{"id":3094,"date":"2025-05-14T16:45:50","date_gmt":"2025-05-14T08:45:50","guid":{"rendered":"https:\/\/www.gnn.club\/?p=3094"},"modified":"2025-05-14T16:47:47","modified_gmt":"2025-05-14T08:47:47","slug":"uni-perceiver-%e5%9f%ba%e4%ba%8e%e5%8f%8c%e7%bc%96%e7%a0%81%e5%99%a8%e7%9a%84%e5%a4%9a%e6%a8%a1%e6%80%81%e7%bb%9f%e4%b8%80%e6%a1%86%e6%9e%b6","status":"publish","type":"post","link":"http:\/\/www.gnn.club\/?p=3094","title":{"rendered":"Uni-Perceiver\u7cfb\u5217: \u57fa\u4e8e\u53cc\u7f16\u7801\u5668\u7684\u591a\u6a21\u6001\u901a\u7528\u6a21\u578b"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">1.\u5f15\u7528<\/h2>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a&nbsp;<strong>Uni-Perceiver<\/strong>&nbsp;\u7684\u901a\u7528\u611f\u77e5\u67b6\u6784\uff0c\u65e8\u5728\u89e3\u51b3\u5f53\u524d\u673a\u5668\u5b66\u4e60\u4efb\u52a1\u7279\u5b9a\u8303\u5f0f\u7684\u5c40\u9650\u6027\uff0c\u5e76\u6a21\u4eff\u751f\u7269\u667a\u80fd\u7cfb\u7edf\u7684\u591a\u6a21\u6001\u6574\u5408\u4e0e\u5e76\u884c\u5904\u7406\u80fd\u529b\u3002\u5f53\u524d\u673a\u5668\u5b66\u4e60\u6a21\u578b\u901a\u5e38\u9488\u5bf9\u5355\u4e00\u4efb\u52a1\u72ec\u7acb\u8bbe\u8ba1\uff0c\u5bfc\u81f4\u4efb\u52a1\u95f4\u534f\u4f5c\u6548\u7387\u4f4e\u4e14\u5f00\u53d1\u65b0\u4efb\u52a1\u7684\u8fb9\u9645\u6210\u672c\u8f83\u9ad8\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u751f\u7269\u7cfb\u7edf\u80fd\u591f\u9ad8\u6548\u6574\u5408\u89c6\u89c9\u3001\u542c\u89c9\u7b49\u591a\u6a21\u6001\u4fe1\u606f\uff0c\u540c\u65f6\u5904\u7406\u591a\u79cd\u4efb\u52a1\u3002Uni-Perceiver \u901a\u8fc7\u7edf\u4e00\u5efa\u6a21\u548c\u5171\u4eab\u53c2\u6570\u5b9e\u73b0\u4e86\u591a\u6a21\u6001\uff08\u5982\u56fe\u50cf\u3001\u6587\u672c\u3001\u97f3\u9891\uff09\u548c\u591a\u4efb\u52a1\uff08\u5982\u5206\u7c7b\u3001\u68c0\u6d4b\u3001\u751f\u6210\uff09\u7684\u534f\u540c\u5904\u7406\u3002\u5176\u5bf9\u6bd4\u5982\u4e0b\u56fe1\u6240\u793a\uff1a<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"913\" height=\"902\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513162630876.png\" alt=\"\" class=\"wp-image-3097\" style=\"width:600px\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513162630876.png 913w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513162630876-300x296.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513162630876-768x759.png 768w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513162630876-100x100.png 100w\" sizes=\"auto, (max-width: 913px) 100vw, 913px\" \/><\/figure><\/div>\n\n\n<p>\u6a21\u578b\u5c42\u9762\uff0cUni-Perceiver \u7684\u6838\u5fc3\u8bbe\u8ba1\u5305\u62ec\u4e09\u4e2a\u90e8\u5206\uff1a\u4e00\u662f\u00a0<strong>\u6a21\u6001\u65e0\u5173\u7684Transformer\u7f16\u7801\u5668<\/strong>\uff0c\u5c06\u4e0d\u540c\u6a21\u6001\u7684\u8f93\u5165\u548c\u76ee\u6807\u6620\u5c04\u5230\u7edf\u4e00\u8868\u793a\u7a7a\u95f4\uff1b\u4e8c\u662f\u00a0<strong>\u8f7b\u91cf\u7ea7\u6a21\u6001\u7279\u5b9a\u5206\u8bcd\u5668<\/strong>\uff0c\u7528\u4e8e\u5c06\u539f\u59cb\u6570\u636e\u8f6c\u6362\u4e3a\u6a21\u578b\u53ef\u5904\u7406\u7684token\uff0c\u4ec5\u9700\u5c11\u91cf\u53c2\u6570\u9002\u914d\uff1b\u4e09\u662f\u00a0<strong>\u4efb\u52a1\u7edf\u4e00\u5efa\u6a21<\/strong>\uff0c\u6240\u6709\u4efb\u52a1\u5747\u88ab\u8f6c\u5316\u4e3a\u8f93\u5165\u4e0e\u76ee\u6807\u7684\u8868\u793a\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u95ee\u9898\u3002\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u6a21\u578b\u901a\u8fc7\u591a\u79cd\u5355\u6a21\u6001\u548c\u591a\u6a21\u6001\u4efb\u52a1\u5b66\u4e60\u901a\u7528\u8868\u793a\uff1b\u5728\u4e0b\u6e38\u4efb\u52a1\u4e2d\uff0c\u5b83\u652f\u6301\u96f6\u6837\u672c\u63a8\u7406\u3001\u5c11\u91cf\u6570\u636e\u63d0\u793a\u8c03\u4f18\u548c\u5168\u6570\u636e\u5fae\u8c03\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u672a\u89c1\u4efb\u52a1\u4e0a\u8868\u73b0\u5408\u7406\uff0c\u800c\u4ec5\u75281%\u4e0b\u6e38\u6570\u636e\u8c03\u4f18\u5373\u53ef\u63a5\u8fd1SOTA\u6027\u80fd\uff0c\u5168\u5fae\u8c03\u540e\u751a\u81f3\u80fd\u8d85\u8d8a\u73b0\u6709\u6700\u4f73\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e00\u7814\u7a76\u7684\u521b\u65b0\u70b9\u5728\u4e8e<strong>\u901a\u8fc7\u7edf\u4e00\u4efb\u52a1\u5efa\u6a21\u548c\u53c2\u6570\u5171\u4eab\uff0c\u663e\u8457\u964d\u4f4e\u4e86\u591a\u4efb\u52a1\u5f00\u53d1\u7684\u8fb9\u9645\u6210\u672c\uff0c\u540c\u65f6\u63d0\u5347\u4e86\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b<\/strong>\u3002Uni-Perceiver \u7684\u7075\u6d3b\u90e8\u7f72\u65b9\u5f0f\uff08\u96f6\u6837\u672c\u3001\u63d0\u793a\u8c03\u4f18\u3001\u5168\u5fae\u8c03\uff09\u4f7f\u5176\u9002\u7528\u4e8e\u4e0d\u540c\u8d44\u6e90\u573a\u666f\uff0c\u4e3a\u6784\u5efa\u901a\u7528AI\u7cfb\u7edf\u63d0\u4f9b\u4e86\u65b0\u601d\u8def\u3002\u672a\u6765\u7814\u7a76\u53ef\u8fdb\u4e00\u6b65\u63a2\u7d22\u6a21\u6001\u7edf\u4e00\u4e0e\u4efb\u52a1\u6027\u80fd\u7684\u5e73\u8861\uff0c\u4ee5\u53ca\u8ba1\u7b97\u6548\u7387\u7684\u4f18\u5316\u3002\u8bba\u6587\u4f5c\u8005\u8fd8\u5f00\u6e90\u4e86\u4ee3\u7801\u548c\u9884\u8bad\u7ec3\u6743\u91cd\uff0c\u63a8\u52a8\u4e86\u901a\u7528\u611f\u77e5\u6a21\u578b\u7684\u540e\u7eed\u53d1\u5c55\u3002\u603b\u4f53\u800c\u8a00\uff0c\u8fd9\u9879\u5de5\u4f5c\u5728\u591a\u6a21\u6001\u591a\u4efb\u52a1\u5b66\u4e60\u9886\u57df\u8fc8\u51fa\u4e86\u91cd\u8981\u4e00\u6b65\uff0c\u4e3a\u9ad8\u6548\u3001\u901a\u7528\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u91cd\u8981\u53c2\u8003\u3002<\/p>\n\n\n\n<p>\u6700\u540e\uff0cuni-perceiver\u662f\u4e00\u4e2a\u7cfb\u5217\u6a21\u578b\uff0c\u4e5f\u5305\u62ecuni-perceiver\uff0cuni-perceiverV2\uff0c\u672c\u535a\u6587\u90fd\u4f1a\u63d0\u53ca\u3002\u539f\u6587\u94fe\u63a5\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>uni-perceiver\u8bba\u6587\u5730\u5740\uff1a<a