{"id":3842,"date":"2021-06-09T17:35:47","date_gmt":"2021-06-09T09:35:47","guid":{"rendered":"http:\/\/www.sniper97.cn\/?p=3842"},"modified":"2021-06-09T17:35:47","modified_gmt":"2021-06-09T09:35:47","slug":"%e3%80%90%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%ac%94%e8%ae%b0%e3%80%91electra","status":"publish","type":"post","link":"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/note-deep-learning\/3842\/","title":{"rendered":"\u3010\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\u3011ELECTRA"},"content":{"rendered":"\n<p>\u672c\u6587\u6765\u81eaGoggle\u7684\u8bba\u6587 <a rel=\"noreferrer noopener\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/openreview.net\/pdf%3Fid%3Dr1xMH1BtvB\" target=\"_blank\">ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generator<\/a> <\/p>\n\n\n<p>\u76f8\u6bd4Bert\uff0cELECTRA\u91c7\u7528\u4e86\u66f4\u4f18\u79c0\u7684mask\u65b9\u6cd5\u4ee5\u53ca\u8bad\u7ec3\u76ee\u6807\u3002<\/p>\n\n\n<h3 class=\"wp-block-heading\">1.\u66f4\u4f18\u79c0\u7684MASK<\/h3>\n\n\n<p>bert\u7684mask\u662f\u968f\u673a\u7684\uff0c\u6bd4\u5982\u201c\u6211\u60f3\u5403\u82f9\u679c\u201d\uff0c\u5f88\u53ef\u80fdmask\u7684\u7ed3\u679c\u662f\u201c\u6211\u60f3\u5403\u82f9MASK\u201d\uff0c\u5bf9\u4e8e\u8bed\u8a00\u6a21\u578b\u6765\u8bf4\uff0c\u9884\u6d4b\u201c\u679c\u201d\u662f\u5f88\u5bb9\u6613\u7684\u3002<\/p>\n\n\n<p>\u4f46\u662f\u5bf9\u4e8e\u201c\u6211MASK\u5403\u82f9\u679c\u201d\u53ef\u80fd\u9884\u6d4b\u201c\u60f3\u201d\u7684\u96be\u5ea6\u5c31\u4f1a\u5f88\u5927\u4e86\u3002<\/p>\n\n\n<p>\u56e0\u6b64ELECTRA\u5f15\u5165\u4e86\u4e00\u4e2a\u5c0f\u578bMLM\u8bed\u8a00\u6a21\u578b\u4f5c\u4e3a\u751f\u6210\u5668\uff08Generator\uff09\uff0c\u7528\u6765\u5b8c\u6210\u90a3\u4e9b\u201c\u7b80\u5355\u201d\u7684MLM\u4efb\u52a1\u3002\u540c\u65f6\u5bf9\u4e8e\u201c\u8f83\u96be\u201d\u7684MLM\u4efb\u52a1\uff0c\u7531\u4e0b\u6e38\u7684\u5224\u522b\u5668\uff08Discriminator\uff09\u6765\u8fdb\u884c\u5224\u65ad\uff0c\u54ea\u4e9b\u662f\u66ff\u6362\u6216\u7684token\u3002<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"653\" height=\"194\" src=\"\/wp-content\/uploads\/2021\/06\/\u56fe\u7247-1.png\" alt=\"\" class=\"wp-image-3843\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2021\/06\/\u56fe\u7247-1.png 653w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2021\/06\/\u56fe\u7247-1-300x89.png 300w\" sizes=\"(max-width: 653px) 100vw, 653px\" \/><\/figure><\/div>\n\n\n<h3 class=\"wp-block-heading\">2.