{"id":1072398,"date":"2023-06-20T21:23:00","date_gmt":"2023-06-21T04:23:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&#038;p=1072398"},"modified":"2024-09-25T03:40:21","modified_gmt":"2024-09-25T10:40:21","slug":"cvpr-2023-mim","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/cvpr-2023-mim\/","title":{"rendered":"CVPR 2023 | \u63a9\u7801\u56fe\u50cf\u5efa\u6a21MIM\u7684\u7406\u89e3\u3001\u5c40\u9650\u4e0e\u6269\u5c55"},"content":{"rendered":"\n<p>\u7f16\u8005\u6309\uff1a\u63a9\u7801\u56fe\u50cf\u5efa\u6a21\uff08Masked Image Modeling, MIM\uff09\u7684\u63d0\u51fa\uff0c\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u6a21\u578b\u8bad\u7ec3\u5f15\u5165\u65e0\u76d1\u7763\u5b66\u4e60\u505a\u51fa\u4e86\u91cd\u8981\u8d21\u732e\u3002\u5f97\u76ca\u4e8e MIM \u7684\u9884\u8bad\u7ec3\u7b97\u6cd5\uff0c\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5728\u8fd1\u5e74\u6765\u6301\u7eed\u8f93\u51fa\u7740\u4f18\u8d28\u7684\u7814\u7a76\u6210\u679c\u3002\u7136\u800c\u6574\u4e2a\u4e1a\u754c\u5bf9 MIM \u673a\u5236\u7684\u7814\u7a76\u4ecd\u5b58\u5728\u4e0d\u8db3\u3002<\/p>\n\n\n\n<p>\u79c9\u6301\u7740\u4e0d\u65ad\u6269\u5c55\u524d\u6cbf\u6280\u672f\u8fb9\u754c\u7684\u63a2\u7d22\u7cbe\u795e\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5728\u7406\u89e3 MIM \u4f5c\u7528\u673a\u5236\uff0c\u4ee5\u53ca\u57fa\u4e8e\u8fd9\u4e9b\u673a\u5236\u63d0\u5347\u73b0\u6709 MIM \u7b97\u6cd5\u7684\u9886\u57df\uff0c\u53d6\u5f97\u4e86\u4e00\u7cfb\u5217\u7684\u521b\u65b0\u6210\u679c\uff0c\u5e76\u83b7\u5f97\u4e86 CVPR 2023 \u7684\u8ba4\u53ef\u3002\u8fd9\u4e9b\u6210\u679c\u5305\u542b\uff1a\u57fa\u4e8e MIM \u9884\u8bad\u7ec3\u65b9\u6cd5\u7684\u6269\u5c55\u6cd5\u5219\u7814\u7a76\u3001\u5206\u6790 MIM \u7684\u5177\u4f53\u6027\u8d28\u4ee5\u53ca\u6709\u6548\u6027\u80cc\u540e\u7684\u539f\u56e0\u3001\u901a\u8fc7\u84b8\u998f\u6280\u672f\u5c06 MIM \u6a21\u578b\u7684\u4f18\u52bf\u62d3\u5c55\u5230\u5c0f\u6a21\u578b\u4e2d\u3002<\/p>\n\n\n\n<p>\u6b63\u503c CVPR 2023 \u5927\u4f1a\u4e3e\u529e\u4e4b\u9645\uff0c\u8ba9\u6211\u4eec\u4e00\u8d77\u4e86\u89e3 MIM \u52a9\u63a8\u8ba1\u7b97\u673a\u89c6\u89c9\u7814\u7a76\u8fdb\u5165\u52a0\u901f\u9053\u7684\u65b0\u53d1\u73b0\u5427\uff01<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u9884\u8bad\u7ec3-\u5fae\u8c03\uff08Pre-training and Fine-tuning\uff09\u662f\u8fc7\u53bb\u5341\u5e74\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u6700\u91cd\u8981\u7684\u5b66\u4e60\u8303\u5f0f\u4e4b\u4e00\uff0c\u5176\u57fa\u672c\u60f3\u6cd5\u662f\u5728\u6d77\u91cf\u6570\u636e\u7684\u4efb\u52a1\u4e2d\uff0c\u5bf9\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u8bad\u7ec3\uff0c\u7136\u540e\u518d\u5c06\u9884\u8bad\u7ec3\u8fc7\u7684\u6a21\u578b\u5728\u4e0b\u6e38\u6570\u636e\u91cf\u8f83\u5c11\u7684\u4efb\u52a1\u4e2d\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u79cd\u65b9\u5f0f\u80fd\u591f\u5c06\u4e0a\u6e38\u5927\u6570\u636e\u4efb\u52a1\u4e2d\u5b66\u5230\u7684\u4fe1\u606f\u8fc1\u79fb\u81f3\u4e0b\u6e38\u6570\u636e\u91cf\u8f83\u5c11\u7684\u4efb\u52a1\u4e0a\uff0c\u7f13\u89e3\u6570\u636e\u91cf\u4e0d\u8db3\u7684\u95ee\u9898\uff0c\u5e76\u663e\u8457\u63d0\u5347\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u9884\u8bad\u7ec3-\u5fae\u8c03\u8303\u5f0f\u7684\u6210\u529f\uff0c\u6e90\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5341\u5e74\u6765\u9884\u8bad\u7ec3\u7b97\u6cd5\u7684\u505c\u6ede\u3002\u81ea2012\u5e74 AlexNet \u63d0\u51fa\u4ee5\u6765\uff0c\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u7684\u9884\u8bad\u7ec3\u7b97\u6cd5\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u88ab\u7b49\u4ef7\u4e8e\u4ee5 ImageNet \u6570\u636e\u96c6\u4e3a\u4ee3\u8868\u7684\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u3002\u5c3d\u7ba1\u56fe\u50cf\u5206\u7c7b\u7684\u6570\u636e\u6807\u6ce8\u6210\u672c\u5df2\u7136\u8f83\u4f4e\uff0c\u4f46\u540e\u7eed\u7684\u6570\u636e\u6e05\u6d17\u3001\u8d28\u91cf\u63a7\u5236\u7b49\u6b65\u9aa4\u4ecd\u5bf9\u6269\u5c55\u56fe\u50cf\u5206\u7c7b\u6570\u636e\u4ea7\u751f\u4e86\u6311\u6218\uff0c\u800c\u6570\u636e\u4e0d\u8db3\u7684\u56f0\u96be\u4e5f\u9650\u5236\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u6a21\u578b\u7684\u8fdb\u4e00\u6b65\u6269\u5927\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u4f7f\u7528\u65e0\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\u8fdb\u884c\u89c6\u89c9\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u9010\u6e10\u6210\u4e3a\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u4e2d\u7684\u6838\u5fc3\u95ee\u9898\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"642\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-1-1024x642.png\" alt=\"\u56fe1\uff1a\u9884\u8bad\u7ec3-\u5fae\u8c03\u8303\u5f0f\" class=\"wp-image-1072509\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-1-1024x642.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-1-300x188.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-1-768x482.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-1-240x151.