href=\"https:\/\/openaccess.thecvf.com\/content\/CVPR2022\/papers\/Zhu_Uni-Perceiver_Pre-Training_Unified_Architecture_for_Generic_Perception_for_Zero-Shot_and_CVPR_2022_paper.pdf\">https:\/\/openaccess.thecvf.com\/content\/CVPR2022\/papers\/Zhu_Uni-Perceiver_Pre-Training_Unified_Architecture_for_Generic_Perception_for_Zero-Shot_and_CVPR_2022_paper.pdf<\/a><\/li>\n\n\n\n<li>uni-perceiver-moe\u8bba\u6587\u5730\u5740\uff1a<a href=\"https:\/\/proceedings.neurips.cc\/paper_files\/paper\/2022\/file\/11fc8c98b46d4cbdfe8157267228f7d7-Paper-Conference.pdf\">https:\/\/proceedings.neurips.cc\/paper_files\/paper\/2022\/file\/11fc8c98b46d4cbdfe8157267228f7d7-Paper-Conference.pdf<\/a><\/li>\n\n\n\n<li>uni-perceiver-v2\u8bba\u6587\u5730\u5740\uff1a<a href=\"http:\/\/openaccess.thecvf.com\/content\/CVPR2023\/papers\/Li_Uni-Perceiver_v2_A_Generalist_Model_for_Large-Scale_Vision_and_Vision-Language_CVPR_2023_paper.pdf\">http:\/\/openaccess.thecvf.com\/content\/CVPR2023\/papers\/Li_Uni-Perceiver_v2_A_Generalist_Model_for_Large-Scale_Vision_and_Vision-Language_CVPR_2023_paper.pdf<\/a><\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">2. Uni-Perceiver \u6a21\u578b<\/h2>\n\n\n\n<p>\u4e0d\u540c\u6a21\u6001\u7684\u6570\u636e\uff0c\u5728\u9001\u8fdb\u6a21\u578b\u4e4b\u524d\u8981\u8fdb\u884c<strong>Token<\/strong>\u5316\uff0c\u5373Embedding\u3002<\/p>\n\n\n\n<p><strong>\u8f93\u5165Token\u5316<\/strong> \u5177\u4f53\u6765\u8bf4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6587\u672c<\/strong>\uff1a\u4f7f\u7528BPE\u5206\u8bcd\u5668\uff08\u5982BERT\u7b49\u6a21\u578b\u91c7\u7528\u7684\u65b9\u5f0f\uff09<\/li>\n\n\n\n<li><strong>\u56fe\u50cf<\/strong>\uff1a\u62c6\u5206\u6210\u56fe\u50cf\u5757\uff08\u7c7b\u4f3cViT\u7684patch\u5212\u5206\uff09<\/li>\n\n\n\n<li><strong>\u89c6\u9891<\/strong>\uff1a\u6309\u65f6\u95f4\u5e27\u5206\u5757\u5904\u7406<\/li>\n\n\n\n<li>\u6bcf\u79cdToken\u4f1a\u9644\u5e26<strong>\u6a21\u6001\u7c7b\u578b\u5d4c\u5165<\/strong>\uff08Modality Type Embedding\uff09\uff0c\u544a\u8bc9\u6a21\u578b\u5f53\u524d\u8f93\u5165\u662f\u6587\u672c\u3001\u56fe\u50cf\u8fd8\u662f\u89c6\u9891\u3002<\/li>\n\n\n\n<li>\u6bcf\u79cdToken\u4e5f\u4f1a\u9644\u5e26<strong>\u4f4d\u7f6e\u5d4c\u5165<\/strong>\uff08Modality Pos Embedding\uff09\uff0c\u5e2e\u52a9attention\u5efa\u6a21\u5355\u6a21\u6001\u81ea\u8eab\u7684\u7279\u5f81\u3002<\/li>\n\n\n\n<li>\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/li>\n<\/ul>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"954\" height=\"911\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513163912208.png\" alt=\"\" class=\"wp-image-3098\" style=\"width:600px\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513163912208.png 954w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513163912208-300x286.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513163912208-768x733.png 768w\" sizes=\"auto, (max-width: 954px) 100vw, 954px\" \/><\/figure><\/div>\n\n\n<p>\u6a21\u578b\u7684\u8f93\u5165\u5e8f\u5217\u53ef\u4ee5\u81ea\u7531\u7ec4\u5408\u4e0d\u540c\u6a21\u6001\u7684Token\uff0c\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u56fe\u6587\u4efb\u52a1<\/strong>&nbsp;\u2192&nbsp;<code>[&lt;SPE&gt;, \u56fe\u50cfToken, \u6587\u672cToken]<\/code><\/li>\n\n\n\n<li><strong>\u7eaf\u89c6\u9891\u4efb\u52a1<\/strong>&nbsp;\u2192&nbsp;<code>[&lt;SPE&gt;, \u89c6\u9891Token]<\/code><br>\u5176\u4e2d<code>&lt;SPE&gt;<\/code>\u662f\u4e00\u4e2a\u7279\u6b8a\u8d77\u59cb\u6807\u8bb0\uff0c\u5b83\u7684\u8f93\u51fa\u7279\u5f81\u4f1a\u4f5c\u4e3a\u6574\u4e2a\u8f93\u5165\u7684\u5168\u5c40\u8868\u793a\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6240\u6709\u6a21\u6001\u7684Token\u5e8f\u5217\u90fd\u7531<strong>\u540c\u4e00\u4e2aTransformer\u6a21\u578b<\/strong>\u5904\u7406\uff0c\u8f93\u51fa\u7edf\u4e00\u8bed\u4e49\u7a7a\u95f4\u4e2d\u7684\u7279\u5f81\u3002\u4e0d\u540c\u4efb\u52a1\uff08\u5982\u5206\u7c7b\u3001\u68c0\u7d22\uff09\u901a\u8fc7\u6bd4\u8f83\u8f93\u5165\u548c\u76ee\u6807\u7684\u7279\u5f81\u76f8\u4f3c\u5ea6\u6765\u5b9e\u73b0\uff08\u4f8b\u5982\uff1a\u8ba1\u7b97\u6700\u5927\u4f3c\u7136\u5f97\u5206\uff09\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1188\" height=\"512\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513164234681.png\" alt=\"\" class=\"wp-image-3099\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513164234681.png 1188w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513164234681-300x129.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513164234681-1024x441.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250513164234681-768x331.png 768w\" sizes=\"auto, (max-width: 1188px) 100vw, 1188px\" \/><\/figure>\n\n\n\n<p>UniPerceiver\u7684\u6838\u5fc3\u8bbe\u8ba1\u601d\u60f3\u2014\u2014<strong>\u7528\u540c\u4e00\u4e2a\u6a21\u578b\u5904\u7406\u4e0d\u540c\u4efb\u52a1\uff08\u5982\u5206\u7c7b\u3001\u68c0\u7d22\uff09\uff0c\u4e14\u652f\u6301\u6587\u672c\u3001\u56fe\u50cf\u3001\u89c6\u9891\u7b49\u591a\u79cd\u8f93\u5165\u5f62\u5f0f<\/strong>\u3002\u5176\u5173\u952e\u662f\u901a\u8fc7<strong>\u6700\u5927\u5316\u8f93\u5165\u4e0e\u76ee\u6807\u7684\u8054\u5408\u6982\u7387<\/strong>\u6765\u7edf\u4e00\u5efa\u6a21\uff0c\u5177\u4f53\u6765\u8bf4\uff0c\u65e0\u8bba\u4ec0\u4e48\u4efb\u52a1\uff08\u4f8b\u5982\u56fe\u50cf\u5206\u7c7b\u3001\u89c6\u9891\u68c0\u7d22\uff09\uff0c\u90fd\u53ef\u4ee5\u62bd\u8c61\u4e3a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8f93\u5165\uff08x\uff09\uff1a\u4e00\u5f20\u56fe\u3001\u4e00\u6bb5\u6587\u672c\u6216\u89c6\u9891<\/li>\n\n\n\n<li>\u5019\u9009\u76ee\u6807\uff08y\uff09\uff1a\u53ef\u80fd\u7684\u7b54\u6848\uff08\u5982\u7c7b\u522b\u6807\u7b7e\u3001\u63cf\u8ff0\u6587\u672c\uff09<\/li>\n<\/ul>\n\n\n\n<p>\u6a21\u578b\u7684\u76ee\u6807\u662f\u627e\u5230\u4e0e\u8f93\u5165\u6700\u5339\u914d\u7684 $\\mathbf{y}$ \uff0c\u516c\u5f0f\u8868\u793a\u4e3a\uff1a<\/p>\n\n\n\n<p>$$<br>\\hat{y}=\\arg \\max _{y \\in \\mathcal{Y}} P(x, y)<br>$$<\/p>\n\n\n\n<p>\u5176\u4e2d\uff0c $\\mathrm{P}(\\mathrm{x}, \\mathrm{y})$ \u8861\u91cf\u8f93\u5165 x \u548c\u76ee\u6807 y \u7684\u5339\u914d\u7a0b\u5ea6\uff0c\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u8ba1\u7b97\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7279\u5f81\u63d0\u53d6\uff1a\u7528\u5171\u4eab\u7684Transformer\u7f16\u7801\u5668\uff08 $f(\\cdot)$ \uff09\u5c06x\u548cy\u6620\u5c04\u5230\u540c\u4e00\u8bed\u4e49\u7a7a\u95f4\u3002<\/li>\n\n\n\n<li>\u76f8\u4f3c\u5ea6\u8ba1\u7b97\uff1a\u7528\u4f59\u5f26\u76f8\u4f3c\u5ea6\u6bd4\u8f83\u4e24\u8005\u7684\u7279\u5f81\uff0c\u5e76\u901a\u8fc7\u6e29\u5ea6\u53c2\u6570 $\\tau$\uff08\u53ef\u5b66\u4e60\uff09\u8c03\u6574\u5206\u5e03\u5e73\u6ed1\u5ea6\uff1a<\/li>\n<\/ul>\n\n\n\n<p>$$<br>P(x, y) \\propto \\exp (\\cos (f(x), f(y)) \/ \\tau)<br>$$<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8ba9\u6a21\u578b\u6cdb\u5316\u5230\u4e0d\u540c\u4efb\u52a1\uff0c\u4f5c\u8005\u7528\u591a\u79cd\u4efb\u52a1\u7684\u6570\u636e\u540c\u65f6\u8bad\u7ec3\uff08\u5982\u56fe\u6587\u5339\u914d\u3001\u89c6\u9891\u5206\u7c7b\u7b49\uff09\u3002\u9884\u8bad\u7ec3\u7684\u635f\u5931\u51fd\u6570\u8bbe\u8ba1\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p>$$<br>L=\\sum_{i=1}^n \\mathbb{E}{{x, y}}\\left[-\\log \\frac{P(x, y)}{\\sum{z \\in \\mathcal{Y}_i} P(x, z)}\\right]<br>$$<\/p>\n\n\n\n<p>\uff0d\u6838\u5fc3\u601d\u60f3\uff1a\u5bf9\u4e8e\u6bcf\u4e2a\u4efb\u52a1\u7684\u6570\u636e\uff0c\u8ba9\u771f\u5b9e\u914d\u5bf9\uff08 $\\mathrm{x}, \\mathrm{y}$ \uff09\u7684 $\\mathrm{P}(\\mathrm{x}, \\mathrm{y})$ \u8fdc\u9ad8\u4e8e\u5176\u4ed6\u5019\u9009\u76ee\u6807 z \uff08\u7c7b\u4f3c\u5bf9\u6bd4\u5b66\u4e60\uff09\u3002<br>\uff0d\u6570\u5b66\u542b\u4e49\uff1a\u6700\u5c0f\u5316\u8be5\u635f\u5931\uff0c\u76f8\u5f53\u4e8e\u8981\u6c42\u6a21\u578b\u5bf9\u6b63\u6837\u672c\u7684\u5339\u914d\u6982\u7387\u5c3d\u53ef\u80fd\u5927\uff0c\u8d1f\u6837\u672c\u5c3d\u53ef\u80fd\u5c0f\u3002<\/p>\n\n\n\n<p><mark style=\"background-color:rgba(0, 0, 0, 0)\" class=\"has-inline-color has-luminous-vivid-orange-color\"><strong>\u4e2a\u4eba\u8ba4\u4e3a\uff0c<\/strong><\/mark>Uni-perceiver\u5c31\u662f\u5728CLIP\u5de5\u4f5c\u7684\u57fa\u7840\u4e4b\u4e0a\u518d\u8fdb\u4e86\u4e00\u6b65\uff0c<strong>\u8d85\u8d8aCLIP\u7684\u5730\u65b9<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6a21\u6001\u65e0\u5173<\/strong>\uff1a\u4e0d\u5c40\u9650\u4e8e\u56fe\u6587\uff0c\u652f\u6301\u4efb\u610f\u7ec4\u5408\u3002<\/li>\n\n\n\n<li><strong>\u4efb\u52a1\u65e0\u5173<\/strong>\uff1a\u5206\u7c7b\u3001\u68c0\u7d22\u3001\u751f\u6210\u7b49\u4efb\u52a1\u7edf\u4e00\u5efa\u6a21\u3002<\/li>\n\n\n\n<li><strong>\u76ee\u6807\u81ea\u7531<\/strong>\uff1ay\u53ef\u4ee5\u662f\u4efb\u4f55\u5f62\u5f0f\uff0c\u53ea\u8981\u80fd\u88abToken\u5316\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5982\u679cCLIP\u662f\u201c\u4e13\u624d\u201d\uff08\u64c5\u957f\u56fe\u6587\u5bf9\u6bd4\uff09\uff0cUniPerceiver\u5c31\u662f\u201c\u901a\u624d\u201d\uff08\u4ec0\u4e48\u4efb\u52a1\u90fd\u80fd\u63d2\u4e00\u624b\uff09<\/p>\n\n\n\n<p><mark style=\"background-color:rgba(0, 0, 0, 0)\" class=\"has-inline-color has-luminous-vivid-orange-color\">\u9664\u6b64\u4e4b\u5916\uff0cCLIP\u7684\u53cc\u5854\u7ed3\u6784\u53c2\u6570\u4e0d\u5171\u4eab\uff0c\u800cUni-perceiver\u7684\u53cc\u5854\u7ed3\u6784\u5219\u662f\u53c2\u6570\u5171\u4eab\u7684\u3002\u8fd9\u6837\u505a\u6709\u4ee5\u4e0b\u51e0\u4e2a\u597d\u5904\uff1a<\/mark><\/p>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u53c2\u6570\u6548\u7387\u66f4\u9ad8<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>CLIP<\/strong>\uff1a\u56fe\u50cf\u548c\u6587\u672c\u7f16\u7801\u5668\u5b8c\u5168\u5206\u79bb\uff0c\u53c2\u6570\u91cf\u968f\u6a21\u6001\u589e\u52a0\u7ebf\u6027\u589e\u957f\uff08\u4f8b\u5982\u65b0\u589e\u97f3\u9891\u6a21\u6001\u9700\u989d\u5916\u8bad\u7ec3\u4e00\u4e2a\u7f16\u7801\u5668\uff09\u3002<\/li>\n\n\n\n<li><strong>UniPerceiver<\/strong>\uff1a<strong>\u6240\u6709\u6a21\u6001\u5171\u7528\u540c\u4e00\u5957\u53c2\u6570<\/strong>\uff0c\u65b0\u589e\u6a21\u6001\u53ea\u9700\u6269\u5c55Tokenizer\uff0c\u4e3b\u5e72\u6a21\u578b\u4fdd\u6301\u4e0d\u53d8\u3002<br>\u2192 \u66f4\u8282\u7701\u663e\u5b58\u548c\u8bad\u7ec3\u8d44\u6e90\uff0c\u5c24\u5176\u9002\u5408\u591a\u6a21\u6001\u6269\u5c55\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u6a21\u6001\u4ea4\u4e92\u66f4\u5f7b\u5e95<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>CLIP<\/strong>\uff1a\u56fe\u50cf\u548c\u6587\u672c\u4ec5\u5728\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u65f6\u4ea4\u4e92\uff0c\u524d\u671f\u7f16\u7801\u8fc7\u7a0b\u5b8c\u5168\u9694\u79bb\uff0c\u53ef\u80fd\u4e22\u5931\u7ec6\u7c92\u5ea6\u5bf9\u9f50\u4fe1\u606f\u3002<\/li>\n\n\n\n<li><strong>UniPerceiver<\/strong>\uff1a<strong>\u8f93\u5165\u5e8f\u5217\u76f4\u63a5\u6df7\u5408\u591a\u6a21\u6001Token<\/strong>\uff08\u5982<code>[\u56fe\u50cfToken][\u6587\u672cToken]<\/code>\uff09\uff0c\u5728Transformer\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e2d\u5b9e\u65f6\u4ea4\u4e92\u3002<br>\u2192 \u66f4\u9002\u5408\u9700\u8981\u6df1\u5ea6\u878d\u5408\u7684\u4efb\u52a1\uff08\u4f8b\u5982\u89c6\u9891\u95ee\u7b54\u9700\u540c\u65f6\u7406\u89e3\u753b\u9762\u548c\u5b57\u5e55\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u907f\u514d\u6a21\u6001\u504f\u5dee<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>CLIP\u7684\u53cc\u5854\u95ee\u9898<\/strong>\uff1a\u72ec\u7acb\u7f16\u7801\u5668\u53ef\u80fd\u5bfc\u81f4\u56fe\u50cf\u548c\u6587\u672c\u7279\u5f81\u5206\u5e03\u4e0d\u4e00\u81f4\uff08\u4f8b\u5982\u6587\u672c\u7279\u5f81\u66f4\u7a00\u758f\uff09\uff0c\u9700\u989d\u5916\u8bbe\u8ba1\u635f\u5931\u51fd\u6570\u5bf9\u9f50\u3002<\/li>\n\n\n\n<li><strong>UniPerceiver\u7684\u89e3\u51b3\u65b9\u6848<\/strong>\uff1a\u5f3a\u5236\u6240\u6709\u6a21\u6001\u901a\u8fc7\u540c\u4e00\u7ec4\u53c2\u6570\u6620\u5c04\u5230<strong>\u5b8c\u5168\u5171\u4eab\u7684\u8bed\u4e49\u7a7a\u95f4<\/strong>\uff0c\u5929\u7136\u5bf9\u9f50\u7279\u5f81\u5206\u5e03\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u7075\u6d3b\u5e94\u5bf9\u975e\u5e38\u89c4\u4efb\u52a1<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>CLIP\u7684\u5c40\u9650<\/strong>\uff1a\u96be\u4ee5\u5904\u7406\u76ee\u6807\u4e0d\u662f\u6587\u672c\u7684\u4efb\u52a1\uff08\u4f8b\u5982\u56fe\u50cf\u2192\u56fe\u50cf\u68c0\u7d22\uff09\u3002<\/li>\n\n\n\n<li><strong>UniPerceiver\u7684\u901a\u7528\u6027<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u76ee\u6807\u53ef\u4ee5\u662f\u4efb\u610f\u5f62\u5f0f\uff08\u5982<code>y<\/code>=\u56fe\u50cf\u3001\u89c6\u9891\u3001\u7ed3\u6784\u5316\u6570\u636e\uff09\u3002<\/li>\n\n\n\n<li>\u751a\u81f3\u652f\u6301<strong>\u591a\u6a21\u6001\u8054\u5408\u76ee\u6807<\/strong>\uff08\u4f8b\u5982\u8f93\u5165\u533b\u5b66\u5f71\u50cf\uff0c\u8f93\u51fa<code>[\u8bca\u65ad\u6587\u672c][\u75c5\u53d8\u533a\u57df\u56fe\u50cf]<\/code>\u7684\u7ec4\u5408\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u5f53\u7136\uff0c\u8fd9\u79cd\u5f62\u5f0f\u4e5f\u5b58\u5728\u5b83\u7684\u7f3a\u70b9\uff0c\u6211\u4e2a\u4eba\u8ba4\u4e3a\uff0c\u56e0\u4e3a\u591a\u4efb\u52a1\u4e4b\u95f4\u7684\u504f\u5dee\u548c\u6570\u636e\u5dee\u5f02\uff0c\u8fd9\u79cd\u5f62\u5f0f\u5927\u6982\u7387\u5b58\u5728\u5de8\u5927\u7684\u8bad\u7ec3\u96be\u5ea6\u3002\u5173\u4e8e\u8bad\u7ec3\u7684\u7ec6\u8282\uff0c\u6211\u5728\u8bba\u6587\u4e2d\u5e76\u672a\u53d1\u73b0\u592a\u591a\u7684\u7ec6\u8282\uff0c\u53ea\u6709\u4e0b\u8ff0\u4e24\u53e5\u7b80\u5355\u7684\u63cf\u8ff0\u3002<\/p>\n\n\n\n<p class=\"has-blush-light-purple-gradient-background has-background\" style=\"font-style:italic;font-weight:500;text-decoration:underline\"><kbd>We pre-train the model with multiple tasks simultaneously. In each iteration, each GPU independently samples a single task and dataset. The gradients of different GPUs are synchronized after the gradient back-propagation.<\/kbd><\/p>\n\n\n\n<h2 class=\"wp-block-heading\">3. Uni-Perceiver-MoE \u6a21\u578b<\/h2>\n\n\n\n<p>Uni-Perceiver-MoE\u7684\u4e3b\u8981\u7684\u8d21\u732e\u5c31\u662f\u52a0\u5165\u4e86\u6df7\u5408\u4e13\u5bb6\u673a\u5236MoE\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u662f\u00a0<strong>\u6761\u4ef6\u6df7\u5408\u4e13\u5bb6\uff08Conditional Mixture-of-Experts, Conditional MoEs\uff09<\/strong>\u00a0\u673a\u5236\uff0c\u901a\u8fc7\u52a8\u6001\u8def\u7531\u7b56\u7565\uff0c\u5b9e\u73b0\u4e86\u591a\u6a21\u6001\u3001\u591a\u4efb\u52a1\u573a\u666f\u4e0b\u7684\u9ad8\u6548\u8ba1\u7b97\u4e0e\u4efb\u52a1\u89e3\u8026\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1194\" height=\"570\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514150211383.png\" alt=\"\" class=\"wp-image-3111\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514150211383.png 1194w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514150211383-300x143.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514150211383-1024x489.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514150211383-768x367.png 768w\" sizes=\"auto, (max-width: 1194px) 100vw, 1194px\" \/><\/figure>\n\n\n\n<p><strong>1\uff0eConditional MoEs \u7684\u6838\u5fc3\u673a\u5236<\/strong><br><strong>1.1\u57fa\u672c\u539f\u578b<\/strong><br>\uff0d\u8f93\u5165\u4e0e\u8def\u7531\uff1a<br>\u5bf9\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e2a token $x_i$ \uff0c\u901a\u8fc7\u8def\u7531\u51fd\u6570 $R(\\cdot)$ \u751f\u6210\u95e8\u63a7\u51b3\u7b56\u5411\u91cf $\\mathcal{G} \\in \\mathbb{R}^E(E$\u4e3a\u4e13\u5bb6\u6570\u91cf\uff09\uff0c\u9009\u62e9 top\uff0d$k$ \u4e2a\u4e13\u5bb6\u6fc0\u6d3b\uff1a<\/p>\n\n\n\n<p>$$<br>\\mathcal{G}=\\operatorname{top}_k\\left(\\operatorname{softmax}\\left(\\mathbf{W}_g \\cdot R\\left(x_i\\right)+\\epsilon\\right)\\right)<br>$$<\/p>\n\n\n\n<p>\uff0d $\\mathbf{W}_g$ \uff1a\u53ef\u8bad\u7ec3\u7684\u95e8\u63a7\u6743\u91cd\u77e9\u9635\u3002<br>$\\circ \\epsilon$ \uff1a\u566a\u58f0\u9879\uff08\u53ef\u80fd\u7528\u4e8e\u589e\u5f3a\u9c81\u68d2\u6027\uff09\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7a00\u758f\u6027\uff1a\u4ec5 $k \\ll E$ \u4e2a\u4e13\u5bb6\u88ab\u6fc0\u6d3b\uff0c\u663e\u8457\u51cf\u5c11\u8ba1\u7b97\u91cf\u3002<\/li>\n\n\n\n<li>\u8f93\u51fa\u8ba1\u7b97\uff1a<\/li>\n<\/ul>\n\n\n\n<p>\u88ab\u9009\u4e2d\u7684\u4e13\u5bb6\u5bf9 $x_i$ \u7684\u7ebf\u6027\u53d8\u6362\u7ed3\u679c\u52a0\u6743\u6c42\u548c\uff1a<\/p>\n\n\n\n<p>$$<br>y_i=\\sum_{e=1}^E \\mathcal{G}_e \\cdot \\mathbf{W}_e \\cdot x_i<br>$$<\/p>\n\n\n\n<p>\uff0d\u672a\u88ab\u9009\u4e2d\u7684\u4e13\u5bb6\uff08 $\\mathcal{G}_e=0$ \uff09\u65e0\u9700\u8ba1\u7b97\u3002<br><strong>1.2 \u8def\u7531\u7b56\u7565 $R(\\cdot)$ \u7684\u4f5c\u7528<\/strong><br>\u8def\u7531\u7b56\u7565\u662f Conditional MoEs \u7684\u5173\u952e\uff0c\u51b3\u5b9a\u4e86\u5982\u4f55\u6839\u636e\u8f93\u5165\u6761\u4ef6\uff08\u5982\u6a21\u6001\u3001\u4efb\u52a1\u3001\u4e0a\u4e0b\u6587\uff09\u5206\u914d\u4e13\u5bb6\u3002\u5176\u8bbe\u8ba1\u76f4\u63a5\u5f71\u54cd\u6a21\u578b\u7684\u591a\u4efb\u52a1\u6cdb\u5316\u80fd\u529b\u548c\u6a21\u6001\u9002\u5e94\u6027\u3002<\/p>\n\n\n\n<p><strong>2\uff0e\u8def\u7531\u7b56\u7565\u7684\u53d8\u4f53<\/strong><br><strong>2.1 