{"id":34522,"date":"2024-12-08T06:56:36","date_gmt":"2024-12-07T22:56:36","guid":{"rendered":"https:\/\/17aitech.com\/?p=34522"},"modified":"2024-12-08T06:56:36","modified_gmt":"2024-12-07T22:56:36","slug":"icml-2024-%e6%a2%af%e5%ba%a6%e6%a3%80%e6%9f%a5%e7%82%b9%e5%a4%aa%e6%85%a2%ef%bc%9f%e4%b8%8d%e9%99%8d%e9%80%9f%e3%80%81%e7%9c%81%e6%98%be%e5%ad%98%ef%bc%8clowmemorybp%e5%a4%a7%e5%b9%85%e6%8f%90","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34522","title":{"rendered":"ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-12-5\" target=\"_blank\">ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1952fb3d985627bbc9cfde0d1aa18f59.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1952fb3d985627bbc9cfde0d1aa18f59.png\"><\/a><\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>AIxiv\u4e13\u680f\u662f<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/section>\n<\/section>\n<\/blockquote>\n<p><strong>\u672c\u6587\u8bba\u6587\u4e00\u4f5c\u662f\u5357\u5f00\u5927\u5b66\u7edf\u8ba1\u4e0e<mark data-type=\"concepts\" data-id=\"3946938d-2b2f-4b48-9ee4-58debad2a6c4\">\u6570\u636e\u79d1\u5b66<\/mark>\u5b66\u9662\u7814\u4e8c\u7855\u58eb\u751f\u6768\u96e8\u8fb0\uff0c\u6307\u5bfc\u8001\u5e08\u4e3a\u5357\u5f00\u5927\u5b66\u7edf\u8ba1\u4e0e<mark data-type=\"concepts\" data-id=\"3946938d-2b2f-4b48-9ee4-58debad2a6c4\">\u6570\u636e\u79d1\u5b66<\/mark>\u5b66\u9662\u7684\u5f90\u541b\u526f\u6559\u6388\u3002\u5f90\u541b\u8001\u5e08\u56e2\u961f\u7684\u7814\u7a76\u91cd\u70b9\u662f<mark data-type=\"tech_tasks\" data-id=\"6e614199-9e49-450e-9078-61fb2b122da9\">\u8ba1\u7b97\u673a\u89c6\u89c9<\/mark>\u3001\u751f\u6210\u5f0f AI \u548c\u9ad8\u6548<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\uff0c\u5e76\u5728\u9876\u7ea7\u4f1a\u8bae\u548c\u671f\u520a\u4e0a\u53d1\u8868\u4e86\u591a\u7bc7\u8bba\u6587\uff0c\u8c37\u6b4c\u5b66\u672f\u5f15\u7528\u8d85\u8fc7 4700 \u6b21\u3002<\/strong><\/p>\n<p>\u81ea\u4ece\u5927\u578b Transformer \u6a21\u578b\u9010\u6e10\u6210\u4e3a\u5404\u4e2a\u9886\u57df\u7684\u7edf\u4e00\u67b6\u6784\uff0c\u5fae\u8c03\u5c31\u6210\u4e3a\u4e86\u5c06\u9884\u8bad\u7ec3\u5927\u6a21\u578b\u5e94\u7528\u5230\u4e0b\u6e38\u4efb\u52a1\u7684\u91cd\u8981\u624b\u6bb5\u3002\u7136\u800c\uff0c\u7531\u4e8e\u6a21\u578b\u7684\u5c3a\u5bf8\u65e5\u76ca\u589e\u5927\uff0c\u5fae\u8c03\u6240\u9700\u8981\u7684\u663e\u5b58\u4e5f\u9010\u6e10\u589e\u52a0\uff0c\u5982\u4f55\u9ad8\u6548\u5730\u964d\u4f4e\u5fae\u8c03\u663e\u5b58\u5c31\u6210\u4e86\u4e00\u4e2a\u91cd\u8981\u7684\u95ee\u9898\u3002\u6b64\u524d\uff0c\u5fae\u8c03 Transformer \u6a21\u578b\u65f6\uff0c\u4e3a\u4e86\u8282\u7701\u663e\u5b58\u5f00\u9500\uff0c\u901a\u5e38\u7684\u505a\u6cd5\u662f\u4f7f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9\uff08gradient checkpointing\uff0c\u4e5f\u53eb\u4f5c\u6fc0\u6d3b\u91cd\u7b97\uff09\uff0c\u4ee5\u727a\u7272\u8bad\u7ec3\u901f\u5ea6\u4e3a\u4ee3\u4ef7\u964d\u4f4e\u53cd\u5411\u4f20\u64ad\uff08Backpropagation, BP\uff09\u8fc7\u7a0b\u4e2d\u7684\u6fc0\u6d3b\u663e\u5b58\u5360\u7528\u3002<\/p>\n<p>\u6700\u8fd1\uff0c\u7531\u5357\u5f00\u5927\u5b66\u7edf\u8ba1\u4e0e<mark