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-1.png 1202w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe1\uff1a\u9884\u8bad\u7ec3-\u5fae\u8c03\u8303\u5f0f<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45397\"><\/p>\n\n\n\n<p>2021\u5e746\u6708\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u7684 BEiT \u65b9\u6cd5\uff0c\u901a\u8fc7\u5f15\u5165\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u4e2d\u7684\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\uff08Masked Language Modeling, MLM\uff09\u7b97\u6cd5\uff0c\u6210\u529f\u5730\u8bc1\u660e\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u65e0\u76d1\u7763\u9884\u8bad\u7ec3\u53ef\u4ee5\u8fbe\u5230\u4e0e\u6709\u76d1\u7763\u9884\u8bad\u7ec3\u76f8\u540c\u751a\u81f3\u66f4\u597d\u7684\u6548\u679c\u30022021\u5e7411\u6708\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u7684 SimMIM \u4e0e Meta \u63d0\u51fa\u7684 MAE \u8fdb\u4e00\u6b65\u7b80\u5316\u4e86 BEiT\uff0c\u5e76\u63d0\u5347\u4e86\u7b97\u6cd5\u6027\u80fd\u3002\u81ea\u6b64\uff0c\u63a9\u7801\u56fe\u50cf\u5efa\u6a21\uff08Masked Image Modeling, MIM\uff09\u7684\u7814\u7a76\u8303\u5f0f\u6b63\u5f0f\u5f00\u542f\u3002<\/p>\n\n\n\n<p>\u867d\u7136\u57fa\u4e8e MIM \u7684\u9884\u8bad\u7ec3\u7b97\u6cd5\u7684\u6210\u679c\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5185\u767e\u82b1\u9f50\u653e\uff0c\u4f46\u5bf9 MIM \u673a\u5236\u7684\u63a2\u7d22\u4ecd\u7136\u5341\u5206\u532e\u4e4f\u3002\u4eca\u5929\u6211\u4eec\u5c06\u4ecb\u7ecd\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u89c6\u89c9\u8ba1\u7b97\u7ec4\u5728\u7406\u89e3 MIM \u4f5c\u7528\u673a\u5236\uff0c\u4ee5\u53ca\u57fa\u4e8e\u8fd9\u4e9b\u673a\u5236\u6269\u5c55\u5e76\u63d0\u5347\u73b0\u6709 MIM \u7b97\u6cd5\u7684\u7cfb\u5217\u5de5\u4f5c\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u63a2\u7d22-mim-\u7684\u6269\u5c55\u6cd5\u5219\u4e0e\u6570\u636e\u53ef\u6269\u5c55\u6027\">\u63a2\u7d22 MIM \u7684\u6269\u5c55\u6cd5\u5219\u4e0e\u6570\u636e\u53ef\u6269\u5c55\u6027<\/h2>\n\n\n\n<p>\u6269\u5c55\u6cd5\u5219\uff08scaling law\uff09\u7684\u6982\u5ff5\u6700\u521d\u7531 OpenAI \u53d1\u8868\u4e8e2020\u5e74\u7684\u201cScaling Laws for Neural Language Models\u201d\uff0c\u6587\u4e2d\u63d0\u51fa\uff1a\u6d4b\u8bd5\u96c6\u4e0a\u7684 Loss \u4f1a\u968f\u7740\u8ba1\u7b97\uff08compute\uff09\u3001\u6570\u636e\u89c4\u6a21\uff08dataset size\uff09\u4e0e\u6a21\u578b\u53c2\u6570\u91cf\uff08parameters\uff09\u7684\u589e\u52a0\u800c\u5448\u73b0\u53ef\u4ee5\u9884\u6d4b\u7684\u4e0b\u964d\u6a21\u5f0f\u3002\u8be5\u53d1\u73b0\u5bf9\u4e8e\u5982\u4f55\u4f18\u5316\u81ea\u7136\u8bed\u8a00\u6a21\u578b\u7684\u8bbe\u8ba1\u4e0e\u8bad\u7ec3\uff0c\u5177\u6709\u91cc\u7a0b\u7891\u5f0f\u7684\u6307\u5bfc\u610f\u4e49\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"313\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-2-1024x313.png\" alt=\"\u56fe2\uff1a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\u7684\u6269\u5c55\u6cd5\u5219\uff1a\u6d4b\u8bd5\u96c6 Loss \u968f\u7740\u8ba1\u7b97\uff0c\u6570\u636e\u89c4\u6a21\u4ee5\u53ca\u6a21\u578b\u53c2\u6570\u7684\u589e\u52a0\u5448\u73b0\u53ef\u4ee5\u9884\u6d4b\u7684\u4e0b\u964d\u6a21\u5f0f\" class=\"wp-image-1072512\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-2-1024x313.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-2-300x92.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-2-768x235.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-2-1536x470.png 1536w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-2-240x73.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-2.png 1622w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe2\uff1a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\u7684\u6269\u5c55\u6cd5\u5219\uff1a\u6d4b\u8bd5\u96c6 Loss \u968f\u7740\u8ba1\u7b97\uff0c\u6570\u636e\u89c4\u6a21\u4ee5\u53ca\u6a21\u578b\u53c2\u6570\u7684\u589e\u52a0\u5448\u73b0\u53ef\u4ee5\u9884\u6d4b\u7684\u4e0b\u964d\u6a21\u5f0f<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45398\"><\/p>\n\n\n\n<p>\u5728\u5165\u9009 CVPR 2023 \u7684\u201cOn Data Scaling in Masked Image Modeling\u201d\uff08\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2206.04664\uff09\u4e00\u6587\u4e2d\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u4e5f\u63a2\u7d22\u4e86\u57fa\u4e8e MIM \u9884\u8bad\u7ec3\u65b9\u6cd5\u7684\u6269\u5c55\u6cd5\u5219\u3002\u5c3d\u7ba1\u5728\u8ba1\u7b97\u4e0e\u6a21\u578b\u5927\u5c0f\u8fd9\u4e24\u4e2a\u7ef4\u5ea6\u4e2d\uff0cMIM \u9884\u8bad\u7ec3\u7b97\u6cd5\u4e5f\u5448\u73b0\u4e86\u8f83\u597d\u7684\u6269\u5c55\u6027\u8d28\uff0c\u4f46\u662f\u5728\u6570\u636e\u7ef4\u5ea6\u4e0a\uff0cMIM \u7b97\u6cd5\u5219\u5448\u73b0\u4e86\u4e0e\u5728 NLP \u4e2d\u622a\u7136\u4e0d\u540c\u7684\u7279\u6027\uff1a\u6d4b\u8bd5\u96c6\u7684 Loss \u968f\u7740\u6570\u636e\u96c6\u5927\u5c0f\u8fbe\u5230\u4e00\u5b9a\u89c4\u6a21\u540e\u4e0d\u518d\u964d\u4f4e\uff0c\u5448\u9971\u548c\u72b6\u3002\u8fd9\u5f15\u53d1\u4e86\u4e00\u4e2a\u5173\u952e\u95ee\u9898\u2014\u2014\u4f5c\u4e3a\u4e00\u4e2a\u65e0\u76d1\u7763\u9884\u8bad\u7ec3\u7b97\u6cd5\uff0cMIM \u662f\u5426\u80fd\u4ece\u66f4\u591a\u7684\u6570\u636e\u4e2d\u53d7\u76ca\uff1f\u6362\u8a00\u4e4b\uff0cMIM \u662f\u5426\u5177\u6709\u6570\u636e\u53ef\u6269\u5c55\u6027\uff1f<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"311\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-3-1024x311.png\" alt=\"\u56fe3\uff1aMIM \u4e2d\u7684\u6269\u5c55\u6cd5\u5219\uff1a\u6d4b\u8bd5\u96c6 Loss \u4ec5\u968f\u7740\u8ba1\u7b97\u4e0e\u6a21\u578b\u53c2\u6570\u7684\u589e\u52a0\u5448\u73b0\u53ef\u9884\u6d4b\u7684\u4e0b\u964d\u6a21\u5f0f\uff0c\u800c\u5728\u6570\u636e\u96c6\u5927\u5c0f\u7ef4\u5ea6\u4e0a\uff0c\u5448\u73b0\u4e86\u9971\u548c\u7684\u73b0\u8c61\" class=\"wp-image-1072515\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-3-1024x311.