Token\uff0dLevel Routing\uff08token\u7ea7\u8def\u7531\uff09<\/strong><br>\uff0d\u7b56\u7565\uff1a\u76f4\u63a5\u4f7f\u7528 token \u81ea\u8eab\u7684\u8868\u793a\u4f5c\u4e3a\u8def\u7531\u4f9d\u636e\uff1a<\/p>\n\n\n\n<p>$$<br>R_{\\text {token }}\\left(x_i\\right)=x_i<br>$$<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7279\u70b9\uff1a<\/li>\n\n\n\n<li>\u7c7b\u4f3c\u4f20\u7edf MoE\uff08\u5982Switch Transformer\uff09\uff0c\u4ec5\u4f9d\u8d56\u5c40\u90e8\u4fe1\u606f\u3002<\/li>\n\n\n\n<li>\u9002\u7528\u4e8e\u540c\u8d28\u5316\u8f93\u5165\uff08\u5982\u5355\u4e00\u6a21\u6001\u6587\u672c\uff09\uff0c\u4f46\u5bf9\u591a\u4efb\u52a1\uff0f\u591a\u6a21\u6001\u53ef\u80fd\u56e0\u4efb\u52a1\u51b2\u7a81\u5bfc\u81f4\u5e72\u6270\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>2.2 Context\uff0dLevel Routing\uff08\u4e0a\u4e0b\u6587\u7ea7\u8def\u7531\uff09<\/strong><br>\uff0d\u7b56\u7565\uff1a\u7ed3\u5408 token \u8868\u793a\u4e0e\u5168\u5c40\u4e0a\u4e0b\u6587\uff08\u901a\u8fc7\u6ce8\u610f\u529b\u6c60\u5316\uff09\uff1a<\/p>\n\n\n\n<p>$$<br>R_{\\text {context }}\\left(x_i\\right)=\\operatorname{concat}\\left(x_i, \\operatorname{attnpool}(X)\\right)<br>$$<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7279\u70b9\uff1a<\/li>\n\n\n\n<li>\u5f15\u5165\u5e8f\u5217\u5168\u5c40\u4fe1\u606f\uff0c\u7f13\u89e3\u76f8\u4f3c token \u5728\u4e0d\u540c\u4efb\u52a1\u4e2d\u7684\u51b2\u7a81\u3002<\/li>\n\n\n\n<li>\u9002\u5408\u9700\u8981\u957f\u8ddd\u79bb\u4f9d\u8d56\u7684\u4efb\u52a1\uff08\u5982\u89c6\u9891\u7406\u89e3\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>2.3 Modality\uff0dLevel Routing\uff08\u6a21\u6001\u7ea7\u8def\u7531\uff09<\/strong><br>\uff0d\u7b56\u7565\uff1a\u6839\u636e token \u7684\u6a21\u6001\u7c7b\u578b\uff08\u5982\u6587\u672c\u3001\u56fe\u50cf\uff09\u5206\u914d\u4e13\u5bb6\uff1a<\/p>\n\n\n\n<p>$$<br>R_{\\text {modal }}\\left(x_i\\right)=\\operatorname{embed}\\left(\\operatorname{id}_{\\text {modal }}\\left(x_i\\right)\\right)<br>$$<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7279\u70b9\uff1a<\/li>\n\n\n\n<li>\u663e\u5f0f\u533a\u5206\u6a21\u6001\uff0c\u7c7b\u4f3c\u4f20\u7edf\u591a\u6a21\u6001\u6a21\u578b\u7684\u72ec\u7acb\u7f16\u7801\u5668\u3002<\/li>\n\n\n\n<li>\u4f46\u53ef\u80fd\u524a\u5f31\u8de8\u6a21\u6001\u4ea4\u4e92\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>2.4 Task\uff0dLevel Routing\uff08\u4efb\u52a1\u7ea7\u8def\u7531\uff09<\/strong><br>\uff0d\u7b56\u7565\uff1a\u57fa\u4e8e\u4efb\u52a1 ID \u5206\u914d\u4e13\u5bb6\uff1a<\/p>\n\n\n\n<p>$$<br>R_{\\text {task }}\\left(x_i\\right)=\\operatorname{embed}\\left(\\operatorname{id}_{\\text {task }}\\left(x_i\\right)\\right)<br>$$<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7279\u70b9\uff1a<\/li>\n\n\n\n<li>\u540c\u4e00\u4efb\u52a1\u7684\u6240\u6709 token \u5171\u4eab\u4e13\u5bb6\uff0c\u5f7b\u5e95\u907f\u514d\u4efb\u52a1\u5e72\u6270\u3002<\/li>\n\n\n\n<li>\u5c40\u9650\u6027\uff1a\u65e0\u6cd5\u6cdb\u5316\u5230\u65b0\u4efb\u52a1\uff08\u9700\u9884\u5b9a\u4e49\u4efb\u52a1 ID\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>2.5 Attribute Routing\uff08\u5c5e\u6027\u8def\u7531\uff09<\/strong><br>\uff0d\u7b56\u7565\uff1a\u7efc\u5408 token \u7684\u591a\u79cd\u5c5e\u6027\uff08\u6a21\u6001\u3001\u4efb\u52a1\u7c7b\u578b\u3001\u56e0\u679c\u6027\u7b49\uff09\u751f\u6210\u8def\u7531\uff1a<\/p>\n\n\n\n<p>$$<br>R_{\\mathrm{attr}}\\left(x_i\\right)=\\text { layernorm }\\left(\\mathbf{W}_{\\mathrm{attr}} \\cdot \\operatorname{attr}\\left(x_i\\right)\\right)<br>$$<\/p>\n\n\n\n<p>\u5c5e\u6027\u4e3a 8 \u7ef4\u4e8c\u8fdb\u5236\u5d4c\u5165\uff08\u5982\u6a21\u6001\u7d22\u5f15\u3001\u4efb\u52a1\u7c7b\u578b\u7b49\uff09\uff0c\u8be6\u89c1\u4e0b\u8868\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"652\" height=\"274\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514151007686.png\" alt=\"\" class=\"wp-image-3112\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514151007686.png 652w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514151007686-300x126.png 300w\" sizes=\"auto, (max-width: 652px) 100vw, 652px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7279\u70b9\uff1a<\/li>\n\n\n\n<li>\u5e73\u8861\u4efb\u52a1\u89e3\u8026\u4e0e\u6cdb\u5316\u80fd\u529b\uff1a\u65e2\u5305\u542b\u4efb\u52a1\u4fe1\u606f\uff0c\u53c8\u907f\u514d\u4f9d\u8d56\u56fa\u5b9a ID\u3002<\/li>\n\n\n\n<li>\u9002\u5408\u901a\u7528\u6a21\u578b\uff08Generalist Models\uff09\uff0c\u5982 Uni\uff0dPerceiver \u7684\u591a\u4efb\u52a1\u573a\u666f\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4e0a\u8ff0MoE\u7684\u4e0d\u540c\u8bbe\u8ba1\u53ef\u4ee5\u4ece\u662f\u5426\u4f9d\u8d56\u6570\u636e\u8fdb\u884c\u5206\u6790\uff0c\u5177\u4f53\u6765\u8bf4\uff1a<\/p>\n\n\n\n<p><strong>\u6570\u636e\u4f9d\u8d56\uff08Data-Dependent\uff09<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Token-Level &amp; Context-Level MoEs<\/strong>\uff1a\u8def\u7531\u51b3\u7b56\u4f9d\u8d56\u8f93\u5165\u6570\u636e\u7684\u5177\u4f53\u5185\u5bb9\uff08\u5982 token \u8868\u793a\u6216\u4e0a\u4e0b\u6587\uff09\u3002<\/li>\n\n\n\n<li><strong>\u4f18\u70b9\uff1a<\/strong>\u4e0d\u635f\u5bb3\u6a21\u578b\u5bf9\u65b0\u4efb\u52a1\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n\n\n\n<li><strong>\u7f3a\u70b9<\/strong>\uff1a\u8bad\u7ec3\u548c\u63a8\u7406\u65f6\u5185\u5b58\u6d88\u8017\u9ad8\uff08\u9700\u52a0\u8f7d\u6240\u6709\u4e13\u5bb6\uff09\uff0c\u9700\u901a\u8fc7\u6a21\u578b\u5e76\u884c\u7f13\u89e3\uff0c\u4f46\u4f1a\u5f15\u5165\u8bbe\u5907\u95f4\u901a\u4fe1\u5f00\u9500\u3002\u6b64\u5916\uff0c\u5ffd\u7565\u4efb\u52a1\u4fe1\u606f\uff0c\u65e0\u6cd5\u89e3\u51b3\u4efb\u52a1\u95f4\u5e72\u6270\uff08Task