\u4e8e\u6765\u81ea$D^n$\u6570\u636e\u96c6\u7684\u6570\u636e\uff0c\u4f7f\u7528\u4f2a\u6807\u8bb0\u751f\u6210\u76ee\u6807\u3002<\/li>\n<li>\u9075\u5faa\u539f\u5219 <strong><em>Known UnKnown<\/em><\/strong> \u6bcf\u4e2a\u4e13\u5bb6\u90fd\u5728\u5e94\u8be5\u77e5\u9053\u7684\u65f6\u5019\u77e5\u9053\uff08\uff1f\uff09<\/li>\n<li>\u4ea4\u53c9\u71b5\u635f\u5931\u5b9a\u4e49\u4e3a $L_{ce}(,y)=-ylogsigma(hat{y}^phi\/tau)$ [\u6ca1\u770b\u61c2\u8fd9\u5757\u662f\u5e72\u4ec0\u4e48\u7528\u7684]\uff1b\u7528\u4e86\u53e6\u4e00\u4e2a\u6b63\u5219\u8868\u8fbe\u5f0f\u6765\u663e\u5f0f\u6291\u5236\u975e\u76ee\u6807\u8f93\u51fa<\/li>\n<\/ul>\n<p>[^3]: Link to:<a href=\"https:\/\/blog.csdn.net\/xg123321123\/article\/details\/80781611\">(6\u6761\u6d88\u606f) \u7b80\u5355\u8c08\u8c08Cross Entropy Loss_\u65f6\u5149\u6742\u8d27\u5e97\u7684\u535a\u5ba2-CSDN\u535a\u5ba2_xentropy loss<\/a><\/p>\n<h3>Training Targets<\/h3>\n<p>\u6ca1\u7406\u89e3\uff0c\u8fd8\u5f97\u518d\u770b\u770b<\/p>\n<h3>Overall Objective<\/h3>\n<p>\u56db\u4e2a\u4e13\u5bb6\u90fd\u4ee5\u7aef\u5230\u7aef\u65b9\u5f0f\u8fdb\u884c\u8054\u5408\u8bad\u7ec3\u3002\u6574\u4f53\u635f\u5931\u5b9a\u4e49\u4e3a$latex mathscr{L}(x,y)=mathscr{L}+{ce}(hat{y}^phi,y)+mathscr{L}+{ce}(hat{y}^psi,y)+mathscr{L}_{reg}(hat{y}^phi)$,<\/p>\n<p>\u5c06\u56db\u4e2a\u4e13\u5bb6\u7684\u8f93\u51fa\u8fdb\u884c\u7ec4\u5408\uff0c\u524d\u4e24\u9879\u5bf9\u5e94\u6279\u5904\u7406\u4e13\u5bb6\u7684\u8d21\u732e\uff0c\u540e\u4e00\u9879\u5bf9\u5e94\u7c7b\u5904\u7406\u4e13\u5bb6\u7684\u8d21\u732e\u3002<\/p>\n<h2>Experiments<\/h2>\n<ul>\n<li>\u4f7f\u7528\u6570\u636e\u96c6\u4e3aCIFAR10\u3001CIFAR100\u3001ImageNet<\/li>\n<li>\n<p>\u6d4b\u8bd5\u65b9\u5f0f <\/p>\n<ol>\n<li><em>task-aware<\/em> :\u6570\u636e\u96c6\u5df2\u77e5\uff0c\u7528\u4e8e\u8bc4\u4f30\u5728\u65b0\u96c6\u4e0a\u6d4b\u8bd5\u5212\u5206\u7684\u6027\u80fd\u3002<\/li>\n<li><em>task-agnostic<\/em>:\u6570\u636e\u96c6\u672a\u77e5\uff0c\u540c\u65f6\u5728\u65b0\u96c6\u548c\u5df2\u77e5\u96c6\u4e0a\u6d4b\u8bd5\u5212\u5206\u6027\u80fd\u3002<\/li>\n<\/ol>\n<\/li>\n<li>Detail\uff1aResNet-18 \u4f5c\u4e3a\u56fe\u50cf\u7f16\u7801\u5668\uff0c\u6709\u6dfb\u52a0\u9002\u5ea6\u7684\u968f\u5373\u88c1\u526a\u3001\u53cd\u8f6c\u3001\u6296\u52a8\u548c\u7070\u5ea6\u3002<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u6d88\u878d\u5b9e\u9a8c\u8bc4\u4f30\u4e86\u6bcf\u4e00\u4e2a\u6a21\u5757\u7684\u4f5c\u7528\u548c\u5f71\u54cd\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>ComEx \u6458\u8981 \u65b0\u7c7b\u53d1\u73b0\u7684\u81ea\u52a8\u5316\u8bc6\u522b\uff0c\u9ad8\u6548\u7684GNCD\u65b9\u5f0f Introduction DL\u9700\u8981\u5927\u91cf\u6570\u636e\u96c6\uff0c\u5c1d\u8bd5 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105,11],"tags":[],"_links":{"self":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/870"}],"collection":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=870"}],"version-history":[{"count":4,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/870\/revisions"}],"predecessor-version":[{"id":936,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/870\/revisions\/936"}],"wp:attachment":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=870"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=870"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=870"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}