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-3-300x91.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-3-768x233.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-3-1536x467.png 1536w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-3-240x73.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-3.png 1758w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe3\uff1aMIM \u4e2d\u7684\u6269\u5c55\u6cd5\u5219\uff1a\u6d4b\u8bd5\u96c6 Loss \u4ec5\u968f\u7740\u8ba1\u7b97\u4e0e\u6a21\u578b\u53c2\u6570\u7684\u589e\u52a0\u5448\u73b0\u53ef\u9884\u6d4b\u7684\u4e0b\u964d\u6a21\u5f0f\uff0c\u800c\u5728\u6570\u636e\u96c6\u5927\u5c0f\u7ef4\u5ea6\u4e0a\uff0c\u5448\u73b0\u4e86\u9971\u548c\u7684\u73b0\u8c61<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45399\"><\/p>\n\n\n\n<p>\u4e3a\u4e86\u56de\u7b54\u8be5\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u5206\u6790\u4e86\u6a21\u578b\u5927\u5c0f\u3001\u6570\u636e\u89c4\u6a21\u4ee5\u53ca\u8bad\u7ec3\u957f\u5ea6\u7684\u5f71\u54cd\uff0c\u53d1\u73b0 MIM \u5177\u6709\u6570\u636e\u53ef\u6269\u5c55\u6027\uff0c\u4f46\u9700\u8981\u6ee1\u8db3\u4e24\u4e2a\u5173\u952e\u7684\u6761\u4ef6\uff1a1\uff09\u9700\u8981\u66f4\u5927\u7684\u6a21\u578b\uff1b2\uff09\u9700\u8981\u914d\u4ee5\u66f4\u957f\u7684\u8bad\u7ec3\u8f6e\u6570\u3002\u8fdb\u4e00\u6b65\u7684\u89c2\u5bdf\u8868\u660e\uff0c\u8be5\u73b0\u8c61\u662f\u7531\u8fc7\u62df\u5408\uff08over-fitting\uff09\u5bfc\u81f4\u7684\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"312\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-4-1024x312.png\" alt=\"\u56fe4\uff1aMIM \u4e2d\u7684\u8fc7\u62df\u5408\u73b0\u8c61\" class=\"wp-image-1072518\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-4-1024x312.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-4-300x91.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-4-768x234.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-4-240x73.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-4.png 1269w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe4\uff1aMIM \u4e2d\u7684\u8fc7\u62df\u5408\u73b0\u8c61<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45400\"><\/p>\n\n\n\n<p>\u5982\u56fe4\u6240\u793a\uff0c\u5bf9\u4e00\u4e9b\u8f83\u5927\u7684\u6a21\u578b\uff0c\u4f7f\u7528\u5c0f\u6570\u636e\u4e0e\u957f\u8bad\u7ec3\u8f6e\u6570\u4f1a\u4f7f\u5f97\u8bad\u7ec3 Loss \u5f02\u5e38\u4e0b\u964d\uff0c\u6d4b\u8bd5 Loss \u4e0e\u4e0b\u6e38\u4efb\u52a1\u4e2d\u7684\u5fae\u8c03\u6027\u80fd\u53d7\u635f\u3002\u540c\u65f6\uff0c\u8fc7\u62df\u5408\u72b6\u6001\u65f6\uff0c\u6a21\u578b\u66f4\u503e\u5411\u4e8e\u5448\u73b0\u8bb0\u5fc6\u56fe\u50cf\u7684\u6027\u8d28\uff1b\u975e\u8fc7\u62df\u5408\u72b6\u6001\u65f6\uff0c\u66f4\u503e\u5411\u5448\u73b0\u63a8\u7406\u7684\u6027\u8d28\u3002\u57fa\u4e8e\u8fd9\u4e9b\u53d1\u73b0\uff0c\u7814\u7a76\u5458\u4eec\u8ba4\u4e3a\u5bf9\u4e8e MIM \u800c\u8a00\uff0c\u6d4b\u8bd5\u96c6 Loss \u6bd4\u8bad\u7ec3 Loss \u66f4\u9002\u5408\u4f5c\u4e3a\u4e0b\u6e38\u4efb\u52a1\u8fc1\u79fb\u80fd\u529b\u7684\u4ee3\u7406\u6307\u6807\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"437\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-5-1024x437.png\" alt=\"\u56fe5\uff1a\u5c55\u793a\u4e86\u8fc7\u62df\u5408\u6a21\u578b\u4e0e\u975e\u8fc7\u62df\u5408\u6a21\u578b\u5bf9\u56fe\u7247\u8865\u5168\u7684\u53ef\u89c6\u5316\u7ed3\u679c\u3002\u8fc7\u62df\u5408\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e0a\u4f1a\u8bb0\u5fc6\u539f\u56fe\uff0c\u800c\u5728\u6d4b\u8bd5\u96c6\u4e0a\u5219\u65e0\u6cd5\u6b63\u786e\u7684\u63a8\u7406\u5185\u5bb9\u3002\u975e\u8fc7\u62df\u5408\u6a21\u578b\u5219\u5728\u8bad\u7ec3\u96c6\u4e0e\u6d4b\u8bd5\u96c6\u56fe\u50cf\u4e0a\u90fd\u8868\u73b0\u51fa\u8f83\u597d\u7684\u63a8\u7406\u80fd\u529b\u3002\" class=\"wp-image-1072521\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-5-1024x437.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-5-300x128.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-5-768x327.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-5-240x102.