Interference\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>\u6570\u636e\u72ec\u7acb\uff08Data-Independent\uff09<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Modality-Level\/Task-Level\/Attribute MoEs<\/strong>\uff1a\u8def\u7531\u51b3\u7b56\u4f9d\u8d56\u9884\u5b9a\u4e49\u7684\u6a21\u6001\/\u4efb\u52a1\/\u5c5e\u6027\u6807\u7b7e\uff0c\u4e0e\u8f93\u5165\u5185\u5bb9\u65e0\u5173\u3002<\/li>\n\n\n\n<li><strong>\u4f18\u70b9<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li><strong>\u5185\u5b58\u9ad8\u6548<\/strong>\uff1a\u53ea\u9700\u6fc0\u6d3b\u4e0e\u5f53\u524d\u6a21\u6001\/\u4efb\u52a1\u76f8\u5173\u7684 top-<em>k<\/em>\u00a0\u4e13\u5bb6\u3002<\/li>\n\n\n\n<li><strong>\u8ba1\u7b97\u4f18\u5316<\/strong>\uff1a\u53ef\u901a\u8fc7\u53c2\u6570\u91cd\u53c2\u6570\u5316\uff08reparameterization\uff09\u5408\u5e76\u4e13\u5bb6\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u7f3a\u70b9\uff1aModality-Level \u548c Task-Level<\/strong>\u4f9d\u8d56\u9884\u5b9a\u4e49\u4efb\u52a1 ID\uff0c\u96be\u4ee5\u6cdb\u5316\u5230\u65b0\u4efb\u52a1\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1286\" height=\"634\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514151055565.png\" alt=\"\" class=\"wp-image-3113\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514151055565.png 1286w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514151055565-300x148.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514151055565-1024x505.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514151055565-768x379.png 768w\" sizes=\"auto, (max-width: 1286px) 100vw, 1286px\" \/><\/figure>\n\n\n\n<p><strong>\u5c5e\u6027\u8def\u7531\uff08Attribute MoEs\uff09<\/strong>\uff1a\u901a\u8fc7\u9884\u5b9a\u4e49\u7684\u5c5e\u6027\uff08\u5982\u6a21\u6001\u3001\u4efb\u52a1\u7c7b\u578b\u7b49\uff09\u7f16\u7801\u4efb\u52a1\u4fe1\u606f\uff0c\u65e2\u7f13\u89e3\u4efb\u52a1\u5e72\u6270\uff0c\u53c8\u652f\u6301\u65b0\u4efb\u52a1\u8fc1\u79fb\uff08\u65e0\u9700\u4fee\u6539\u6a21\u578b\u7ed3\u6784\uff09\uff0c\u662f\u5e73\u8861\u6548\u7387\u4e0e\u6cdb\u5316\u7684\u6700\u4f73\u9009\u62e9\u3002\u6587\u4e2d\u4e5f\u5c55\u793a\u4e86\u4e0d\u540c\u8def\u7531\u65b9\u5f0f\u7684\u6d88\u878d\u5b9e\u9a8c\uff0c\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1337\" height=\"536\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514152200936.png\" alt=\"\" class=\"wp-image-3114\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514152200936.png 1337w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514152200936-300x120.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514152200936-1024x411.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514152200936-768x308.png 768w\" sizes=\"auto, (max-width: 1337px) 100vw, 1337px\" \/><\/figure>\n\n\n\n<p>\u4e0d\u7ba1\u662f\u7406\u8bba\u4e0a\uff0c\u8fd8\u662f\u5b9e\u9a8c\u4e0a\uff0c\u5c5e\u6027\u7ea7\u522b\u8def\u7531\u9009\u62e9\u90fd\u662f\u6700\u4f18\u7684\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c<mark style=\"background-color:rgba(0, 0, 0, 0)\" class=\"has-inline-color has-luminous-vivid-orange-color\">\u4e2a\u4eba\u8ba4\u4e3a<\/mark>\uff0cModality-Level Routing\uff08\u6a21\u6001\u7ea7\u8def\u7531\uff09\u5e94\u8be5\u4e0eVLMO\u5de5\u4f5c\u975e\u5e38\u7c7b\u4f3c\u3002\u6216\u8005\u8bf4\uff0c<strong>VLMo \u53ef\u89c6\u4e3a Modality-Level Routing \u7684\u4e00\u79cd\u6781\u7aef\u60c5\u51b5\uff08\u65e0\u52a8\u6001\u5b66\u4e60\uff09<\/strong>\u3002\u82e5\u5c06 Uni-Perceiver-MoE \u7684 Modality-Level Routing \u8bbe\u7f6e\u4e3a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u95e8\u63a7\u6743\u91cd\u00a0<strong>W<\/strong><em>g<\/em>\u200b\u00a0\u56fa\u5b9a\u4e3a\u6a21\u6001\u5230\u4e13\u5bb6\u7684\u786c\u6620\u5c04\u3002<\/li>\n\n\n\n<li>\u6a21\u6001\u5d4c\u5165\uff08<code>embed(id_modal)<\/code>\uff09\u4e3a one-hot \u5411\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5219\u9000\u5316\u4e3a VLMo \u7684\u9759\u6001\u6a21\u6001\u4e13\u5bb6\u5206\u914d\u3002<strong>VLMo<\/strong>\u00a0\u662f\u9762\u5411<strong>\u89c6\u89c9-\u8bed\u8a00\u4efb\u52a1<\/strong>\u7684\u4e13\u7528\u8bbe\u8ba1\uff0c\u901a\u8fc7\u9759\u6001\u5206\u79bb\u4e13\u5bb6\u6700\u5927\u5316\u6a21\u6001\u7279\u5f02\u6027\uff0c\u4f46\u727a\u7272\u4e86\u7075\u6d3b\u6027\u548c\u6269\u5c55\u6027\u3002<\/p>\n\n\n\n<p>\u82e5\u9700\u6784\u5efa\u7edf\u4e00\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\uff0cUni-Perceiver-MoE \u7684\u7b56\u7565\u66f4\u5177\u4f18\u52bf\uff1b\u82e5\u4ec5\u9700\u9ad8\u6548\u7684\u53cc\u6a21\u6001\u5904\u7406\uff08\u5982\u56fe\u6587\u68c0\u7d22\uff09\uff0cVLMo \u7684\u7b80\u5355\u6027\u53ef\u80fd\u66f4\u5408\u9002\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">4. Uni-perceiver v2<\/h2>\n\n\n\n<p>UniPerceiver v2\u4f5c\u4e3a\u591a\u6a21\u6001\u7edf\u4e00\u611f\u77e5\u6846\u67b6\u7684\u5347\u7ea7\u7248\u672c\uff0c\u5176\u6838\u5fc3\u521b\u65b0\u4f53\u73b0\u5728<strong>\u6570\u636e\u8868\u5f81\u7684\u7ec6\u7c92\u5ea6\u91cd\u6784<\/strong>\u4e0e<strong>\u591a\u4efb\u52a1\u8bad\u7ec3\u7684\u68af\u5ea6\u5e73\u8861\u673a\u5236<\/strong>\u4e24\u5927\u7ef4\u5ea6\u3002\u4ee5\u4e0b\u4ece\u6280\u672f\u539f\u7406\u4e0e\u5de5\u7a0b\u5b9e\u8df5\u89d2\u5ea6\u5c55\u5f00\u5206\u6790\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u6a21\u578b\u7ed3\u6784<\/h3>\n\n\n\n<p>\u4f20\u7edf\u89c6\u89c9\u7f16\u7801\u5668\u503e\u5411\u4e8e\u63d0\u53d6\u5168\u5c40\u7279\u5f81\uff08\u5982CNN\u7684\u9ad8\u5c42\u8bed\u4e49\u7279\u5f81\uff09\uff0c\u4f46\u5728\u76ee\u6807\u68c0\u6d4b\u3001\u5b9e\u4f8b\u5206\u5272\u7b49\u4efb\u52a1\u4e2d\uff0c<strong>\u7a7a\u95f4\u5b9a\u4f4d\u7cbe\u5ea6<\/strong>\u4e0e<strong>\u5c40\u90e8\u8bed\u4e49\u533a\u5206\u5ea6<\/strong>\u76f4\u63a5\u5f71\u54cd\u6a21\u578b\u6027\u80fd\u3002UniPerceiver v2\u901a\u8fc7\u4ee5\u4e0b\u53cc\u91cd\u6539\u8fdb\u5b9e\u73b0\u7279\u5f81\u5c42\u7ea7\u7684\u7ec6\u7c92\u5ea6\u89e3\u8026\uff0c\u5176\u6a21\u578b\u7ed3\u6784\u5982\u4e0b\uff1a<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1388\" height=\"1357\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514162451745.png\" alt=\"\" class=\"wp-image-3115\" style=\"width:600px\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514162451745.png 1388w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514162451745-300x293.