{"id":3929,"date":"2024-06-20T12:31:26","date_gmt":"2024-06-20T04:31:26","guid":{"rendered":"https:\/\/nullthought.net\/?p=3929"},"modified":"2025-02-23T11:08:03","modified_gmt":"2025-02-23T03:08:03","slug":"transformer%e9%9c%80%e8%a6%81%e6%88%b4%e4%b8%8a%e7%9c%bc%e9%95%9c%f0%9f%98%8e","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=3929","title":{"rendered":"Transformer\u9700\u8981\u6234\u4e0a\u773c\u955c\ud83d\ude0e"},"content":{"rendered":"\n<p>Federico Barbero\u7b49\u8fd1\u671f\u53d1\u8868\u8bba\u6587\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2406.04267\" target=\"_blank\" rel=\"noreferrer noopener\">Transformers need glasses\ud83d\udc53: Information over-squashing in language tasks<\/a>\uff0c\u7814\u7a76\u4e86\u4ec5\u89e3\u7801\u5668\u7ed3\u6784\u7684Transformer(decoder-only Transformers)\u5728\u8bed\u8a00\u4efb\u52a1\u4e2d\u7684\u4fe1\u606f\u4f20\u64ad\u65b9\u5f0f\u3002\u901a\u8fc7\u7406\u8bba\u4fe1\u53f7\u4f20\u64ad\u5206\u6790\uff0c\u53d1\u73b0\u4e86\u4e00\u79cd\u8868\u793a\u5d29\u6e83(representational collapse)\u73b0\u8c61\uff1a\u4e0d\u540c\u7684\u8f93\u5165\u5e8f\u5217\u5728Transformer\u6700\u7ec8\u5c42\u7684\u6700\u540e\u4e00\u4e2atoken\u8868\u793a\u4e2d\u53ef\u80fd\u4f1a\u53d8\u5f97\u975e\u5e38\u63a5\u8fd1\u3002\u8fd9\u79cd\u73b0\u8c61\u5728\u73b0\u4ee3LLMs\u4e2d\u4f7f\u7528\u7684\u4f4e\u7cbe\u5ea6\u6d6e\u70b9\u683c\u5f0f\u4e0b\u5c24\u4e3a\u660e\u663e\uff0c\u5bfc\u81f4\u6a21\u578b\u5728\u5904\u7406\u8fd9\u4e9b\u5e8f\u5217\u65f6\u51fa\u73b0\u9519\u8bef\u3002\u6b64\u5916\uff0c\u8bba\u6587\u8fd8\u6307\u51fa\u4ec5\u89e3\u7801\u5668Transformer\u8bed\u8a00\u6a21\u578b\u53ef\u80fd\u4f1a\u5bf9\u8f93\u5165\u4e2d\u7684\u7279\u5b9atoken\u5931\u53bb\u654f\u611f\u6027\uff0c\u8fd9\u4e0e\u56fe\u795e\u7ecf\u7f51\u7edc\u4e2d\u7684\u8fc7\u5ea6\u538b\u7f29\u73b0\u8c61\u7c7b\u4f3c\u3002\u8bba\u6587\u901a\u8fc7\u5b9e\u9a8c\u8bc1\u660e\u4e86\u8fd9\u4e9b\u95ee\u9898\u7684\u5b58\u5728\uff0c\u5e76\u63d0\u51fa\u4e86\u7b80\u5355\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u8bba\u6587\u4f5c\u8005:<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Federico Barbero, University of Oxford<\/li>\n\n\n\n<li>Andrea Banino, Steven Kapturowski, Dharshan Kumaran, Jo\u00e3o G.M. Ara\u00fajo, Alex Vitvitskyi, Razvan Pascanu, Petar Veli\u010dkovi\u0107, Google DeepMind<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"820\" height=\"425\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/06\/image-8.png\" alt=\"\" class=\"wp-image-3931\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/06\/image-8.png 820w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/06\/image-8-300x155.