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-5.png 1412w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe5\uff1a\u5c55\u793a\u4e86\u8fc7\u62df\u5408\u6a21\u578b\u4e0e\u975e\u8fc7\u62df\u5408\u6a21\u578b\u5bf9\u56fe\u7247\u8865\u5168\u7684\u53ef\u89c6\u5316\u7ed3\u679c\u3002\u8fc7\u62df\u5408\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e0a\u4f1a\u8bb0\u5fc6\u539f\u56fe\uff0c\u800c\u5728\u6d4b\u8bd5\u96c6\u4e0a\u5219\u65e0\u6cd5\u6b63\u786e\u7684\u63a8\u7406\u5185\u5bb9\u3002\u975e\u8fc7\u62df\u5408\u6a21\u578b\u5219\u5728\u8bad\u7ec3\u96c6\u4e0e\u6d4b\u8bd5\u96c6\u56fe\u50cf\u4e0a\u90fd\u8868\u73b0\u51fa\u8f83\u597d\u7684\u63a8\u7406\u80fd\u529b\u3002<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45401\"><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u66f4\u6df1\u5165\u5730\u7406\u89e3-mim-\u53ca\u5176\u6709\u6548\u6027\">\u66f4\u6df1\u5165\u5730\u7406\u89e3 MIM \u53ca\u5176\u6709\u6548\u6027<\/h2>\n\n\n\n<p>MIM \u5c55\u793a\u4e86\u5176\u5728\u9884\u8bad\u7ec3-\u5fae\u8c03\u8303\u5f0f\u4e0b\u7684\u5e7f\u6cdb\u6709\u6548\u6027\u3002\u4f20\u7edf\u89c6\u89d2\u901a\u5e38\u8ba4\u4e3a\u6a21\u578b\u7684\u6709\u6548\u6027\u53d6\u51b3\u4e8e\u5176\u63d0\u53d6\u7684\u7279\u5f81\u8d28\u91cf\u3002\u7136\u800c\uff0c\u8fdb\u4e00\u6b65\u5b9e\u9a8c\u53d1\u73b0\uff0c\u5728\u56fa\u5b9a\u7f51\u7edc\u6743\u91cd\u7684\u8bbe\u5b9a\u65f6\uff0cMIM \u7684\u6027\u80fd\u8fdc\u900a\u8272\u4e8e\u5176\u4ed6\u9884\u8bad\u7ec3\u7b97\u6cd5\u3002\u8fd9\u8bf4\u660e MIM \u7684\u6709\u6548\u6027\u6e90\u81ea\u5176\u4ed6\u56e0\u7d20\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"268\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-6-1024x268.png\" alt=\"\u88681\uff1a\u5728\u56fa\u5b9a\u7f51\u7edc\u6743\u91cd\uff08frozen setting\uff09\u4e0e\u5fae\u8c03\u5168\u7f51\u7edc\u6743\u91cd\uff08full fine-tuning setting\uff09\u4e0b\uff0c\u6bd4\u8f83\u4e0d\u540c\u9884\u8bad\u7ec3\u7b97\u6cd5\u7684\u6027\u80fd\" class=\"wp-image-1072524\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-6-1024x268.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-6-300x78.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-6-768x201.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-6-240x63.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-6.png 1476w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u88681\uff1a\u5728\u56fa\u5b9a\u7f51\u7edc\u6743\u91cd\uff08frozen setting\uff09\u4e0e\u5fae\u8c03\u5168\u7f51\u7edc\u6743\u91cd\uff08full fine-tuning setting\uff09\u4e0b\uff0c\u6bd4\u8f83\u4e0d\u540c\u9884\u8bad\u7ec3\u7b97\u6cd5\u7684\u6027\u80fd<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45402\"><\/p>\n\n\n\n<p>\u4e8e\u662f\uff0c\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u5165\u9009 CVPR 2023 \u7684\u53e6\u4e00\u7bc7\u8bba\u6587\u201cRevealing the Dark Secrets of Masked Image Modeling\u201d\u4e2d\uff08\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2205.13543\uff09\uff0c\u7814\u7a76\u5458\u4eec\u5bf9 MIM \u7684\u6027\u8d28\u4ee5\u53ca\u6709\u6548\u6027\u80cc\u540e\u7684\u539f\u56e0\u8fdb\u884c\u4e86\u66f4\u7ec6\u81f4\u7684\u7814\u7a76\u4e0e\u5206\u6790\uff0c\u53d6\u5f97\u4e86\u5982\u4e0b\u53d1\u73b0\uff1a<\/p>\n\n\n\n<p>1\uff09 \u6709\u76d1\u7763\u9884\u8bad\u7ec3\u4ee5\u53ca\u57fa\u4e8e\u5bf9\u6bd4\u5b66\u4e60\u7684\u9884\u8bad\u7ec3\u65b9\u6cd5\u7684\u6df1\u5c42\u7f51\u7edc\u4ec5\u5efa\u6a21\u957f\u7a0b\u4fe1\u606f\uff0c\u76f8\u6bd4\u4e4b\u4e0b\uff0cMIM \u80fd\u591f\u5bf9\u5c40\u90e8\u4fe1\u606f\u4e0e\u957f\u7a0b\u4fe1\u606f\u540c\u65f6\u5efa\u6a21\uff0c\u5982\u56fe6\u6240\u793a\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"237\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-7-1024x237.png\" alt=\"\u56fe6\uff1a\u4e0d\u540c\u6a21\u578b\u4e2d\u7684\u6ce8\u610f\u529b\u8ddd\u79bb\uff08attention distance\uff09\u3002\u6709\u76d1\u7763\u9884\u8bad\u7ec3\u4e0e\u5bf9\u6bd4\u5b66\u4e60\u9884\u8bad\u7ec3\u7b97\u6cd5\u5728\u7f51\u7edc\u7684\u6df1\u5c42\u53ea\u5173\u6ce8\u957f\u7a0b\u4fe1\u606f\uff0c\u800c MIM \u65b9\u6cd5\u540c\u65f6\u5173\u6ce8\u957f\u7a0b\u4fe1\u606f\u4e0e\u5c40\u90e8\u4fe1\u606f\u3002\" class=\"wp-image-1072527\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-7-1024x237.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-7-300x69.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-7-768x178.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-7-1536x355.png 1536w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-7-240x55.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-7.png 1972w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe6\uff1a\u4e0d\u540c\u6a21\u578b\u4e2d\u7684\u6ce8\u610f\u529b\u8ddd\u79bb\uff08attention distance\uff09\u3002\u6709\u76d1\u7763\u9884\u8bad\u7ec3\u4e0e\u5bf9\u6bd4\u5b66\u4e60\u9884\u8bad\u7ec3\u7b97\u6cd5\u5728\u7f51\u7edc\u7684\u6df1\u5c42\u53ea\u5173\u6ce8\u957f\u7a0b\u4fe1\u606f\uff0c\u800c MIM \u65b9\u6cd5\u540c\u65f6\u5173\u6ce8\u957f\u7a0b\u4fe1\u606f\u4e0e\u5c40\u90e8\u4fe1\u606f\u3002<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45403\"><\/p>\n\n\n\n<p>2\uff09 MIM \u4e2d\u4e0d\u540c\u6ce8\u610f\u529b\u5934\uff08attention