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514162451745-1024x1001.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514162451745-768x751.png 768w\" sizes=\"auto, (max-width: 1388px) 100vw, 1388px\" \/><\/figure><\/div>\n\n\n<p>\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u6211\u4eec\u4e3b\u8981\u770b\u770b\u89c6\u89c9\u7aef\u53e3\u7684\u5904\u7406\uff0c\u5206\u522b\u6709\u4e24\u4e2a\u5b50\u7f51\u7edc\u6784\u6210\uff1aBackbone \u548c Region Proposal Network\uff08\u5373\u56fe\u4e2d\u6a59\u8272\u80cc\u666f\u7684Transformer\u77e9\u5f62\uff09\u3002\u6839\u636e\u539f\u6587\u63cf\u8ff0\uff1a<\/p>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u4e3b\u5e72\u7f51\u7edc\uff08Backbone\uff09<\/strong><\/h5>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>\u8f93\u5165\u5904\u7406<\/strong><br>\u8f93\u5165\u56fe\u50cf\u5c3a\u5bf8\u4e3aH\u00d7W\uff0c\u901a\u8fc7ResNet\u6216Swin-Transformer\u7b49\u4e3b\u5e72\u7f51\u7edc\u63d0\u53d6<strong>\u591a\u5c3a\u5ea6\u7279\u5f81\u56fe<\/strong>\uff0c\u51714\u4e2a\u5c42\u7ea7\uff08L=4\uff09\u3002\u5404\u5c42\u7ea7\u7279\u5f81\u56fe\u5c3a\u5bf8\u4f9d\u6b21\u4e3a\uff1a\n<ul class=\"wp-block-list\">\n<li>Level 0: H\/4 \u00d7 W\/4<\/li>\n\n\n\n<li>Level 1: H\/8 \u00d7 W\/8<\/li>\n\n\n\n<li>Level 2: H\/16 \u00d7 W\/16<\/li>\n\n\n\n<li>Level 3: H\/32 \u00d7 W\/32<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u7279\u5f81\u8c03\u6574<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u75281\u00d71\u5377\u79ef\u5bf9\u9f50\u7279\u5f81\u56fe\u901a\u9053\u7ef4\u5ea6\uff08\u9690\u85cf\u7ef4\u5ea6d=256\uff09\u3002<\/li>\n\n\n\n<li>\u5bf9Level 3\u7279\u5f81\u989d\u5916\u5e94\u75283\u00d73\u5377\u79ef\uff08\u6b65\u957f2\uff09\uff0c\u751f\u6210\u66f4\u5c0f\u5c3a\u5bf8\u7684Level 4\u7279\u5f81\u56fe\uff08H\/64 \u00d7 W\/64\uff09\u3002<\/li>\n\n\n\n<li>\u5168\u5c40\u6ce8\u610f\u529b\u6c60\u5316\u4fdd\u7559\u524d10\u4e2a\u663e\u8457\u7279\u5f81\uff08M0=10\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>Region Proposal Networ<\/strong>k<\/h5>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>\u7f16\u7801\u9636\u6bb5<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u8f93\u5165\u4e3aLevel 1\u81f3Level 4\u7684\u591a\u5c3a\u5ea6\u7279\u5f81\u56fe\u3002<\/li>\n\n\n\n<li>\u4f7f\u7528<strong>\u53ef\u53d8\u5f62Transformer\u7f16\u7801\u5668<\/strong>\u63d0\u53d6\u7f16\u7801\u7279\u5f81\uff0c\u4fdd\u7559\u539f\u59cb\u7a7a\u95f4\u5c3a\u5bf8\u548c\u7ef4\u5ea6\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u89e3\u7801\u4e0e\u5019\u9009\u751f\u6210<\/strong>\n<ul class=\"wp-block-list\">\n<li><strong>\u5019\u9009\u67e5\u8be2\u6784\u5efa<\/strong>\uff1a<br>\u5bf9\u7f16\u7801\u7279\u5f81\u56fe\u8fdb\u884c\u9010\u50cf\u7d20\u9884\u6d4b\uff0c\u751f\u6210\u76ee\u6807\u7f6e\u4fe1\u5ea6\uff08objectness score\uff09\u548c\u8fb9\u754c\u6846\uff0c\u4ece\u4e2d\u9009\u62e9\u7f6e\u4fe1\u5ea6\u6700\u9ad8\u7684N=900\u4e2a\u7279\u5f81\u70b9\u4f5c\u4e3a\u5019\u9009\u57fa\u3002<\/li>\n\n\n\n<li><strong>Transformer\u89e3\u7801\u5668<\/strong>\uff1a<br>\u5c06\u5019\u9009\u57fa\u4e0e\u968f\u673a\u521d\u59cb\u5316\u7684\u5bf9\u8c61\u67e5\u8be2\u878d\u5408\uff0c\u751f\u6210N\u4e2a\u5019\u9009\u63d0\u8bae\u3002\u6bcf\u4e2a\u63d0\u8bae\u5305\u542b\uff1a\n<ul class=\"wp-block-list\">\n<li><strong>\u8bed\u4e49\u7279\u5f81<\/strong>\uff08q_sem\uff09\uff1a256\u7ef4\u5411\u91cf<\/li>\n\n\n\n<li><strong>\u8fb9\u754c\u6846\u5750\u6807<\/strong>\uff08q_box\uff09\uff1a4\u7ef4\u5750\u6807<\/li>\n\n\n\n<li><strong>\u5206\u5272\u63a9\u7801<\/strong>\uff08q_mask\uff09\uff1aH\u00d7W\u4e8c\u503c\u63a9\u7801\uff08\u901a\u8fc7\u591a\u5c42\u5377\u79ef\u4e0e\u4e0a\u91c7\u6837\u751f\u6210\uff09<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u7279\u5f81\u878d\u5408<\/strong><br>\u5c06\u4e09\u8005\u878d\u5408\u4e3a\u6700\u7ec8\u533a\u57df\u8868\u793a\uff1a<\/li>\n<\/ol>\n\n\n\n<p>$$<br>q_{\\text {proposal }}=q_{\\text {sem }}+B\\left(q_{\\text {box }}\\right)+M\\left(q_{\\text {mask }}\\right)<br>$$<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>B\uff08\u2022\uff09\uff1a\u8fb9\u754c\u6846\u5750\u6807\u7684\u4f4d\u7f6e\u7f16\u7801<\/li>\n\n\n\n<li>M\uff08\u2022\uff09\uff1a\u63a9\u7801\u81ea\u9002\u5e94\u6c60\u5316\u81f3 $28 \\times 28$ \u5c3a\u5bf8\u540e\u7ebf\u6027\u6295\u5f71<\/li>\n\n\n\n<li>\u6839\u636e\u76ee\u6807\u7f6e\u4fe1\u5ea6\u7b5b\u9009 200\u4e2a\u5019\u9009\uff0c\u4f5c\u4e3a\u6700\u7ec8\u533a\u57df\u8868\u793a\u4ee5\u964d\u4f4e\u8ba1\u7b97\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u8bad\u7ec3\u7b56\u7565<\/h3>\n\n\n\n<p>\u5728Uni-Perceiver v2\u4e2d\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a\u00a0<strong>Task-Balanced Gradient