\u8bad\u7ec3\u76ee\u6807<\/h3>\n\n\n<p>\u8bad\u7ec3\u76ee\u6807\u5f53\u7136\u53ef\u4ee5\u548cBert\u76f8\u540c\uff0c\u6839\u636e\u662f\u5426\u53d8\u5316\u5bf9\u5e94Bert\u4e2dMASK\u7684\u66ff\u6362\u548c\u4e0d\u53d8\u3002<\/p>\n\n\n<p>\u4f46\u662fELECTRA\u4e0d\u4e00\u6837\u5c31\u4e0d\u4e00\u6837\u5728\u4ed6\u66f4\u6362\u4e86\u8bad\u7ec3\u76ee\u6807\u3002<\/p>\n\n\n<p>\u5982\u4e0a\u56fe\uff0c\u4f7f\u7528\u4e00\u4e2a\u4e8c\u5206\u7c7b\u53bb\u5224\u65ad\u6bcf\u4e2atoken\u662f\u5426\u88ab\u6362\u8fc7\u3002\u8fd9\u5c31\u628aLM\u4efb\u52a1\u8f6c\u6362\u4e3a\u4e86\u5e8f\u5217\u6807\u6ce8\uff08\u5206\u7c7b\uff09\u4efb\u52a1\u3002\u4e5f\u5c31\u5e26\u6765\u4e86\u4e24\u4e2a\u597d\u5904\uff1a<\/p>\n\n\n<ul><li>\u6bcf\u4e2atoken\u90fd\u80fd\u6709\u4e00\u5b9a\u7684\u8d21\u732e\uff08 contribute to some extent \uff09<\/li><li>\u7f29\u5c0f\u4e86\u8ba1\u7b97\u91cf\uff08\u76f8\u6bd4\u9884\u6d4btoken\uff0c\u4e8c\u5206\u7c7b\u8ba1\u7b97\u91cf\u5c0f\u591a\u4e86\uff09<\/li><\/ul>\n\n\n<p> \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u5224\u522b\u5668\u7684loss\u4e0d\u4f1a\u53cd\u5411\u4f20\u64ad\u5230\u751f\u6210\u5668\uff0c\u5728pre-training\u4e4b\u540e\uff0c\u6211\u4eec\u53ea\u4f7f\u7528 \u5224\u522b\u5668 \u8fdb\u884cfine-tuning. \uff08 \u4e24\u8005\u540c\u65f6\u8bad\u7ec3\uff0c\u4f46\u5224\u522b\u5668\u7684\u68af\u5ea6\u4e0d\u4f1a\u4f20\u7ed9\u751f\u6210\u5668 \uff09<\/p>\n\n\n<p>\u7136\u540e\u5728\u5fae\u8c03\u7684\u8fc7\u7a0b\u4e2d\u820d\u5f03\u751f\u6210\u5668\u5373\u53ef\u3002<\/p>\n\n\n<p>\u4ece\u8bba\u6587\u4ee3\u7801\u4e0a\u6765\u8bf4\uff0c<a rel=\"noreferrer noopener\" aria-label=\"\u5224\u522b\u5668\uff08\u5728\u65b0\u7a97\u53e3\u6253\u5f00\uff09\" href=\"https:\/\/github.com\/google-research\/electra\/blob\/master\/run_pretraining.py#L102\" target=\"_blank\">\u5224\u522b\u5668<\/a>\u548c<a rel=\"noreferrer noopener\" aria-label=\"\u751f\u6210\u5668\uff08\u5728\u65b0\u7a97\u53e3\u6253\u5f00\uff09\" href=\"https:\/\/github.com\/google-research\/electra\/blob\/master\/run_pretraining.py#L80\" target=\"_blank\">\u751f\u6210\u5668<\/a>\u90fd\u662fTransformer based\uff0c\u4e5f\u5c31\u662f\u90fd\u662f\u7c7bbert\u7ed3\u6784\u7684\uff08\u751a\u81f3\u4ee3\u7801\u91cc\u76f4\u63a5\u4f7f\u7528\u4e86Bert\u76f8\u5173\u53d8\u91cf\uff09\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u6765\u81eaGoggle\u7684\u8bba\u6587 ELECTRA: Pre-training Text Encoders  [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_mi_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[18,9],"tags":[31,37,39],"views":4035,"_links":{"self":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3842"}],"collection":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/comments?post=3842"}],"version-history":[{"count":0,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3842\/revisions"}],"wp:attachment":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/media?parent=3842"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/categories?post=3842"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/tags?post=3842"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}