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/06\/image-8-768x398.png 768w\" sizes=\"auto, (max-width: 820px) 100vw, 820px\" \/><figcaption class=\"wp-element-caption\">\u8868\u793a\u5d29\u6e83(Representational Collapse)\u3001\u8fc7\u5ea6\u538b\u7f29(Over-squashing)<br><a href=\"https:\/\/arxiv.org\/abs\/2406.04267\" target=\"_blank\" rel=\"noreferrer noopener\">Transformers need glasses\ud83d\udc53: Information over-squashing in language tasks<\/a><\/figcaption><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u5f15\u8a00<\/h3>\n\n\n\n<p>\u8fd1\u5e74\u6765\uff0cNLP\u9886\u57df\u901a\u8fc7\u5f15\u5165\u57fa\u4e8eTransformer\u7684\u67b6\u6784\u53d6\u5f97\u4e86\u9769\u547d\u6027\u8fdb\u5c55\u3002\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728\u591a\u79cd\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u5305\u62ec\u5bf9\u8bdd\u4ee3\u7406\u3001\u591a\u6a21\u6001\u8f93\u5165\u7406\u89e3\u548c\u4ee3\u7801\u8865\u5168\u7b49\u3002\u7136\u800c\uff0cTransformer\u6a21\u578b\u5728\u67d0\u4e9b\u7b80\u5355\u4efb\u52a1\u4e2d\u4ecd\u7136\u8868\u73b0\u51fa\u660e\u663e\u7684\u4e0d\u8db3\uff0c\u4f8b\u5982\u8ba1\u6570\u548c\u590d\u5236\u64cd\u4f5c\u3002\u672c\u6587\u65e8\u5728\u901a\u8fc7\u7406\u8bba\u5206\u6790\u548c\u5b9e\u9a8c\u8bc1\u5b9e\u8fd9\u4e9b\u4e0d\u8db3\uff0c\u5e76\u63d0\u4f9b\u6539\u5584\u8fd9\u4e9b\u95ee\u9898\u7684\u65b9\u6848\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u80cc\u666f<\/h3>\n\n\n\n<p>\u672c\u6587\u4e3b\u8981\u7814\u7a76\u4ec5\u89e3\u7801\u5668\u7ed3\u6784\u7684Transformer\uff0c\u8fd9\u7c7b\u7ed3\u6784\u662f\u5927\u591a\u6570\u5f53\u524dLLMs\u7684\u57fa\u7840\u67b6\u6784\u3002\u4f7f\u7528\u8fd9\u4e9b\u6a21\u578b\u8fdb\u884c\u4fe1\u53f7\u4f20\u64ad\u5206\u6790\uff0c\u7279\u522b\u662f\u5206\u6790\u6700\u7ec8\u5c42\u6700\u540e\u4e00\u4e2atoken\u7684\u8868\u793a\uff0c\u8fd9\u662f\u7528\u4e8e\u4e0b\u4e00\u4e2atoken\u9884\u6d4b\u7684\u8868\u793a\u3002\u901a\u8fc7\u5206\u6790\uff0c\u672c\u6587\u53d1\u73b0\u4e86\u4e00\u79cd\u8868\u793a\u5d29\u6e83\u73b0\u8c61\uff1a\u67d0\u4e9b\u4e0d\u540c\u7684\u8f93\u5165\u5e8f\u5217\u5728Transformer\u6700\u7ec8\u5c42\u7684\u6700\u540e\u4e00\u4e2atoken\u8868\u793a\u4e2d\u53ef\u80fd\u4f1a\u53d8\u5f97\u975e\u5e38\u63a5\u8fd1\uff0c\u5bfc\u81f4\u6a21\u578b\u65e0\u6cd5\u533a\u5206\u8fd9\u4e9b\u5e8f\u5217\u3002\u6b64\u5916\uff0c\u4f4e\u7cbe\u5ea6\u6d6e\u70b9\u683c\u5f0f\u52a0\u5267\u4e86\u8fd9\u4e00\u73b0\u8c61\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u52a8\u673a\u793a\u4f8b<\/h3>\n\n\n\n<p>\u672c\u6587\u901a\u8fc7\u4e00\u7cfb\u5217\u5b9e\u9a8c\u5c55\u793a\u4e86\u73b0\u4ee3\u4ec5\u89e3\u7801\u5668Transformer\u67b6\u6784\u5728\u7b80\u5355\u7684\u590d\u5236\u548c\u8ba1\u6570\u4efb\u52a1\u4e2d\u7684\u60ca\u4eba\u5931\u8d25\u6848\u4f8b\u3002\u8fd9\u4e9b\u5b9e\u9a8c\u63ed\u793a\u4e86\u6a21\u578b\u5728\u5904\u7406\u8fd9\u4e9b\u57fa\u672c\u4efb\u52a1\u65f6\u7684\u4e0d\u8db3\uff0c\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u672c\u6587\u7406\u8bba\u5206\u6790\u7684\u5fc5\u8981\u6027\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3.1 \u590d\u5236\u4efb\u52a1<\/h4>\n\n\n\n<p>\u5728\u590d\u5236\u4efb\u52a1\u4e2d\uff0c\u6a21\u578b\u9700\u8981\u6839\u636e\u63d0\u793a\u590d\u5236\u7279\u5b9a\u4f4d\u7f6e\u7684token\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5f53\u63d0\u793a\u6a21\u578b\u590d\u5236\u5e8f\u5217\u672b\u5c3e\u7684\u5143\u7d20\u65f6\uff0c\u6a21\u578b\u8868\u73b0\u4e0d\u4f73\uff0c\u800c\u590d\u5236\u5e8f\u5217\u5f00\u5934\u7684\u5143\u7d20\u65f6\u8868\u73b0\u8f83\u597d\u3002\u8fd9\u8868\u660e\u4fe1\u606f\u5728\u5e8f\u5217\u4e2d\u7684\u4f20\u64ad\u5b58\u5728\u95ee\u9898\uff0c\u7279\u522b\u662f\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3.2 \u8ba1\u6570\u4efb\u52a1<\/h4>\n\n\n\n<p>\u5728\u8ba1\u6570\u4efb\u52a1\u4e2d\uff0c\u6a21\u578b\u9700\u8981\u8ba1\u7b97\u7279\u5b9atoken\u5728\u5e8f\u5217\u4e2d\u51fa\u73b0\u7684\u6b21\u6570\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u968f\u7740\u5e8f\u5217\u957f\u5ea6\u7684\u589e\u52a0\uff0c\u6a21\u578b\u7684\u8868\u73b0\u6025\u5267\u4e0b\u964d\uff0c\u9519\u8bef\u7387\u663e\u8457\u4e0a\u5347\u3002\u8fd9\u8868\u660e\u6a21\u578b\u5728\u5904\u7406\u9700\u8981\u7cbe\u786e\u8ba1\u6570\u7684\u4efb\u52a1\u65f6\u5b58\u5728\u4e25\u91cd\u95ee\u9898\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4. \u8868\u793a\u5d29\u6e83<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">4.1 \u7406\u8bba\u5206\u6790<\/h4>\n\n\n\n<p>\u672c\u6587\u901a\u8fc7\u7406\u8bba\u5206\u6790\u63ed\u793a\u4e86\u8868\u793a\u5d29\u6e83\u73b0\u8c61\uff1a\u5728\u67d0\u4e9b\u6761\u4ef6\u4e0b\uff0c\u4e24\u4e2a\u4e0d\u540c\u7684\u8f93\u5165\u5e8f\u5217\u5728Transformer\u6700\u7ec8\u5c42\u7684\u8868\u793a\u4f1a\u53d8\u5f97\u975e\u5e38\u63a5\u8fd1\uff0c\u7279\u522b\u662f\u5f53\u5e8f\u5217\u957f\u5ea6\u589e\u52a0\u65f6\u3002\u8fd9\u79cd\u73b0\u8c61\u5728\u4f4e\u6d6e\u70b9\u7cbe\u5ea6\u4e0b\u66f4\u4e3a\u660e\u663e\uff0c\u5bfc\u81f4\u6a21\u578b\u65e0\u6cd5\u533a\u5206\u8fd9\u4e9b\u5e8f\u5217\u3002\u7406\u8bba\u5206\u6790\u8fd8\u6307\u51fa\uff0c\u8fd9\u79cd\u73b0\u8c61\u4e0e\u56fe\u795e\u7ecf\u7f51\u7edc\u4e2d\u7684\u8fc7\u5ea6\u538b\u7f29\u95ee\u9898\u6709\u5173\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">4.2 \u5b9e\u9a8c\u8bc1\u660e<\/h4>\n\n\n\n<p>\u901a\u8fc7\u5bf9\u73b0\u6709LLMs\u7684\u5b9e\u9a8c\u8bc1\u5b9e\u4e86\u7406\u8bba\u5206\u6790\u7684\u5b9e\u9645\u5f71\u54cd\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u8868\u793a\u5d29\u6e83\u73b0\u8c61\u5728\u957f\u5e8f\u5217\u4e2d\u786e\u5b9e\u5b58\u5728\uff0c\u5e76\u4e14\u4f1a\u5bfc\u81f4\u6a21\u578b\u5728\u590d\u5236\u548c\u8ba1\u6570\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u4e0d\u4f73\u3002\u7279\u522b\u662f\u5f53\u5e8f\u5217\u4e2d\u6709\u91cd\u590dtoken\u65f6\uff0c\u8868\u793a\u5d29\u6e83\u73b0\u8c61\u66f4\u52a0\u660e\u663e\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5. \u4fe1\u606f\u8fc7\u5ea6\u538b\u7f29<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">5.1 \u7406\u8bba\u5206\u6790<\/h4>\n\n\n\n<p>\u4fe1\u606f\u8fc7\u5ea6\u538b\u7f29\u73b0\u8c61\u63cf\u8ff0\u4e86\u8f93\u5165\u5e8f\u5217\u4e2d\u8f83\u65e9\u7684token\u4fe1\u606f\u66f4\u5bb9\u6613\u4f20\u64ad\u5230\u6700\u7ec8\u8868\u793a\u4e2d\uff0c\u800c\u8f83\u665a\u7684token\u4fe1\u606f\u5219\u53ef\u80fd\u4e27\u5931\u3002\u672c\u6587\u901a\u8fc7\u5206\u6790\u63ed\u793a\u4e86\u8fd9\u79cd\u73b0\u8c61\u5728\u4ec5\u89e3\u7801\u5668Transformer\u4e2d\u7684\u8868\u73b0\uff0c\u5e76\u6307\u51fa\u8fd9\u79cd\u73b0\u8c61\u4e0e\u56fe\u795e\u7ecf\u7f51\u7edc\u4e2d\u7684\u8fc7\u5ea6\u538b\u7f29\u95ee\u9898\u7c7b\u4f3c\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">5.2 \u5b9e\u9a8c\u8bc1\u660e<\/h4>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u8f93\u5165\u5e8f\u5217\u4e2d\u8f83\u65e9\u7684token\u4fe1\u606f\u66f4\u5bb9\u6613\u4fdd\u7559\u5230\u6700\u7ec8\u8868\u793a\u4e2d\uff0c\u800c\u8f83\u665a\u7684token\u4fe1\u606f\u5219\u66f4\u5bb9\u6613\u4e27\u5931\u3002\u8fd9\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u6a21\u578b\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\u8868\u73b0\u4e0d\u4f73\uff0c\u7279\u522b\u662f\u5728\u8ba1\u6570\u548c\u590d\u5236\u4efb\u52a1\u4e2d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6. \u8ba1\u6570\u95ee\u9898<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">6.1 \u7406\u8bba\u5206\u6790<\/h4>\n\n\n\n<p>\u8ba1\u6570\u4efb\u52a1\u9700\u8981\u6a21\u578b\u80fd\u591f\u7cbe\u786e\u5730\u5904\u7406\u548c\u4f20\u64ad\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u4fe1\u606f\u3002\u7136\u800c\uff0cTransformer\u6a21\u578b\u4e2d\u7684softmax\u5f52\u4e00\u5316\u673a\u5236\u548c\u4f4d\u7f6e\u7f16\u7801\u4f7f\u5f97\u8ba1\u6570\u4efb\u52a1\u53d8\u5f97\u975e\u5e38\u56f0\u96be\u3002\u672c\u6587\u901a\u8fc7\u7406\u8bba\u5206\u6790\u63ed\u793a\u4e86\u8fd9\u79cd\u673a\u5236\u5bfc\u81f4\u7684\u8868\u793a\u95ee\u9898\uff0c\u5e76\u6307\u51fa\u8fd9\u79cd\u95ee\u9898\u5728\u4f4e\u7cbe\u5ea6\u6d6e\u70b9\u683c\u5f0f\u4e0b\u66f4\u52a0\u660e\u663e\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">6.2 \u5b9e\u9a8c\u8bc1\u660e<\/h4>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0cTransformer\u6a21\u578b\u5728\u5904\u7406\u8ba1\u6570\u4efb\u52a1\u65f6\u8868\u73b0\u4e0d\u4f73\uff0c\u7279\u522b\u662f\u5728\u5e8f\u5217\u957f\u5ea6\u589e\u52a0\u65f6\u3002\u6a21\u578b\u5728\u957f\u5e8f\u5217\u4e2d\u7684\u8868\u73b0\u8fc5\u901f\u6076\u5316\uff0c\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u7406\u8bba\u5206\u6790\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7. \u89e3\u51b3\u65b9\u6848<\/h3>\n\n\n\n<p>\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u4e9b\u7b80\u5355\u7684\u89e3\u51b3\u65b9\u6848\u6765\u7f13\u89e3\u8868\u793a\u5d29\u6