head\uff09\u5173\u6ce8\u7684\u4fe1\u606f\u5177\u6709\u591a\u6837\uff0c\u5982\u56fe7\u6240\u793a\uff0c\u5728\u6709\u76d1\u7763\u9884\u8bad\u7ec3\u4e0e\u5bf9\u6bd4\u5b66\u4e60\u9884\u8bad\u7ec3\u7b97\u6cd5\u4e2d\uff0c\u7f51\u7edc\u7684\u8f83\u6df1\u5c42\u6ce8\u610f\u529b\u6a21\u5757\u91cc\uff0c\u4e0d\u540c\u6ce8\u610f\u529b\u5934\u5173\u6ce8\u7684\u4fe1\u606f\u662f\u8d8b\u540c\u7684\uff0c\u800c\u5728 MIM \u4e2d\uff0c\u4e0d\u540c\u6ce8\u610f\u529b\u5934\u5173\u6ce8\u7684\u4fe1\u606f\u66f4\u591a\u6837\uff0c\u8fd9\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u907f\u514d\u4e86\u6a21\u578b\u584c\u9677\uff08model collapse\uff09\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"239\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-8-1024x239.png\" alt=\"\u56fe7\uff1a\u4e0d\u540c\u6a21\u578b\u4e2d\u6ce8\u610f\u529b\u5934\u5bf9\u5e94\u7684\u6ce8\u610f\u529b\u5730\u56fe\uff08attention map\uff09\u7684\u591a\u6837\u6027\u5206\u6790\" class=\"wp-image-1072530\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-8-1024x239.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-8-300x70.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-8-768x179.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-8-1536x358.png 1536w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-8-240x56.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-8.png 1990w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe7\uff1a\u4e0d\u540c\u6a21\u578b\u4e2d\u6ce8\u610f\u529b\u5934\u5bf9\u5e94\u7684\u6ce8\u610f\u529b\u5730\u56fe\uff08attention map\uff09\u7684\u591a\u6837\u6027\u5206\u6790<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45404\"><\/p>\n\n\n\n<p>3\uff09 MIM \u5bf9\u8bed\u4e49\u4fe1\u606f\u7684\u523b\u753b\u8f83\u5c11\uff0c\u4f46\u662f\u5bf9\u51e0\u4f55\u4fe1\u606f\u7684\u523b\u753b\u8f83\u591a\u3002\u56e0\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u5bf9\u6bd4\u4e86\u76d1\u7763\u5b66\u4e60\u4e0e MIM \u5728\u8bed\u4e49\u5206\u7c7b\u4efb\u52a1\u4e2d\u548c\u51e0\u4f55\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\u8868\u73b0\u3002\u7ed3\u679c\u663e\u793a MIM \u5728\u8bed\u4e49\u5206\u7c7b\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\u8868\u73b0\u8f83\u5dee\uff0c\u4f46\u662f\u5728\u51e0\u4f55\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\u8868\u73b0\u8f83\u597d\uff0c\u5982\u88682\u6240\u793a\u3002\u540c\u65f6\uff0c\u7814\u7a76\u5458\u4eec\u8fd8\u8003\u5bdf\u4e86\u5728\u6df7\u5408\u4efb\u52a1\uff08\u5982\u7269\u4f53\u68c0\u6d4b\uff09\u4e2d\uff0cMIM \u4e0e\u6709\u76d1\u7763\u9884\u8bad\u7ec3\u5728\u5206\u7c7b\u4e0e\u5b9a\u4f4d\u4e24\u4e2a\u5b50\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u53d8\u5316\u60c5\u51b5\uff08\u5982\u56fe8\u6240\u793a\uff09\u3002\u7ed3\u679c\u4e5f\u663e\u793a MIM \u5728\u5206\u7c7b\u4efb\u52a1\u4e0a\u7684\u6536\u655b\u901f\u5ea6\u6bd4\u6709\u76d1\u7763\u9884\u8bad\u7ec3\u5dee\uff0c\u4f46\u662f\u5728\u5b9a\u4f4d\u4efb\u52a1\u4e0a\u6536\u655b\u6027\u66f4\u597d\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"940\" height=\"266\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-9.png\" alt=\"\u88682\uff1a\uff08\u5de6\uff09\u8bed\u4e49\u5206\u7c7b\u4efb\u52a1\u7684\u6027\u80fd\u6bd4\u8f83\uff1b\uff08\u53f3\uff09\u51e0\u4f55\u4efb\u52a1\u7684\u6027\u80fd\u6bd4\u8f83\" class=\"wp-image-1072533\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-9.png 940w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-9-300x85.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-9-768x217.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-9-240x68.png 240w\" sizes=\"auto, (max-width: 940px) 100vw, 940px\" \/><figcaption class=\"wp-element-caption\">\u88682\uff1a\uff08\u5de6\uff09\u8bed\u4e49\u5206\u7c7b\u4efb\u52a1\u7684\u6027\u80fd\u6bd4\u8f83\uff1b\uff08\u53f3\uff09\u51e0\u4f55\u4efb\u52a1\u7684\u6027\u80fd\u6bd4\u8f83<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45405\"><\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"326\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-10-1024x326.png\" alt=\"\u56fe8\uff1a\u7269\u4f53\u68c0\u6d4b\u4efb\u52a1\u4e2d\u7684\u5206\u7c7b\u635f\u5931\u4e0e\u5b9a\u4f4d\u635f\u5931\" class=\"wp-image-1072536\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-10-1024x326.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-10-300x96.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-10-768x245.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-10-240x76.