data-type=\"concepts\" data-id=\"3946938d-2b2f-4b48-9ee4-58debad2a6c4\">\u6570\u636e\u79d1\u5b66<\/mark>\u5b66\u9662\u5f90\u541b\u8001\u5e08\u56e2\u961f\u53d1\u8868\u5728 ICML 2024 \u4e0a\u7684\u8bba\u6587\u300aReducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation\u300b\u63d0\u51fa\u901a\u8fc7\u66f4\u6539\u53cd\u5411\u4f20\u64ad\uff08BP\uff09\u8fc7\u7a0b\uff0c\u5728\u4e0d\u589e\u52a0\u8ba1\u7b97\u91cf\u7684\u60c5\u51b5\u4e0b\uff0c\u663e\u8457\u51cf\u5c11\u5cf0\u503c\u6fc0\u6d3b\u663e\u5b58\u5360\u7528\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-f58ac08064af1a01a6f0c5414df52cc2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-f58ac08064af1a01a6f0c5414df52cc2.png\"><\/a><\/p>\n<ul>\n<li>\n<p>\u8bba\u6587\uff1aReducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2406.16282<\/p>\n<\/li>\n<li>\n<p>\u9879\u76ee\u94fe\u63a5\uff1ahttps:\/\/github.com\/yyyyychen\/LowMemoryBP<\/p>\n<\/li>\n<\/ul>\n<p>\u6587\u7ae0\u63d0\u51fa\u4e86\u4e24\u79cd\u53cd\u5411\u4f20\u64ad\u6539\u8fdb\u7b56\u7565\uff0c\u5206\u522b\u662f Approximate Backpropagation\uff08Approx-BP\uff09\u548c Memory-Sharing Backpropagation\uff08MS-BP\uff09\u3002Approx-BP \u548c MS-BP \u5206\u522b\u4ee3\u8868\u4e86\u4e24\u79cd\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u4e2d\u5185\u5b58\u6548\u7387\u7684\u65b9\u6848\uff0c\u53ef\u4ee5\u5c06\u5176\u7edf\u79f0\u4e3a LowMemoryBP\u3002\u65e0\u8bba\u662f\u5728\u7406\u8bba\u8fd8\u662f\u5b9e\u8df5\u610f\u4e49\u4e0a\uff0c\u6587\u7ae0\u90fd\u5bf9\u66f4\u9ad8\u6548\u7684\u53cd\u5411\u4f20\u64ad\u8bad\u7ec3\u63d0\u4f9b\u4e86\u5f00\u521b\u6027\u7684\u6307\u5bfc\u3002<\/p>\n<p>\u5728\u7406\u8bba\u663e\u5b58\u5206\u6790\u4e2d\uff0cLowMemoryBP \u53ef\u4ee5\u5927\u5e45\u964d\u4f4e\u6765\u81ea<mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\u548c\u6807\u51c6\u5316\u5c42\u7684\u6fc0\u6d3b\u663e\u5b58\u5360\u7528\uff0c\u4ee5 ViT \u548c LLaMA \u4e3a\u4f8b\uff0c\u53ef\u4ee5\u5bf9 ViT \u5fae\u8c03\u964d\u4f4e 39.47% \u7684\u6fc0\u6d3b\u663e\u5b58\uff0c\u53ef\u4ee5\u5bf9 LLaMA \u5fae\u8c03\u964d\u4f4e 29.19% \u7684\u6fc0\u6d3b\u663e\u5b58\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-de369d8f2348fa4bb202ea98f8ff2976.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-de369d8f2348fa4bb202ea98f8ff2976.png\"><\/a><\/p>\n<p>\u5728\u5b9e\u9645\u5b9e\u9a8c\u4e2d\uff0cLowMemoryBP \u53ef\u4ee5\u6709\u6548\u5730\u4f7f\u5305\u62ec ViT, LLaMA, RoBERTa, BERT, Swin \u5728\u5185\u7684 Transformer \u6a21\u578b\u5fae\u8c03\u5cf0\u503c\u663e\u5b58\u5360\u7528\u964d\u4f4e 20%~30%\uff0c\u5e76\u4e14\u4e0d\u4f1a\u5e26\u6765\u8bad\u7ec3\u541e\u5410\u91cf\u548c\u6d4b\u8bd5\u7cbe\u5ea6\u7684\u635f\u5931\u3002<\/p>\n<p><strong>Approx-BP<\/strong><\/p>\n<p>\u5728\u4f20\u7edf\u53cd\u5411\u4f20\u64ad\u8bad\u7ec3\u4e2d\uff0c<mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\u68af\u5ea6\u7684\u53cd\u5411\u56de\u4f20\u662f\u4e25\u683c\u5bf9\u5e94\u5176\u5bfc\u51fd\u6570\u7684\uff0c\u5bf9\u4e8e