e83\u548c\u4fe1\u606f\u8fc7\u5ea6\u538b\u7f29\u95ee\u9898\u3002\u4f8b\u5982\uff0c\u901a\u8fc7\u5f15\u5165\u989d\u5916\u7684token(\ud83d\udc53)\u6765\u4fdd\u6301\u8868\u793a\u7684\u8ddd\u79bb\uff0c\u6216\u901a\u8fc7\u8c03\u6574\u6d6e\u70b9\u7cbe\u5ea6\u6765\u51cf\u5c11\u8868\u793a\u5d29\u6e83\u7684\u5f71\u54cd\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u8fd9\u4e9b\u89e3\u51b3\u65b9\u6848\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u6539\u5584\u4e86\u6a21\u578b\u5728\u590d\u5236\u548c\u8ba1\u6570\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">8. \u7ed3\u8bba\u4e0e\u672a\u6765\u5de5\u4f5c<\/h3>\n\n\n\n<p>\u672c\u6587\u63ed\u793a\u4e86\u4ec5\u89e3\u7801\u5668Transformer\u5728\u5904\u7406\u67d0\u4e9b\u57fa\u672c\u4efb\u52a1\u65f6\u7684\u8868\u793a\u95ee\u9898\uff0c\u5e76\u63d0\u51fa\u4e86\u7f13\u89e3\u8fd9\u4e9b\u95ee\u9898\u7684\u89e3\u51b3\u65b9\u6848\u3002\u672a\u6765\u7684\u7814\u7a76\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63a2\u7d22\u5982\u4f55\u6d4b\u91cf\u548c\u6539\u5584Transformer\u6a21\u578b\u4e2d\u7684\u4fe1\u606f\u4f20\u64ad\u95ee\u9898\uff0c\u4ee5\u63d0\u5347\u73b0\u6709\u8bed\u8a00\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e00\u4e9b\u7ec6\u8282\u5185\u5bb9\uff1a<\/h3>\n\n\n\n<h5 class=\"wp-block-heading\">1)\u4f4e\u7cbe\u5ea6\u6743\u91cd\/Quantisation\u53ef\u663e\u8457\u52a0\u901f\u63a8\u7406\u901f\u5ea6\uff0c\u4f46\u4e5f\u6709\u4ee3\u4ef7<\/h5>\n\n\n\n<p>\u4e00\u79cd\u5e38\u7528\u4e8e\u52a0\u901fLLM\u63a8\u7406\u7684\u6280\u672f\u662fQuantisation\u3002Quantisation\u662f\u4e00\u79cd\u6784\u5efaLLM\u8fd1\u4f3c\u7248\u672c\u7684\u8fc7\u7a0b\uff0c\u8fd9\u79cd\u8fd1\u4f3c\u7248\u672c\u4f7f\u7528\u4f4e\u7cbe\u5ea6\u7684\u6570\u636e\u7c7b\u578b\u8fd0\u884c\u3002\u8fd9\u6709\u52a9\u4e8e\u663e\u8457\u63d0\u9ad8LLM\u7684\u63a8\u7406\u901f\u5ea6\uff0c\u56e0\u4e3a\u73b0\u4ee3\u52a0\u901f\u5668\u5728\u4f4e\u7cbe\u5ea6\u6570\u636e\u7c7b\u578b\u4e0b\u80fd\u4ea7\u751f\u663e\u8457\u66f4\u591a\u7684FLOPs\u3002\u5f53\u7136\uff0cQuantisation\u901a\u5e38\u662f\u6709\u4ee3\u4ef7\u7684\u3002\u6211\u4eec\u7684\u7406\u8bba\u5206\u6790\u8868\u660e\uff0cQuantisation\u53ef\u80fd\u5bfc\u81f4\u8868\u793a\u7684\u6f5c\u5728\u707e\u96be\u6027\u4e27\u5931\u3002\u7279\u522b\u662f\uff0c\u66f4\u4f4e\u7684\u673a\u5668\u7cbe\u5ea6\u610f\u5473\u7740\u8868\u793a\u6536\u655b\u4f1a\u53d1\u751f\u5f97\u66f4\u65e9\uff0c\u56e0\u6b64LLM\u751a\u81f3\u65e0\u6cd5\u533a\u5206\u66f4\u77ed\u7684\u5e8f\u5217\u3002<\/p>\n\n\n\n<p>\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0ctokenisation\u4f7f\u5f97\u7406\u8bba\u7ed3\u679c\u7684\u76f4\u63a5\u5e94\u7528\u53d8\u5f97\u66f4\u52a0\u590d\u6742\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a\u91cd\u590dtoken\u7684\u5e8f\u5217\u201c11111\u201d\u4e0d\u4e00\u5b9a\u4f1a\u88abtokenised\u62105\u4e2a\u72ec\u7acb\u7684\u201c1\u201