Normalization<\/strong>\u00a0\u7684\u68af\u5ea6\u4f18\u5316\u7b56\u7565\uff0c\u7528\u4e8e\u89e3\u51b3\u591a\u4efb\u52a1\u8054\u5408\u8bad\u7ec3\u4e2d\u7684\u68af\u5ea6\u4e0d\u7a33\u5b9a\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u5728\u591a\u4efb\u52a1\u5b66\u4e60\u4e2d\uff0c\u5e38\u89c1\u7684\u4e24\u79cd\u6570\u636e\u91c7\u6837\u7b56\u7565\u5b58\u5728\u7f3a\u9677\uff1a<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>\u6df7\u5408\u91c7\u6837\uff08Mixed Sampling\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u4e2a\u8bad\u7ec3\u8fed\u4ee3\uff08iteration\uff09\u6df7\u5408\u6240\u6709\u4efb\u52a1\u7684\u6570\u636e\u3002<\/li>\n\n\n\n<li><strong>\u95ee\u9898<\/strong>\uff1a\u6bcf\u4e2a\u4efb\u52a1\u7684\u6279\u6b21\u5927\u5c0f\uff08batch-size\uff09\u88ab\u8feb\u51cf\u5c0f\uff0c\u5c24\u5176\u4e0d\u5229\u4e8e\u9700\u8981\u5927\u6279\u6b21\u7684\u4efb\u52a1\uff08\u5982\u56fe\u6587\u68c0\u7d22\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u975e\u6df7\u5408\u91c7\u6837\uff08Unmixed Sampling\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u4e2a\u8fed\u4ee3\u4ec5\u91c7\u6837\u4e00\u4e2a\u4efb\u52a1\u7684\u6570\u636e\uff0c\u53ef\u6700\u5927\u5316\u5355\u4efb\u52a1\u7684\u6279\u6b21\u5927\u5c0f\u3002<\/li>\n\n\n\n<li><strong>\u95ee\u9898<\/strong>\uff1a\u4e0d\u540c\u4efb\u52a1\u7684\u68af\u5ea6\u65b9\u5411\u548c\u91cf\u7ea7\u5dee\u5f02\u5927\uff0c\u5bfc\u81f4\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u548c\u6027\u80fd\u4e0b\u964d\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p><strong>Task-Balanced Gradient Normalization<\/strong> \u7684\u6838\u5fc3\u601d\u60f3\u662f\u901a\u8fc7\u00a0<strong>\u68af\u5ea6\u5f52\u4e00\u5316<\/strong>\u00a0\u548c\u00a0<strong>\u4efb\u52a1\u8865\u507f\u7cfb\u6570<\/strong>\u00a0\u5e73\u8861\u4e0d\u540c\u4efb\u52a1\u7684\u68af\u5ea6\u8d21\u732e\u3002\u4ee5\u4e0b\u662f\u516c\u5f0f\u7684\u9010\u6b65\u62c6\u89e3\uff1a<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1136\" height=\"856\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514164303766.png\" alt=\"\" class=\"wp-image-3119\" style=\"width:600px\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514164303766.png 1136w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514164303766-300x226.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514164303766-1024x772.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/05\/20250514164303766-768x579.png 768w\" sizes=\"auto, (max-width: 1136px) 100vw, 1136px\" \/><\/figure><\/div>\n\n\n<p>\u5173\u952e\u6539\u8fdb\u70b9\u89e3\u6790\uff1a<br>1\uff0e\u68af\u5ea6\u5f52\u4e00\u5316\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5c06\u5f53\u524d\u4efb\u52a1\u7684\u68af\u5ea6 $\\nabla L_{t, k}$ \u9664\u4ee5\u5176\u8303\u6570 $\\left|\\nabla L_{t, k}\\right|$ \uff0c\u8f6c\u5316\u4e3a\u5355\u4f4d\u5411\u91cf\uff08\u65b9\u5411\u4e0d\u53d8\uff0c\u91cf\u7ea7\u7edf\u4e00\u4e3a 1 \uff09\u3002<\/li>\n\n\n\n<li>\u4f5c\u7528\uff1a\u6d88\u9664\u4e0d\u540c\u4efb\u52a1\u68af\u5ea6\u91cf\u7ea7\u7684\u5dee\u5f02\uff0c\u907f\u514d\u67d0\u4e9b\u4efb\u52a1\u4e3b\u5bfc\u66f4\u65b0\u65b9\u5411\u3002<\/li>\n<\/ul>\n\n\n\n<p>2\uff0e\u5e73\u8861\u7cfb\u6570 $\\omega_k$ \uff1a<br>\u53ef\u624b\u52a8\u8bbe\u5b9a\u7684\u6743\u91cd\uff0c\u63a7\u5236\u4efb\u52a1 $k$ \u7684\u68af\u5ea6\u8d21\u732e\u3002\u82e5\u6240\u6709\u4efb\u52a1\u91cd\u8981\u6027\u76f8\u540c\uff0c\u5219\u8bbe $\\omega_k=1$ \u3002<br><\/p>\n\n\n\n<p>3\uff0e\u4efb\u52a1\u8865\u507f\u7cfb\u6570 $1 \/ s_k$ \uff1a<br>$s_k$ \u662f\u4efb\u52a1 $k$ \u7684\u91c7\u6837\u6982\u7387\uff08\u5982\u4efb\u52a1A\u5360\u8bad\u7ec3\u6570\u636e\u7684 $20 \\%$ \uff0c\u5219 $s_k=0.2$ \uff09\u3002<br>\uff0d\u4f5c\u7528\uff1a\u5bf9\u4f4e\u9891\u4efb\u52a1\uff08 $s_k$ \u5c0f\uff09\u7684\u68af\u5ea6\u8fdb\u884c\u653e\u5927 $\\left(\\frac{1}{s_k}>1\\right)$ \uff0c\u907f\u514d\u5176\u68af\u5ea6\u88ab\u9ad8\u9891\u4efb\u52a1\u6df9\u6ca1\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">5. \u603b\u7ed3<\/h2>\n\n\n\n<p>UniPerceiver v2\uff0c\u8fd9\u662f\u7b2c\u4e00\u4e2a\u5728\u4e3b\u8981\u7684\u5927\u89c4\u6a21\u89c6\u89c9\u548c\u89c6\u89c9\u8bed\u8a00\u4efb\u52a1\u4e0a\u53d6\u5f97\u7ade\u4e89\u6027\u7ed3\u679c\u7684\u901a\u7528\u6a21\u578b\u3002\u5728\u5355\u6a21\u6001\u548c\u591a\u6a21\u6001\u4efb\u52a1\u4e0a\u8fdb\u884c\u8054\u5408\u8bad\u7ec3\u540e\uff0cUniPerceiver v2 \u5728\u5e7f\u6cdb\u7684\u4e0b\u6e38\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u5177\u6709\u7ade\u4e89\u529b\u7684\u6027\u80fd\u3002\u81f3\u4e8e\u5c40\u9650\u6027\uff0c\u7531\u4e8e\u8ba1\u7b97\u8d44\u6e90\u6709\u9650\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u5c1a\u672a\u5728\u56fe\u50cf\u751f\u6210\u4efb\u52a1\u4e0a\u5f97\u5230\u9a8c\u8bc1\u3002<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>1.\u5f15\u7528 \u8fd9\u7bc7\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a&nbsp;Uni-Perceiver&nbsp;\u7684\u901a\u7528\u611f\u77e5\u67b6\u6784\uff0c\u65e8\u5728\u89e3\u51b3\u5f53\u524d [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3123,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[29,18],"tags":[],"class_list":["post-3094","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-29","category-18"],"_links":{"self":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3094","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3094"}],"version-history":[{"count":16,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3094\/revisions"}],"predecessor-version":[{"id":3124,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3094\/revisions\/3124"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/media\/3123"}],"wp:attachment":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3094"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3094"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3094"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}