Transformer \u6a21\u578b\u4e2d\u5e38\u7528\u7684 GELU \u548c SiLU \u51fd\u6570\uff0c\u8fd9\u610f\u5473\u7740\u9700\u8981\u5c06\u8f93\u5165\u7279\u5f81<mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u5b8c\u6574\u5730\u5b58\u5165\u6fc0\u6d3b\u663e\u5b58\u4e2d\u3002\u800c\u672c\u6587\u7684\u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u5957\u53cd\u5411\u4f20\u64ad\u8fd1\u4f3c\u7406\u8bba\uff0c\u5373 Approx-BP \u7406\u8bba\u3002\u5728\u8be5\u7406\u8bba\u7684\u6307\u5bfc\u4e0b\uff0c\u4f5c\u8005\u4f7f\u7528\u5206\u6bb5\u7ebf\u6027\u51fd\u6570\u903c\u8fd1<mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\uff0c\u5e76\u7528\u5206\u6bb5\u7ebf\u6027\u51fd\u6570\u7684<mark data-type=\"concepts\" data-id=\"59a40d6c-220d-4a14-847c-0156b45a2f74\">\u5bfc\u6570<\/mark>\uff08\u9636\u68af\u51fd\u6570\uff09\u66ff\u4ee3 GELU\/SiLU \u68af\u5ea6\u7684\u53cd\u5411\u56de\u4f20\u3002\u8fd9\u4e2a\u65b9\u6cd5\u5bfc\u51fa\u4e86\u4e24\u4e2a\u975e\u5bf9\u79f0\u7684\u5185\u5b58\u9ad8\u6548<mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\uff1aReGELU2 \u548c ReSiLU2\u3002\u8fd9\u7c7b<mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\u7531\u4e8e\u4f7f\u7528 4 \u6bb5\u9636\u68af\u51fd\u6570\u8fdb\u884c\u53cd\u5411\u56de\u4f20\uff0c\u4ece\u800c\u4f7f\u5f97\u6fc0\u6d3b\u5b58\u50a8\u53ea\u9700\u8981\u4f7f\u7528 2bit \u6570\u636e\u7c7b\u578b\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-3cb3a8e56a0726ad16d52c9dde285e10.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-3cb3a8e56a0726ad16d52c9dde285e10.png\"><\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b9c7c78cdcc614a44a756e4d26548980.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b9c7c78cdcc614a44a756e4d26548980.png\"><\/a><\/p>\n<p><strong>MS-BP<\/strong><\/p>\n<p>BP \u7f51\u7edc\u6bcf\u4e00\u5c42\u901a\u5e38\u90fd\u4f1a\u5c06\u8f93\u5165<mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u5b58\u5165\u6fc0\u6d3b\u663e\u5b58\u4ee5\u7528\u4f5c\u53cd\u5411\u4f20\u64ad\u8ba1\u7b97\u3002\u4f5c\u8005\u6307\u51fa\u5982\u679c\u53ef\u4ee5\u5c06\u67d0\u4e00\u5c42\u7684\u53cd\u5411\u4f20\u64ad\u6539\u5199\u6210\u4f9d\u8d56\u8f93\u51fa\u7684\u5f62\u5f0f\uff0c\u90a3\u4e48\u8fd9\u4e00\u5c42\u548c\u540e\u4e00\u5c42\u5c31\u53ef\u4ee5\u5171\u4eab\u540c\u4e00\u4e2a\u6fc0\u6d3b<mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\uff0c\u4ece\u800c\u964d\u4f4e\u6fc0\u6d3b\u5b58\u50a8\u7684\u5197\u4f59\u3002<\/p>\n<p>\u800c\u6587\u7ae0\u6307\u51fa Transformer \u6a21\u578b\u4e2d\u5e38\u7528\u7684 LayerNorm \u548c RMSNorm\uff0c\u5728\u5c06\u4eff\u5c04<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u5408\u5e76\u5230\u540e\u4e00\u5c42\u7684\u7ebf\u6027\u5c42\u4e4b\u540e\uff0c\u53ef\u4ee5\u5f88\u597d\u5730\u7b26\u5408 MS-BP \u7b56\u7565\u7684\u8981\u6c42\u3002\u7ecf\u8fc7\u91cd\u65b0\u8bbe\u8ba1\u7684 MS-LayerNorm \u548c MS-RMSNorm \u4e0d\u518d\u4ea7\u751f\u72ec\u7acb\u7684\u6fc0\u6d3b\u663e\u5b58\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-f384899977c64bf773461f1dad03b0c6.