dtoken\u3002\u4ece\u539f\u5219\u4e0a\u8bb2\uff0c\u8fd9\u5e94\u8be5\u6709\u52a9\u4e8e\u7f13\u89e3\u8868\u793a\u7684\u76f4\u63a5\u5d29\u6e83\u3002tokenisation\u603b\u4f53\u4e0a\u589e\u52a0\u4e86\u7814\u7a76\u8fd9\u79cd\u73b0\u8c61\u7684\u96be\u5ea6\uff0c\u56e0\u4e3a\u5b83\u4e3a\u5b9e\u9a8c\u5206\u6790\u589e\u52a0\u4e86\u989d\u5916\u7684\u590d\u6742\u6027\u3002\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u8003\u8651\u4e86tokenisation\u5e76\u8bd5\u56fe\u51cf\u8f7b\u5176\u5f71\u54cd\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2)\u9488\u5bf9<em>\u8868\u793a\u5d29\u6e83<\/em>\u7684\u7b80\u5355\u89e3\u51b3\u65b9\u6848<\/h5>\n\n\n\n<p>Transformer\u5728\u5904\u7406\u957f\u5e8f\u5217\u7684\u91cd\u590dtoken\u65f6\u4f1a\u9047\u5230\u6311\u6218\u3002\u5bf9\u6b64\u95ee\u9898\u7684\u4e00\u4e2a\u5b9e\u9645\u89e3\u51b3\u65b9\u6848\u662f\u5728\u6574\u4e2a\u5e8f\u5217\u4e2d\u5f15\u5165\u989d\u5916\u7684token\uff0c\u4ee5\u5e2e\u52a9\u4fdd\u6301\u8868\u793a\u7684\u8ddd\u79bb\u3002\u4f8b\u5982\u5f53\u5904\u7406\u4e00\u4e2a\u5168\u662f1\u7684\u5e8f\u5217\u65f6\uff08c\uff09\u53d1\u751f\u4e86\u8868\u793a\u5d29\u6e83\uff0c\u800c\u6bcf\u9694\u4e09\u4e2a\u6570\u5b57\u6dfb\u52a0\u9017\u53f7\u5219\u6709\u52a9\u4e8e\u4fdd\u6301\u8868\u793a\u7684\u5206\u79bb\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Federico Barbero\u7b49\u8fd1\u671f\u53d1\u8868\u8bba\u6587\uff1aTransformers need glasses\ud83d\udc53: Inf [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","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":"","ast-disable-related-posts":"","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":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"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":""},"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":[8,36],"tags":[39,95,90,86,66],"class_list":["post-3929","post","type-post","status-publish","format-standard","hentry","category-tech","category-36","tag-ai","tag-transformer","tag-multimodal","tag-llm","tag-google"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a> <a href=\"https:\/\/nullthought.net\/?cat=36\" rel=\"category\">\u79d1\u5b66<\/a>","rttpg_excerpt":"Federico Barbero\u7b49\u8fd1\u671f\u53d1\u8868\u8bba\u6587\uff1aTransformers need glasses\ud83d\udc53: Inf&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/3929","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3929"}],"version-history":[{"count":3,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/3929\/revisions"}],"predecessor-version":[{"id":3935,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/3929\/revisions\/3935"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3929"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3929"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3929"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}