
{"id":11235,"date":"2026-08-20T03:26:26","date_gmt":"2026-08-19T19:26:26","guid":{"rendered":"https:\/\/infernews.com\/blog\/mixture-of-experts-vision-encoder\/"},"modified":"2026-08-20T03:26:26","modified_gmt":"2026-08-19T19:26:26","slug":"mixture-of-experts-vision-encoder","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/mixture-of-experts-vision-encoder\/","title":{"rendered":"moe_vie \u8996\u89ba\u7de8\u78bc\u5668\uff1aCLIP \u898f\u6a21 MoE \u5316\uff0c\u6700\u7d30\u7248\u672c\u4e5f\u8cbc\u8fd1 SOTA"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-bbd622884ec9.jpg\" alt=\"Repository image for facebookresearch\/moe_vie\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8996\u89ba\u7de8\u78bc\u5668\u6108\u8b8a\u6108\u5927\uff0c\u6e96\u78ba\u5ea6\u63db\u4f86\u7684\u662f\u63a8\u8ad6\u6210\u672c\u8207\u5ef6\u9072\u540c\u6642\u81a8\u8139\uff0c\u7279\u5225\u662f\u9ad8\u89e3\u6790\u5ea6\u5716\u7247\u8207\u9577\u5f71\u7247\u60c5\u5883\u3002MoE-ViE \u60f3\u6253\u7834\u9019\u500b\u53d6\u6368\uff1a\u628a Mixture-of-Experts\uff08MoE\uff09\u6a5f\u5236\u5f15\u5165 CLIP \u98a8\u683c\u7684\u5c0d\u6bd4\u9810\u8a13\u7df4\u6d41\u7a0b\uff0c\u8b93\u6a21\u578b\u7e3d\u5bb9\u91cf\u64f4\u5927\uff0c\u4f46\u6bcf\u5f35\u5716\u3001\u6bcf\u6bb5\u7247\u53ea\u555f\u7528\u4e00\u90e8\u5206\u5c08\u5bb6\uff0c\u628a\u5bb9\u91cf\u8207\u5be6\u969b\u8a08\u7b97\u812b\u52fe\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8207\u4e00\u822c MoE \u4e0d\u540c\uff0cMoE-ViE \u63a1\u7528\u7d30\u7c92\u5ea6\u5c08\u5bb6\u62d3\u6a38\uff08fine-grained MoE topology\uff09\uff0c\u4e26\u63d0\u51fa\u4e00\u500b\u7121\u8f14\u52a9\u640d\u5931\uff08auxiliary-loss-free\uff09\u7684\u8b8a\u9ad4\u4f86\u5e73\u8861\u5c08\u5bb6\u8ca0\u8f09\uff0c\u907f\u514d\u50b3\u7d71 MoE \u5e38\u898b\u7684\u8ca0\u8f09\u4e0d\u5747\u554f\u984c\u3002\u540c\u6642\uff0c\u4f5c\u8005\u91dd\u5c0d MoE \u984d\u5916\u958b\u92b7\u64b0\u5beb\u4e86\u5c08\u9580\u7684 Triton kernel\uff0c\u628a\u63a8\u8ad6\u5ef6\u9072\u62c9\u56de\u63a5\u8fd1\u5bc6\u96c6\u6a21\u578b\u7684\u6c34\u6e96\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u4ee5 Vision Transformer \u9aa8\u5e79\u642d\u914d\u4e0a\u8ff0 MoE \u8a2d\u8a08\uff0c\u914d\u5408\u7a69\u5065\u7684\u5c0d\u6bd4\u9810\u8a13\u7df4\u6d41\u7a0b\uff08contrastive pretraining recipe\uff09\uff0c\u4e00\u53e3\u6c23\u91cb\u51fa B \/ L \/ H \u4e09\u500b\u5c3a\u5bf8\u7684\u6a21\u578b\uff0c\u8986\u84cb 224 \/ 384 \/ 448px \u89e3\u6790\u5ea6\u3002\u5728 ImageNet-1k\u3001ObjectNet\u3001COCO \u6587\u5b57\u6aa2\u7d22\u5716\u7247\u3001Kinetics-400\u3001MSR-VTT \u6587\u5b57\u6aa2\u7d22\u5f71\u7247\u7b49\u57fa\u6e96\u4e0a\uff0c\u96f6\u6a23\u672c\u8868\u73fe\u8207\u5404\u9805\u5f37 CLIP baseline \u76f8\u7576\u751a\u81f3\u66f4\u4f73\uff0c\u5176\u4e2d H\/14 \u7248\u672c\u5728\u4fdd\u6301\u7d04 76% \u5ef6\u9072\u7684\u60c5\u6cc1\u4e0b\uff0c\u8cbc\u8fd1\u9ad4\u578b\u5927 1.7 \u500d\u7684 SOTA \u7de8\u78bc\u5668\u3002\u7576\u5c0d\u63a5 LLM \u7d44\u6210\u8996\u89ba\u8a9e\u8a00\u6a21\u578b\u6642\uff0cMoE-ViE \u5728\u5716\u50cf\u8207\u5f71\u7247\u57fa\u6e96\u4e0a\u4ea6\u52dd\u904e\u591a\u500b\u555f\u7528\u53c3\u6578\u591a\u9054 5 \u500d\u7684\u7de8\u78bc\u5668\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u4e00\u500b\u91dd\u5c0d\u5f71\u7247\u80fd\u529b\u7684\u8655\u7406\u65b9\u5f0f\u662f frame-level distillation \u52a0\u4e0a\u4e00\u500b\u65b0\u8a2d\u8a08\u7684\u51cd\u7d50\u6a5f\u5236\uff0c\u76ee\u7684\u662f\u5728\u5f15\u5165\u5f71\u7247\u7406\u89e3\u7684\u540c\u6642\u4fdd\u7559\u539f\u672c\u5b78\u5230\u7684\u5716\u50cf\u77e5\u8b58\u3002\u9019\u500b\u6b65\u9a5f\u4ee4\u540c\u4e00\u7d44\u8996\u89ba\u7de8\u78bc\u5668\u53ef\u4ee5\u540c\u6642\u670d\u52d9\u5716\u50cf\u8207\u5f71\u7247\u4efb\u52d9\uff0c\u800c\u4e0d\u5fc5\u8a13\u7df4\u5169\u5957\u6a21\u578b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cd\u9ede\u6458\u8981\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>MoE \u5316\u7684 CLIP \u98a8\u683c\u8996\u89ba\u7de8\u78bc\u5668<\/strong>\uff1a\u4ee5 Mixture-of-Experts Vision Transformer \u914d\u5408\u5c0d\u6bd4\u9810\u8a13\u7df4\uff0c\u652f\u63f4 B \/ L \/ H \u4e09\u7a2e\u898f\u6a21\u3002<\/li>\n<li><strong>\u7d30\u7c92\u5ea6\u5c08\u5bb6\u62d3\u6a38 + \u7121\u8f14\u52a9\u640d\u5931\u5e73\u8861<\/strong>\uff1a\u907f\u514d\u50b3\u7d71 MoE \u7684\u5c08\u5bb6\u8ca0\u8f09\u4e0d\u5747\uff0c\u4e26\u4ee5\u81ea\u8a02 Triton kernel \u63a7\u5236\u63a8\u8ad6\u5ef6\u9072\u3002<\/li>\n<li><strong>\u5716\u50cf\u8207\u5f71\u7247\u517c\u9867<\/strong>\uff1aframe-level distillation \u914d\u5408\u65b0\u51cd\u7d50\u6a5f\u5236\uff0c\u8b93\u55ae\u4e00\u7de8\u78bc\u5668\u540c\u6642\u8655\u7406\u5716\u50cf\u8207\u5f71\u7247\u7406\u89e3\u3002<\/li>\n<li><strong>\u5c0d\u63a5 LLM \u5f8c\u4ecd\u5177\u7af6\u722d\u529b<\/strong>\uff1a\u5728\u591a\u9805\u5716\u50cf\u8207\u5f71\u7247\u57fa\u6e96\u4e0a\uff0c\u8868\u73fe\u52dd\u904e\u555f\u7528\u53c3\u6578\u591a\u9054 5 \u500d\u7684\u65e2\u6709\u7de8\u78bc\u5668\u3002<\/li>\n<li><strong>\u5b98\u65b9\u91cb\u51fa\u4ee3\u78bc\u8207\u96f6\u6a23\u672c\u8a55\u6e2c\u5957\u4ef6<\/strong>\uff1a\u63d0\u4f9b\u6a21\u578b\u5b9a\u7fa9\u3001\u914d\u7f6e\u6a94\u8207\u53ef\u91cd\u73fe\u7684\u96f6\u6a23\u672c\u8a55\u4f30\u6d41\u7a0b\uff0c\u60df\u6388\u6b0a\u70ba CC BY-NC 4.0\uff0c\u9700\u7559\u610f\u975e\u5546\u7528\u9650\u5236\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u672c\u9805\u76ee\u7531 Meta \u5718\u968a\u767c\u4f48\u3002\u9700\u8981\u7559\u610f\u7684\u662f\u6388\u6b0a\u70ba CC BY-NC 4.0\uff0c\u50c5\u9650\u975e\u5546\u7528\u5834\u666f\u4f7f\u7528\uff1b\u8a13\u7df4\u7d30\u7bc0\u3001\u8cc7\u6599\u7d44\u6210\u8207\u5c0d\u9f4a LLM \u7684\u5be6\u4f5c\u5c6c\u7814\u7a76\u914d\u7f6e\uff0c\u91cd\u73fe\u6210\u672c\u4e0d\u4f4e\uff0c\u8f03\u9069\u5408\u4f5c\u7814\u7a76\u57fa\u6e96\u800c\u975e\u5373\u53d6\u5373\u7528\u7684\u7522\u54c1\u5143\u4ef6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/facebookresearch\/moe_vie\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/facebook\/MoEViE-L16-384\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Meta \u5718\u968a\u628a Mixture-of-Experts \u5e36\u9032 CLIP \u98a8\u683c\u8996\u89ba\u7de8\u78bc\u5668\uff0c\u7528\u7d30\u7c92\u5ea6\u5c08\u5bb6\u8def\u7531\u914d\u5408\u81ea\u8a02 Triton kernel\uff0c\u5728\u96f6\u6a23\u672c\u5206\u985e\u8207\u6aa2\u7d22\u4e0a\u8cbc\u8fd1\u9ad4\u578b\u5927 1.7 \u500d\u7684 SOTA \u7de8\u78bc\u5668\u3002<\/p>\n","protected":false},"author":8,"featured_media":11234,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,119,127,188],"tags":[],"class_list":["post-11235","post","type-post","status-publish","format-standard","hentry","category-133","category-119","category-127","category-meta"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11235","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=11235"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11235\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11234"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11235"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11235"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11235"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}