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-f384899977c64bf773461f1dad03b0c6.png\"><\/a><\/p>\n<p><strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong><\/p>\n<p>\u4f5c\u8005\u5bf9<mark data-type=\"tech_tasks\" data-id=\"6e614199-9e49-450e-9078-61fb2b122da9\">\u8ba1\u7b97\u673a\u89c6\u89c9<\/mark>\u548c<mark data-type=\"tech_tasks\" data-id=\"c8ff5114-6cbb-49ca-8a89-3ee2826be0b4\">\u81ea\u7136\u8bed\u8a00\u5904\u7406<\/mark>\u9886\u57df\u7684\u82e5\u5e72\u4e2a\u4ee3\u8868\u6a21\u578b\u8fdb\u884c\u4e86\u5fae\u8c03\u5b9e\u9a8c\u3002\u5176\u4e2d\uff0c\u5728 ViT\uff0cLLaMA \u548c RoBERTa \u7684\u5fae\u8c03\u5b9e\u9a8c\u4e2d\uff0c\u6587\u7ae0\u63d0\u51fa\u7684\u65b9\u6cd5\u5206\u522b\u5c06\u5cf0\u503c\u663e\u5b58\u5360\u7528\u964d\u4f4e\u4e86 27%\uff0c29% \u548c 21%\uff0c\u5e76\u4e14\u6ca1\u6709\u5e26\u6765\u8bad\u7ec3\u6548\u679c\u548c\u8bad\u7ec3\u901f\u5ea6\u7684\u635f\u5931\u3002\u6ce8\u610f\u5230\uff0c\u4f5c\u4e3a\u5bf9\u6bd4\u7684 Mesa\uff08\u4e00\u4e2a 8-bit Activation Compressed Training \u65b9\u6cd5\uff09\u4f7f\u8bad\u7ec3\u901f\u5ea6\u964d\u4f4e\u4e86\u7ea6 20%\uff0c\u800c\u6587\u7ae0\u63d0\u51fa\u7684 LowMemoryBP \u65b9\u6cd5\u5219\u5b8c\u5168\u4fdd\u6301\u4e86\u8bad\u7ec3\u901f\u5ea6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-04b430169cc589c6aa839b903d045008.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-04b430169cc589c6aa839b903d045008.png\"><\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6660760c50cd351c6e2de28b4ea156d7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6660760c50cd351c6e2de28b4ea156d7.png\"><\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-983b135602619923c66c389cdb334a46.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-983b135602619923c66c389cdb334a46.png\"><\/a><\/p>\n<p><strong>\u7ed3\u8bba\u53ca\u610f\u4e49<\/strong><\/p>\n<p>\u6587\u7ae0\u63d0\u51fa\u7684\u4e24\u79cd BP \u6539\u8fdb\u7b56\u7565\uff0cApprox-BP \u548c MS-BP\uff0c\u5747\u5728\u4fdd\u6301\u8bad\u7ec3\u6548\u679c\u548c\u8bad\u7ec3\u901f\u5ea6\u7684\u540c\u65f6\uff0c\u5b9e\u73b0\u4e86\u6fc0\u6d3b\u663e\u5b58\u7684\u663e\u8457\u8282\u7701\u3002\u8fd9\u610f\u5473\u7740\u4ece BP \u539f\u7406\u4e0a\u8fdb\u884c\u4f18\u5316\u662f\u975e\u5e38\u6709\u524d\u666f\u7684\u663e\u5b58\u8282\u7701\u65b9\u6848\u3002\u6b64\u5916\uff0c\u6587\u7ae0\u63d0\u51fa\u7684 Approx-BP \u7406\u8bba\u7a81\u7834\u4e86\u4f20\u7edf<mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u7684\u4f18\u5316\u6846\u67b6\uff0c\u4e3a\u4f7f\u7528\u975e\u914d\u5bf9<mark data-type=\"concepts\" data-id=\"59a40d6c-220d-4a14-847c-0156b45a2f74\">\u5bfc\u6570<\/mark>\u63d0\u4f9b\u4e86\u7406\u8bba\u53ef\u884c\u6027\u3002\u5176\u5bfc\u51fa\u7684 ReGELU2 \u548c ReSiLU2 \u5c55\u73b0\u4e86\u8fd9\u4e00\u505a\u6cd5\u7684\u91cd\u8981\u5b9e\u8df5\u4ef7\u503c\u3002<\/p>\n<p>\u6b22\u8fce\u5927\u5bb6\u9605\u8bfb\u8bba\u6587\u6216\u8005\u4ee3\u7801\u53bb\u4e86\u89e3\u7b97\u6cd5\u7684\u8be6\u7ec6\u7ec6\u8282\uff0cLowMemoryBP \u9879\u76ee\u7684 github \u4ed3\u5e93\u4e0a\u5df2\u7ecf\u5f00\u6e90\u76f8\u5173\u7684\u6a21\u5757\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-12-5\" target=\"_blank\">ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:ICML 2024 | [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68,55,67,66],"class_list":["post-34522","post","type-post","status-publish","format-standard","hentry","category-news","tag-68","tag-55","tag-67","tag-66"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=34522\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34522\",\"url\":\"https:\/\/17aitech.com\/?p=34522\",\"name\":\"ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34522#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34522#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1952fb3d985627bbc9cfde0d1aa18f59.png\",\"datePublished\":\"2024-12-07T22:56:36+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34522#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34522\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34522#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1952fb3d985627bbc9cfde0d1aa18f59.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1952fb3d985627bbc9cfde0d1aa18f59.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34522#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=34522","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34522","url":"https:\/\/17aitech.com\/?p=34522","name":"ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34522#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34522#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1952fb3d985627bbc9cfde0d1aa18f59.png","datePublished":"2024-12-07T22:56:36+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34522#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34522"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34522#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1952fb3d985627bbc9cfde0d1aa18f59.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1952fb3d985627bbc9cfde0d1aa18f59.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34522#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"ICML 2024 | \u68af\u5ea6\u68c0\u67e5\u70b9\u592a\u6162\uff1f\u4e0d\u964d\u901f\u3001\u7701\u663e\u5b58\uff0cLowMemoryBP\u5927\u5e45\u63d0\u5347\u53cd\u5411\u4f20\u64ad\u663e\u5b58\u6548\u7387"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34522","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=34522"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34522\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34522"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34522"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34522"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}