{"id":41193,"date":"2023-03-11T17:32:02","date_gmt":"2023-03-11T09:32:02","guid":{"rendered":"https:\/\/swarma.org\/?p=41193"},"modified":"2023-03-11T17:32:02","modified_gmt":"2023-03-11T09:32:02","slug":"nat-mach-intell-%e9%80%9f%e9%80%92%ef%bc%9a%e5%a4%a7%e8%a7%84%e6%a8%a1%e9%a2%84%e8%ae%ad%e7%bb%83%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%9a%84%e5%8f%82%e6%95%b0%e6%9c%89%e6%95%88%e5%be%ae%e8%b0%83","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=41193","title":{"rendered":"Nat. Mach. Intell. \u901f\u9012\uff1a\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u7684\u53c2\u6570\u6709\u6548\u5fae\u8c03"},"content":{"rendered":"
\n
\n
\n
\n

<\/p>\n


<\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/span><\/p>\n

\u5173\u952e\u8bcd\uff1a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff0c\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff0c\u53c2\u6570\u5fae\u8c03<\/span><\/strong><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/span><\/strong><\/strong><\/strong><\/p>\n<\/section>\n

\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n


<\/p>\n


<\/p>\n<\/section>\n

\n
<\/section>\n<\/section>\n
\n
\n
\n
\n
\u8bba\u6587\u9898\u76ee\uff1a<\/span>Parameter-efficient fine-tuning of large-scale pre-trained language models<\/span><\/section>\n
\u8bba\u6587\u6765\u6e90\uff1aNature Machine Intelligence<\/em><\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/www.nature.com\/articles\/s42256-023-00626-4<\/span><\/span><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/blockquote>\n