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-10.png 1074w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe8\uff1a\u7269\u4f53\u68c0\u6d4b\u4efb\u52a1\u4e2d\u7684\u5206\u7c7b\u635f\u5931\u4e0e\u5b9a\u4f4d\u635f\u5931<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45406\"><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u6269\u5c55-mim-\u5728\u5c0f\u6a21\u578b\u4e0a\u7684\u6709\u6548\u6027\">\u6269\u5c55 MIM \u5728\u5c0f\u6a21\u578b\u4e0a\u7684\u6709\u6548\u6027<\/h2>\n\n\n\n<p>\u5728\u5173\u4e8e MIM \u7684\u65e9\u671f\u8bba\u6587\u4e2d\uff0c\u79d1\u7814\u4eba\u5458\u666e\u904d\u53d1\u73b0 MIM \u65b9\u6cd5\u5bf9\u5927\u6a21\u578b\u66f4\u52a0\u53cb\u597d\uff0c\u76f4\u63a5\u5728\u5c0f\u6a21\u578b\u4e2d\u4f7f\u7528\u7684\u6709\u6548\u6027\u6b20\u4f73\u3002\u5982\u88683\u6240\u793a\uff0c\u5728 ViT-T \u7b49\u8f83\u5c0f\u7684\u6a21\u578b\u4e2d\u4f7f\u7528 MIM \u9884\u8bad\u7ec3\u7b97\u6cd5\uff0c\u5176\u6027\u80fd\u751a\u81f3\u843d\u540e\u4e8e\u968f\u673a\u521d\u59cb\u5316\u7684\u6a21\u578b\u3002\u5982\u4f55\u5c06 MIM \u5e94\u7528\u4e8e\u5c0f\u6a21\u578b\uff0c\u662f\u9886\u57df\u4e2d\u4e00\u4e2a\u91cd\u8981\u7684\u5f00\u653e\u6027\u95ee\u9898\u3002\u5728\u53e6\u5916\u4e00\u7bc7\u5165\u9009 CVPR 2023 \u7684\u5de5\u4f5c\u201cTinyMIM: An Empirical Study of Distilling MIM Pre-trained Models\u201d\u4e2d\uff08\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2301.01296\uff09\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u901a\u8fc7\u84b8\u998f\uff08distillation\uff09\u6280\u672f\uff0c\u6210\u529f\u5c06 MIM \u6a21\u578b\u7684\u4f18\u52bf\u62d3\u5c55\u5230\u4e86\u5c0f\u6a21\u578b\u4e2d\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"720\" height=\"258\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-11.png\" alt=\"\u88683\uff1a\u4e0d\u540c\u5927\u5c0f\u6a21\u578b\u4e0b\uff0c\u4f7f\u7528 MIM \u9884\u8bad\u7ec3\u4e0e\u968f\u673a\u521d\u59cb\u5316\u6a21\u578b\u7684\u6027\u80fd\u6bd4\u8f83\" class=\"wp-image-1072539\" style=\"width:555px;height:auto\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-11.png 720w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-11-300x108.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-11-240x86.png 240w\" sizes=\"auto, (max-width: 720px) 100vw, 720px\" \/><figcaption class=\"wp-element-caption\">\u88683\uff1a\u4e0d\u540c\u5927\u5c0f\u6a21\u578b\u4e0b\uff0c\u4f7f\u7528 MIM \u9884\u8bad\u7ec3\u4e0e\u968f\u673a\u521d\u59cb\u5316\u6a21\u578b\u7684\u6027\u80fd\u6bd4\u8f83<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45407\"><\/p>\n\n\n\n<p>TinyMIM \u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u7cfb\u7edf\u6027\u5730\u7814\u7a76\u4e86\u5982\u4f55\u4f7f\u7528\u7ecf\u8fc7 MIM \u9884\u8bad\u7ec3\u7684\u6a21\u578b\u84b8\u998f\u81f3\u5c0f\u6a21\u578b\u4e2d\u3002\u5176\u7814\u7a76\u5bf9\u8c61\u5305\u62ec\u8f93\u5165\u5f62\u5f0f\uff0c\u84b8\u998f\u5bf9\u8c61\uff0c\u4ee5\u53ca\u84b8\u998f\u65b9\u6cd5\u4e09\u4e2a\u65b9\u9762\u3002\u901a\u8fc7\u5e7f\u6cdb\u7684\u5b9e\u9a8c\uff0c\u7814\u7a76\u4eba\u5458\u53d1\u73b0\uff1a<\/p>\n\n\n\n<p>1\uff09 \u76f4\u63a5\u84b8\u998f\u5143\u7d20\u95f4\u7684\u5173\u7cfb\uff08relation\uff09\u662f MIM \u4e2d\u6700\u6709\u6548\u7684\u84b8\u998f\u65b9\u5f0f\uff0c\u5176\u6027\u80fd\u53ef\u4ee5\u6bd4\u84b8\u998f CLS Token \u5728 ViT-T \u4e0a\u597d4.2 Top-1 Acc\uff0c\u5728 ViT-B \u4e0a\u597d1.6 Top-1 Acc.<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"842\" height=\"802\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-12.png\" alt=\"\u88684\uff1a\u4e0d\u540c\u84b8\u998f\u5bf9\u8c61\u5bf9\u7ed3\u679c\u7684\u5f71\u54cd\" class=\"wp-image-1072542\" style=\"width:465px;height:auto\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-12.png 842w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-12-300x286.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-12-768x732.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-12-189x180.png 189w\" sizes=\"auto, (max-width: 842px) 100vw, 842px\" \/><figcaption class=\"wp-element-caption\">\u88684\uff1a\u4e0d\u540c\u84b8\u998f\u5bf9\u8c61\u5bf9\u7ed3\u679c\u7684\u5f71\u54cd<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45408\"><\/p>\n\n\n\n<p>2\uff09 \u5728\u84b8\u998f\u65f6\u5f15\u5165 MIM \u4efb\u52a1\u4f1a\u635f\u5bb3\u6027\u80fd\u3002\u5982\u88685\u6240\u793a\uff0c\u4f7f\u7528\u63a9\u7801\u56fe\u50cf\u4f5c\u4e3a\u8f93\u5165\uff0c\u4ee5\u53ca\u5728\u84b8\u998f\u65f6\u5f15\u5165\u56fe\u50cf\u91cd\u6784\u4efb\u52a1\uff0c\u90fd\u4f1a\u635f\u5bb3\u6a21\u578b\u7684\u84b8\u998f\u6548\u679c\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"830\" height=\"242\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-13.png\" alt=\"\u88685\uff1a\u84b8\u998f\u65f6\u5f15\u5165 MIM \u4efb\u52a1\u5bf9\u6027\u80fd\u7684\u5f71\u54cd\u3002\" class=\"wp-image-1072545\" style=\"width:481px;height:auto\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-13.png 830w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-13-300x87.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-13-768x224.