{"id":5725,"date":"2024-11-08T05:59:00","date_gmt":"2024-11-07T20:59:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=5725"},"modified":"2024-11-08T05:59:00","modified_gmt":"2024-11-07T20:59:00","slug":"lora-vs-full-fine-tuning-an-illusion-of-equivalence","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=5725","title":{"rendered":"LoRA vs Full Fine-tuning: An Illusion of Equivalence\u00a0"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>LoRA vs Full Fine-tuning: An Illusion of Equivalence\u00a0<\/strong>[76.1]<br>\u672c\u7814\u7a76\u3067\u306f, \u7570\u306a\u308b\u5fae\u8abf\u6574\u6cd5\u304c, \u30b9\u30da\u30af\u30c8\u30eb\u7279\u6027\u306e\u30ec\u30f3\u30ba\u3092\u7528\u3044\u3066\u30e2\u30c7\u30eb\u306e\u91cd\u307f\u884c\u5217\u3092\u89e3\u6790\u3059\u308b\u3053\u3068\u306b\u3088\u308a, \u4e8b\u524d\u5b66\u7fd2\u30e2\u30c7\u30eb\u3092\u5909\u5316\u3055\u305b\u308b\u65b9\u6cd5\u306b\u3064\u3044\u3066\u691c\u8a0e\u3057\u305f\u3002 \u5358\u4e00\u5024\u5206\u89e3\u304c\u5168\u304f\u7570\u306a\u308b\u69cb\u9020\u3092\u6301\u3064\u5168\u5fae\u8abf\u6574\u304a\u3088\u3073LoRA\u53ce\u91cf\u884c\u5217\u304c\u5f97\u3089\u308c\u305f\u3002 \u30a4\u30f3\u30c8\u30eb\u30fc\u30c0\u6b21\u5143\u304cLoRA\u306e\u5fae\u8abf\u6574\u30e2\u30c7\u30eb\u306b\u306a\u305c\u73fe\u308c\u308b\u306e\u304b\u3001\u306a\u305c\u305d\u308c\u3089\u304c\u671b\u307e\u3057\u304f\u306a\u3044\u306e\u304b\u3001\u305d\u3057\u3066\u3069\u306e\u3088\u3046\u306b\u3057\u3066\u305d\u306e\u52b9\u679c\u3092\u6700\u5c0f\u5316\u3067\u304d\u308b\u304b\u3092\u691c\u8a0e\u3059\u308b\u3053\u3068\u3067\u7d50\u8ad6\u3092\u4e0b\u3059\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2410.21228v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2410.21228v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Mon, 28 Oct 2024 17:14:01 GMT)<\/li>\n\n\n\n<li>LoRA\u3067\u5f97\u3089\u308c\u305fWeight\u3068\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3067\u5f97\u3089\u308c\u305fWeight\u306e\u5dee\u7570\u3092\u5206\u6790\u3001\u300cMore specifically, we first show that the weight matrices trained with LoRA have new, high-ranking singular vectors, which we call intruder dimensions. Intruder dimensions do not appear during full fine-tuning. Second, we show that LoRA models with intruder dimensions, despite achieving similar performance to full fine-tuning on the target task, become worse models of the pre-training distribution and adapt less robustly to multiple tasks sequentially. \uff1a\u3068\u306e\u3053\u3068\u3002<\/li>\n\n\n\n<li>\u8208\u5473\u6df1\u3044\u6027\u8cea\u3067\u3042\u308b\u3068\u601d\u3046\u306e\u3068\u3001\u9811\u5065\u6027\u3092\u8a55\u4fa1\u3059\u308b\u306e\u306f\u5927\u5909\u306a\u306e\u3067\u554f\u984c\u304c\u898b\u904e\u3054\u3055\u308c\u3084\u3059\u305d\u3046\u306a\u306e\u304c\u82e5\u5e72\u6016\u3044\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[150,227],"class_list":["post-5725","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-fine-tuning","tag-lora"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/5725","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5725"}],"version-history":[{"count":0,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/5725\/revisions"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5725"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5725"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5725"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}