<\/section>\n
\u968f\u7740\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b<\/span>\uff08pre-trained language model, PLM\uff09<\/span>\u548c\u9884\u8bad\u7ec3-\u5fae\u8c03\u8303\u5f0f\u7684\u76db\u884c\uff0c\u4eba\u4eec\u4e0d\u65ad\u53d1\u73b0\u66f4\u5927\u7684\u6a21\u578b\u5f80\u5f80\u80fd\u4ea7\u751f\u66f4\u597d\u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u968f\u7740\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u89c4\u6a21\u7684\u6269\u5927\uff0c\u5fae\u8c03\u548c\u5b58\u50a8\u6240\u6709\u53c2\u6570\u7684\u6210\u672c\u8fc7\u9ad8\uff0c\u4f7f\u5f97\u6a21\u578b\u7684\u90e8\u7f72\u53d8\u5f97\u4e0d\u53ef\u884c\u3002\u8fd9\u5c31\u9700\u8981\u4e00\u4e2a\u65b0\u7684\u7814\u7a76\u5206\u652f\uff0c\u4e13\u6ce8\u4e8e\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u7684\u53c2\u6570\u9ad8\u6548\u9002\u5e94\u6027\uff0c\u53ea\u4f18\u5316\u4e00\u5c0f\u90e8\u5206\u6a21\u578b\u53c2\u6570\uff0c\u540c\u65f6\u4fdd\u6301\u5176\u4f59\u53c2\u6570\u7684\u56fa\u5b9a\uff0c\u8fd9\u80fd\u5927\u5927\u964d\u4f4e\u8ba1\u7b97\u548c\u5b58\u50a8\u6210\u672c\u3002\u603b\u7684\u6765\u8bf4\uff0c\u8fd9\u4e00\u7814\u7a76\u9886\u57df\u8868\u660e\u5927\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u53ef\u4ee5\u901a\u8fc7\u4f18\u5316\u5c11\u91cf\u53c2\u6570\u6765\u6709\u6548\u6fc0\u6d3b\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u5c3d\u7ba1\u5df2\u7ecf\u5b58\u5728\u5404\u79cd\u5404\u6837\u7684\u8bbe\u8ba1\uff0c\u8fd9\u7bc7\u6587\u7ae0\u7528\u4e00\u4e2a\u66f4\u4e00\u81f4\u548c\u66f4\u5bb9\u6613\u7406\u89e3\u7684\u672f\u8bed\u201cdelta-tuning\u201d\u6765\u8ba8\u8bba\u548c\u5206\u6790\u8fd9\u4e9b\u65b9\u6cd5\uff0c\u5176\u4e2d\u201cdelta\u201d \u662f\u4e00\u4e2a\u7ecf\u5e38\u7528\u6765\u8868\u793a\u53d8\u5316\u7684\u6570\u5b66\u7b26\u53f7\uff0c\u88ab\u501f\u7528\u6765\u6307\u8bad\u7ec3\u4e2d\u201c\u6539\u53d8\u201d\u7684\u90a3\u90e8\u5206\u53c2\u6570\u3002\u8fd9\u9879\u7814\u7a76\u6b63\u5f0f\u63cf\u8ff0\u4e86\u8fd9\u4e2a\u95ee\u9898\uff0c\u5e76\u4e3a\u73b0\u6709\u7684 delta-tuning \u65b9\u6cd5\u63d0\u51fa\u4e86\u4e00\u4e2a\u7edf\u4e00\u7684\u5206\u7c7b\u6807\u51c6\uff0c\u4ee5\u63a2\u7d22\u5b83\u4eec\u4e4b\u95f4\u7684\u5173\u8054\u548c\u5dee\u5f02\u3002delta-tuning \u5fae\u8c03\u65b9\u6cd5\u53ea\u66f4\u65b0\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b<\/span>\uff08PLM\uff09<\/span>\u7684\u6700\u540e\u4e00\u5c42\uff0c\u800c\u4fdd\u6301\u5176\u4ed6\u5c42\u4e0d\u53d8\u3002\u7814\u7a76\u8bc1\u660e delta-tuning \u53ef\u4ee5\u5728\u4e0d\u635f\u5931\u6027\u80fd\u7684\u60c5\u51b5\u4e0b\uff0c\u663e\u8457\u51cf\u5c11\u5fae\u8c03\u7684\u53c2\u6570\u6570\u91cf\u548c\u8ba1\u7b97\u6210\u672c\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u4f5c\u8005\u8fd8\u8ba8\u8bba\u4e86 delta-tuning \u6709\u6548\u6027\u7684\u7406\u8bba\u539f\u5219\uff0c\u5e76\u4ece\u4f18\u5316\u548c\u6700\u4f18\u63a7\u5236\u7684\u89d2\u5ea6\u5bf9\u5176\u8fdb\u884c\u89e3\u91ca\u3002\u6b64\u5916\uff0c\u4f5c\u8005\u8fd8\u5bf9 100 \u591a\u4e2a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u8fdb\u884c\u4e86\u5168\u9762\u7684\u5b9e\u8bc1\u7814\u7a76\uff0c\u5e76\u7814\u7a76\u4e86 delta-tuning \u7684\u5404\u4e2a\u65b9\u9762\u3002\u5728\u591a\u4e2a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e0a\uff0c\u7528 delta-tuning \u5fae\u8c03\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u53ef\u4ee5\u8fbe\u5230\u548c\u5168\u6a21\u578b\u5fae\u8c03\u76f8\u5f53\u6216\u66f4\u597d\u7684\u6548\u679c\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/section>\n

\u56fe1. T5(BASE) \u5728\u4e0d\u540c\u8bad\u7ec3\u6b65\u9aa4\u4e2d\u91c7\u7528\u4e0d\u540c\u589e\u91cf\u8c03\u6574\u65b9\u6cd5\uff08LR\u3001AP \u548c PF\uff09\u548c\u5fae\u8c03 (FT) \u7684\u6027\u80fd\u3002<\/span><\/p>\n


<\/span><\/section>\n
<\/section>\n
\u56fe2. \u6bcf\u4e2adelta-tuning\u65b9\u6cd5\u548c\u5fae\u8c03\u6240\u6d88\u8017\u7684GPU\u5185\u5b58\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/section>\n
\u56fe3. delta-tuning \u7684\u5206\u7c7b\u6807\u51c6\u3002\u8fd9\u91cc \u0398 \u8868\u793a\u9884\u8bad\u7ec3\u53c2\u6570\uff0c\u0398’ \u8868\u793a\u8c03\u6574\u597d\u7684\u53c2\u6570\u3002<\/span><\/section>\n


<\/p>\n


<\/p>\n

\n
\n
\n
\n

\u590d\u6742\u79d1\u5b66\u6700\u65b0\u8bba\u6587<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n