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-13-240x70.png 240w\" sizes=\"auto, (max-width: 830px) 100vw, 830px\" \/><figcaption class=\"wp-element-caption\">\u88685\uff1a\u84b8\u998f\u65f6\u5f15\u5165 MIM \u4efb\u52a1\u5bf9\u6027\u80fd\u7684\u5f71\u54cd\u3002<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45409\"><\/p>\n\n\n\n<p>3\uff09 \u5e8f\u5217\u5316\u84b8\u998f\u53ef\u4ee5\u8fdb\u4e00\u6b65\u964d\u4f4e\u96be\u5ea6\uff0c\u63d0\u5347\u6027\u80fd\u3002\u5e8f\u5217\u5316\u84b8\u998f\u6307\u7684\u662f\u4f7f\u7528\u5c0f\u6a21\u578b\u84b8\u998f\u5927\u6a21\u578b\u7684\u8fc7\u7a0b\u4e2d\uff0c\u5f15\u5165\u4e2d\u7b49\u89c4\u6a21\u6a21\u578b\u8fdb\u884c\u84b8\u998f\uff0c\u5373\u5148\u84b8\u998f\u51fa\u4e00\u4e2a\u4e2d\u7b49\u5927\u5c0f\u7684\u6a21\u578b\uff0c\u518d\u5229\u7528\u8be5\u6a21\u578b\u53bb\u84b8\u998f\u5c0f\u6a21\u578b\u3002\u8fd9\u6837\u7684\u84b8\u998f\u65b9\u5f0f\u53ef\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\uff0c\u5982\u88686\u6240\u793a\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"826\" height=\"370\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-14.png\" alt=\"\u88686\uff1a\u5e8f\u5217\u5316\u84b8\u998f\u5bf9\u7ed3\u679c\u7684\u5f71\u54cd\" class=\"wp-image-1072548\" style=\"width:512px;height:auto\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-14.png 826w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-14-300x134.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-14-768x344.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-14-240x108.png 240w\" sizes=\"auto, (max-width: 826px) 100vw, 826px\" \/><figcaption class=\"wp-element-caption\">\u88686\uff1a\u5e8f\u5217\u5316\u84b8\u998f\u5bf9\u7ed3\u679c\u7684\u5f71\u54cd<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45410\"><\/p>\n\n\n\n<p>\u7ed3\u5408\u4e0a\u8ff0\u53d1\u73b0\uff0cTinyMIM \u5728\u4e00\u7cfb\u5217\u4e2d\u5c0f\u578b\u6a21\u578b\u4e2d\u5747\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\uff0c\u76f8\u8f83\u4e8e\u5176\u4ed6\u76f4\u63a5\u8bad\u7ec3\u7684\u5c0f\u6a21\u578b\uff0c\u5982 MobileViT \u7b49\uff0c\u4e5f\u53d6\u5f97\u4e86\u66f4\u597d\u7684\u4e0b\u6e38\u4efb\u52a1\u8fc1\u79fb\u80fd\u529b\uff0c\u5982\u56fe9\u6240\u793a\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"346\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-15-1024x346.png\" alt=\"\u56fe9\uff1aTinyMIM \u76f8\u8f83\u4e8e MAE \u4e0e\u5176\u4ed6\u5c0f\u6a21\u578b\u8bbe\u8ba1\u65b9\u6cd5\u5747\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u4f18\u52bf\" class=\"wp-image-1072551\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-15-1024x346.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-15-300x101.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-15-768x259.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-15-240x81.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/cvpr-2023-mim-15.png 1332w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe9\uff1aTinyMIM \u76f8\u8f83\u4e8e MAE \u4e0e\u5176\u4ed6\u5c0f\u6a21\u578b\u8bbe\u8ba1\u65b9\u6cd5\u5747\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u4f18\u52bf<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45411\"><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u5c55\u671b\u672a\u6765\">\u5c55\u671b\u672a\u6765<\/h2>\n\n\n\n<p>\u968f\u7740\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u9884\u8bad\u7ec3\u8303\u5f0f\u4ece\u6709\u76d1\u7763\u5b66\u4e60\u9010\u6e10\u6f14\u53d8\u81f3\u81ea\u76d1\u7763\u5b66\u4e60\uff0c\u79d1\u7814\u4eba\u5458\u5bf9\u89c6\u89c9\u667a\u80fd\u7684\u8ba4\u8bc6\u4e0e\u7406\u89e3\u4e5f\u5728\u4e0d\u65ad\u6539\u53d8\u4e0e\u6df1\u5316\u3002\u73b0\u9636\u6bb5\uff0c\u57fa\u4e8e\u63a9\u7801\u56fe\u50cf\u5efa\u6a21\uff08MIM\uff09\u7684\u56fe\u50cf\u9884\u8bad\u7ec3\u7b97\u6cd5\u5df2\u7ecf\u5c55\u73b0\u51fa\u65e0\u76d1\u7763\u9884\u8bad\u7ec3\u5f3a\u5927\u7684\u6f5c\u529b\uff0c\u4f46\u662f\u5426\u5b58\u5728\u66f4\u9002\u5408\u89c6\u89c9\u4fe1\u53f7\u7684\u9884\u8bad\u7ec3\u65b9\u6cd5\u4ecd\u7136\u662f\u9886\u57df\u5185\u6700\u91cd\u8981\u7684\u5f00\u653e\u95ee\u9898\u4e4b\u4e00\u3002\u6b64\u5916\uff0c\u5728\u89c6\u89c9\u4e0e\u8bed\u8a00\u5927\u4e00\u7edf\u7684\u53d1\u5c55\u8d8b\u52bf\u4e4b\u4e0b\uff0c\u5982\u4f55\u6709\u6548\u5229\u7528\u63a9\u7801\u4fe1\u53f7\u5efa\u6a21\u7b49\u9884\u8bad\u7ec3\u7b97\u6cd5\u9ad8\u6548\u8fde\u63a5\u8bed\u8a00\u4e0e\u89c6\u89c9\u4fe1\u53f7\u7684\u95ee\u9898\u4e5f\u4ecd\u9700\u63a2\u7d22\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5e0c\u671b\u968f\u7740\u5bf9\u63a9\u7801\u56fe\u50cf\u5efa\u6a21\u9884\u8bad\u7ec3\u7b97\u6cd5\u7406\u89e3\u4e0e\u8ba4\u8bc6\u7684\u6df1\u5316\uff0c\u7814\u7a76\u5e76\u63d0\u51fa\u66f4\u9ad8\u6548\u7684\u9884\u8bad\u7ec3\u7b97\u6cd5\uff0c\u4fc3\u8fdb\u89c6\u89c9\u667a\u80fd\u8fc8\u5165\u4e0b\u4e00\u4e2a\u53d1\u5c55\u65b0\u9636\u6bb5\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7f16\u8005\u6309\uff1a\u63a9\u7801\u56fe\u50cf\u5efa\u6a21\uff08Masked Image Modeling, MIM\uff09\u7684\u63d0\u51fa\uff0c\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u6a21\u578b\u8bad\u7ec3\u5f15\u5165\u65e0\u76d1\u7763\u5b66\u4e60\u505a\u51fa\u4e86\u91cd\u8981\u8d21\u732e\u3002\u5f97\u76ca\u4e8e MIM \u7684\u9884\u8bad\u7ec3\u7b97\u6cd5\uff0c\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5728\u8fd1\u5e74\u6765\u6301\u7eed\u8f93\u51fa\u7740\u4f18\u8d28\u7684\u7814\u7a76\u6210\u679c\u3002\u7136\u800c\u6574\u4e2a\u4e1a\u754c\u5bf9 MIM \u673a\u5236\u7684\u7814\u7a76\u4ecd\u5b58\u5728\u4e0d\u8db3\u3002 \u79c9\u6301\u7740\u4e0d\u65ad\u6269\u5c55\u524d\u6cbf\u6280\u672f\u8fb9\u754c\u7684\u63a2\u7d22\u7cbe\u795e\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5728\u7406\u89e3 MIM \u4f5c\u7528\u673a\u5236\uff0c\u4ee5\u53ca\u57fa\u4e8e\u8fd9\u4e9b\u673a\u5236\u63d0\u5347\u73b0\u6709 MIM \u7b97\u6cd5\u7684\u9886\u57df\uff0c\u53d6\u5f97\u4e86\u4e00\u7cfb\u5217\u7684\u521b\u65b0\u6210\u679c\uff0c\u5e76\u83b7\u5f97\u4e86 CVPR 2023 \u7684\u8ba4\u53ef\u3002\u8fd9\u4e9b\u6210\u679c\u5305\u542b\uff1a\u57fa\u4e8e MIM \u9884\u8bad\u7ec3\u65b9\u6cd5\u7684\u6269\u5c55\u6cd5\u5219\u7814\u7a76\u3001\u5206\u6790 MIM \u7684\u5177\u4f53\u6027\u8d28\u4ee5\u53ca\u6709\u6548\u6027\u80cc\u540e\u7684\u539f\u56e0\u3001\u901a\u8fc7\u84b8\u998f\u6280\u672f\u5c06 MIM \u6a21\u578b\u7684\u4f18\u52bf\u62d3\u5c55\u5230\u5c0f\u6a21\u578b\u4e2d\u3002 \u6b63\u503c CVPR 2023 \u5927\u4f1a\u4e3e\u529e\u4e4b\u9645\uff0c\u8ba9\u6211\u4eec\u4e00\u8d77\u4e86\u89e3 MIM \u52a9\u63a8\u8ba1\u7b97\u673a\u89c6\u89c9\u7814\u7a76\u8fdb\u5165\u52a0\u901f\u9053\u7684\u65b0\u53d1\u73b0\u5427\uff01 \u9884\u8bad\u7ec3-\u5fae\u8c03\uff08Pre-training and Fine-tuning\uff09\u662f\u8fc7\u53bb\u5341\u5e74\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u6700\u91cd\u8981\u7684\u5b66\u4e60\u8303\u5f0f\u4e4b\u4e00\uff0c\u5176\u57fa\u672c\u60f3\u6cd5\u662f\u5728\u6d77\u91cf\u6570\u636e\u7684\u4efb\u52a1\u4e2d\uff0c\u5bf9\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u8bad\u7ec3\uff0c\u7136\u540e\u518d\u5c06\u9884\u8bad\u7ec3\u8fc7\u7684\u6a21\u578b\u5728\u4e0b\u6e38\u6570\u636e\u91cf\u8f83\u5c11\u7684\u4efb\u52a1\u4e2d\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u79cd\u65b9\u5f0f\u80fd\u591f\u5c06\u4e0a\u6e38\u5927\u6570\u636e\u4efb\u52a1\u4e2d\u5b66\u5230\u7684\u4fe1\u606f\u8fc1\u79fb\u81f3\u4e0b\u6e38\u6570\u636e\u91cf\u8f83\u5c11\u7684\u4efb\u52a1\u4e0a\uff0c\u7f13\u89e3\u6570\u636e\u91cf\u4e0d\u8db3\u7684\u95ee\u9898\uff0c\u5e76\u663e\u8457\u63d0\u5347\u6a21\u578b\u7684\u6027\u80fd\u3002 \u9884\u8bad\u7ec3-\u5fae\u8c03\u8303\u5f0f\u7684\u6210\u529f\uff0c\u6e90\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5341\u5e74\u6765\u9884\u8bad\u7ec3\u7b97\u6cd5\u7684\u505c\u6ede\u3002\u81ea2012\u5e74 AlexNet \u63d0\u51fa\u4ee5\u6765\uff0c\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u7684\u9884\u8bad\u7ec3\u7b97\u6cd5\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u88ab\u7b49\u4ef7\u4e8e\u4ee5 ImageNet \u6570\u636e\u96c6\u4e3a\u4ee3\u8868\u7684\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u3002\u5c3d\u7ba1\u56fe\u50cf\u5206\u7c7b\u7684\u6570\u636e\u6807\u6ce8\u6210\u672c\u5df2\u7136\u8f83\u4f4e\uff0c\u4f46\u540e\u7eed\u7684\u6570\u636e\u6e05\u6d17\u3001\u8d28\u91cf\u63a7\u5236\u7b49\u6b65\u9aa4\u4ecd\u5bf9\u6269\u5c55\u56fe\u50cf\u5206\u7c7b\u6570\u636e\u4ea7\u751f\u4e86\u6311\u6218\uff0c\u800c\u6570\u636e\u4e0d\u8db3\u7684\u56f0\u96be\u4e5f\u9650\u5236\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u6a21\u578b\u7684\u8fdb\u4e00\u6b65\u6269\u5927\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u4f7f\u7528\u65e0\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\u8fdb\u884c\u89c6\u89c9\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u9010\u6e10\u6210\u4e3a\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u4e2d\u7684\u6838\u5fc3\u95ee\u9898\u3002 2021\u5e746\u6708\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u7684 BEiT \u65b9\u6cd5\uff0c\u901a\u8fc7\u5f15\u5165\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u4e2d\u7684\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\uff08Masked Language Modeling, MLM\uff09\u7b97\u6cd5\uff0c\u6210\u529f\u5730\u8bc1\u660e\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u65e0\u76d1\u7763\u9884\u8bad\u7ec3\u53ef\u4ee5\u8fbe\u5230\u4e0e\u6709\u76d1\u7763\u9884\u8bad\u7ec3\u76f8\u540c\u751a\u81f3\u66f4\u597d\u7684\u6548\u679c\u30022021\u5e7411\u6708\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u7684 SimMIM \u4e0e Meta \u63d0\u51fa\u7684 MAE \u8fdb\u4e00\u6b65\u7b80\u5316\u4e86 BEiT\uff0c\u5e76\u63d0\u5347\u4e86\u7b97\u6cd5\u6027\u80fd\u3002\u81ea\u6b64\uff0c\u63a9\u7801\u56fe\u50cf\u5efa\u6a21\uff08Masked Image Modeling, MIM\uff09\u7684\u7814\u7a76\u8303\u5f0f\u6b63\u5f0f\u5f00\u542f\u3002 \u867d\u7136\u57fa\u4e8e MIM \u7684\u9884\u8bad\u7ec3\u7b97\u6cd5\u7684\u6210\u679c\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5185\u767e\u82b1\u9f50\u653e\uff0c\u4f46\u5bf9 [&hellip;]<\/p>\n","protected":false},"author":34512,"featured_media":1072506,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1072398","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072398","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":3,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072398\/revisions"}],"predecessor-version":[{"id":1072554,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072398\/revisions\/1072554"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media\/1072506"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1072398"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1072398"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1072398"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1072398"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}