<\/p>\n<\/section>\n

\u96c6\u667a\u6591\u56fe\u9876\u520a\u8bba\u6587\u901f\u9012\u680f\u76ee\u4e0a\u7ebf\u4ee5\u6765\uff0c\u6301\u7eed\u6536\u5f55\u6765\u81eaNature\u3001Science<\/span>\u7b49\u9876\u520a\u7684\u6700\u65b0\u8bba\u6587\uff0c\u8ffd\u8e2a\u590d\u6742\u7cfb\u7edf\u3001\u7f51\u7edc\u79d1\u5b66\u3001\u8ba1\u7b97\u793e\u4f1a\u79d1\u5b66\u7b49\u9886\u57df\u7684\u524d\u6cbf\u8fdb\u5c55\u3002\u73b0\u5728\u6b63\u5f0f\u63a8\u51fa\u8ba2\u9605\u529f\u80fd\uff0c\u6bcf\u5468\u901a\u8fc7\u5fae\u4fe1\u670d\u52a1\u53f7\u300c\u6211\u7684\u96c6\u667a\u300d\u63a8\u9001\u8bba\u6587\u4fe1\u606f\u3002\u626b\u63cf\u4e0b\u65b9\u4e8c\u7ef4\u7801\u5373\u53ef\u4e00\u952e\u8ba2\u9605\uff1a<\/span>
<\/section>\n


<\/p>\n

<\/p>\n


<\/section>\n

<\/section>\n

\u63a8\u8350\u9605\u8bfb<\/span><\/strong><\/strong><\/p>\n

1. <\/strong>\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5982\u4f55\u5b66\u4e60\u4e8b\u5b9e\u6027\u77e5\u8bc6\uff1f\u517c\u8bba\u4e00\u79cd\u56e0\u679c\u5e72\u9884\u7684\u8bc4\u4f30\u65b9\u6cd5<\/a><\/strong><\/span><\/section>\n
2. <\/strong>ChatGPT \u4e3a\u5565\u90a3\u4e48\u725b\uff1f\u8bed\u8a00\u6a21\u578b\u8db3\u591f\u5927\u5c31\u4f1a\u6d8c\u73b0\u51fa\u65b0\u80fd\u529b<\/a><\/strong><\/span><\/section>\n
3. <\/strong><\/strong>\u7406\u89e3\u5927\u8bed\u8a00\u6a21\u578b\u2014\u201410\u7bc7\u8bba\u6587\u7684\u7b80\u660e\u6e05\u5355<\/a><\/strong><\/strong><\/span><\/section>\n
4. \u300a\u5f20\u6c5f\u00b7\u590d\u6742\u79d1\u5b66\u524d\u6cbf27\u8bb2\u300b\u5b8c\u6574\u4e0a\u7ebf\uff01<\/strong><\/a><\/strong><\/strong><\/strong><\/section>\n
5. \u6210\u4e3a\u96c6\u667aVIP\uff0c\u89e3\u9501\u5168\u7ad9\u8bfe\u7a0b\uff0f\u8bfb\u4e66\u4f1a<\/span><\/strong><\/a><\/strong><\/strong><\/strong><\/strong><\/section>\n
6. \u52a0\u5165\u96c6\u667a\uff0c\u4e00\u8d77\u590d\u6742\uff01<\/strong><\/a><\/strong><\/strong><\/strong><\/section>\n


<\/p>\n


<\/p>\n

\u70b9\u51fb\u201c\u9605\u8bfb\u539f\u6587\u201d\uff0c\u8ffd\u8e2a\u590d\u6742\u79d1\u5b66\u9876\u520a\u8bba\u6587<\/strong><\/strong><\/span><\/strong><\/section>\n

<\/mp-style-type><\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

\u5173\u952e\u8bcd\uff1a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff0c\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff0c\u53c2\u6570\u5fae\u8c03 \u8bba\u6587\u9898\u76ee\uff1aParameter-efficient fine-tuning of large-scale pre-trained language models \u8bba\u6587\u6765\u6e90\uff1aNature Machine Intelligence \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/…<\/p>\n","protected":false},"author":0,"featured_media":41187,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[1],"tags":[],"special":[],"_links":{"self":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/41193"}],"collection":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=41193"}],"version-history":[{"count":0,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/41193\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/media\/41187"}],"wp:attachment":[{"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=41193"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=41193"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=41193"},{"taxonomy":"special","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fspecial&post=41193"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}