{"id":56784,"date":"2025-02-13T11:59:21","date_gmt":"2025-02-13T03:59:21","guid":{"rendered":"https:\/\/swarma.org\/?p=56784"},"modified":"2025-02-13T11:59:21","modified_gmt":"2025-02-13T03:59:21","slug":"max-tegmark%e7%bb%84%e6%96%b0%e5%b7%a5%e4%bd%9c%ef%bc%9a%e5%88%a9%e7%94%a8%e8%b0%83%e5%92%8c%e6%8d%9f%e5%a4%b1%e8%ae%ad%e7%bb%83%e5%8f%af%e8%a7%a3%e9%87%8a%e7%9a%84ai%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=56784","title":{"rendered":"Max Tegmark\u7ec4\u65b0\u5de5\u4f5c\uff1a\u5229\u7528\u8c03\u548c\u635f\u5931\u8bad\u7ec3\u53ef\u89e3\u91ca\u7684AI\u6a21\u578b"},"content":{"rendered":"
\n
\n
\n
\n
\n
\n
\n
\n

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

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

\u6458\u8981<\/strong><\/p>\n<\/section>\n<\/section>\n

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

\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u635f\u5931\u51fd\u6570\u8c03\u548c\u635f\u5931\uff08Harmonic Loss\uff09\uff0c\u4f5c\u4e3a\u6807\u51c6\u4ea4\u53c9\u71b5\u635f\u5931\uff08cross-entropy loss\uff09\u7684\u66ff\u4ee3\u65b9\u6848\uff0c\u7528\u4e8e\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u548c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u3002\u8c03\u548c\u635f\u5931\u7531\u4e8e\u5176\u5c3a\u5ea6\u4e0d\u53d8\u6027\u548c\u6709\u9650\u6536\u655b\u70b9\u7684\u8bbe\u8ba1\uff0c\u4f7f\u5f97\u6a21\u578b\u5177\u6709\u66f4\u5f3a\u7684\u53ef\u89e3\u91ca\u6027\u5e76\u5b9e\u73b0\u66f4\u5feb\u7684\u6536\u655b\uff0c\u5176\u4e2d\u6709\u9650\u6536\u655b\u70b9\u53ef\u4ee5\u88ab\u89e3\u91ca\u4e3a\u7c7b\u522b\u4e2d\u5fc3\u3002\u6211\u4eec\u9996\u5148\u5728\u7b97\u6cd5\u4efb\u52a1\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u548c\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e0a\u9a8c\u8bc1\u4e86 Harmonic Loss \u7684\u6027\u80fd\u3002\u901a\u8fc7\u4e00\u7cfb\u5217\u5b9e\u9a8c\uff0c\u6211\u4eec\u8bc1\u660e\u4e86\u91c7\u7528\u8c03\u548c\u635f\u5931\u8bad\u7ec3\u7684\u6a21\u578b\u5728\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\u4f18\u4e8e\u6807\u51c6\u6a21\u578b\uff1a<\/span><\/span><\/strong><\/span><\/strong>(a) \u63d0\u9ad8\u53ef\u89e3\u91ca\u6027\uff0c<\/span><\/span><\/strong><\/span><\/strong>(b) \u964d\u4f4e\u5bf9\u5927\u91cf\u8bad\u7ec3\u6570\u636e\u7684\u4f9d\u8d56\uff0c<\/span><\/span><\/strong><\/span><\/strong>(c) \u51cf\u5c11 grokking \u73b0\u8c61\uff08\u5ef6\u8fdf\u6cdb\u5316\uff09\u3002<\/span><\/span><\/strong><\/span><\/strong>\u6b64\u5916\uff0c\u6211\u4eec\u5bf9\u6bd4\u4e86\u91c7\u7528 Harmonic Loss \u8bad\u7ec3\u7684 GPT-2 \u4e0e\u6807\u51c6 GPT-2\uff0c\u7ed3\u679c\u8868\u660e Harmonic Loss \u4f7f\u6a21\u578b\u80fd\u591f\u5b66\u4e60\u5230\u66f4\u5177\u53ef\u89e3\u91ca\u6027\u7684\u8868\u793a\u3002\u5c55\u671b\u672a\u6765\uff0cHarmonic Loss \u6709\u671b\u6210\u4e3a\u6570\u636e\u53d7\u9650\u9886\u57df\u6216\u9ad8\u98ce\u9669\u5e94\u7528\u573a\u666f\uff08\u5982\u533b\u7597\u3001\u91d1\u878d\uff09\u4e2d\u7684\u91cd\u8981\u5de5\u5177\uff0c\u4fc3\u8fdb\u66f4\u7a33\u5065\u3001\u9ad8\u6548\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u7684\u53d1\u5c55\u3002<\/span><\/span><\/strong><\/span><\/strong><\/p>\n


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

Max Tegmark\u4e5f\u66fe\u53d7\u9080\u5728\u96c6\u667a\u4ff1\u4e50\u90e8\u5206\u4eab\uff1a\u6784\u5efa\u53ef\u63a7\u3001\u53ef\u89e3\u91ca\u7684AI\u7cfb\u7edf\uff1ahttps:\/\/pattern.swarma.org\/study_group_issue\/503<\/span><\/span><\/strong><\/em><\/span><\/p>\n<\/section>\n

\n

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

\u7814\u7a76\u9886\u57df\uff1a\u8c03\u548c\u635f\u5931\u3001\u53ef\u89e3\u91ca\u6027\u3001\u795e\u7ecf\u7f51\u7edc\u3001\u901a\u7528\u5316\u3001grokking\u73b0\u8c61\u3001\u5927\u8bed\u8a00\u6a21\u578b<\/span><\/strong><\/strong><\/span><\/strong><\/strong><\/strong><\/section>\n<\/section>\n
\n
<\/strong><\/strong><\/strong><\/section>\n<\/section>\n
\n
<\/strong><\/strong><\/section>\n
<\/strong><\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

 <\/span><\/span><\/p>\n

<\/section>\n
\n
\u8bba\u6587\u9898\u76ee\uff1aHarmonic Loss Trains Interpretable AI Models<\/span><\/section>\n
\u53d1\u8868\u65f6\u95f4\uff1a2025\u5e742\u67083\u65e5<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2502.01628<\/span><\/section>\n<\/blockquote>\n
<\/section>\n
\u8fd1\u5e74\u6765\uff0c\u795e\u7ecf\u7f51\u7edc\u548c\u5927\u578b\u8bed\u8a00\u6a21\u578b<\/span>\uff08Large Language Models, LLMs\uff09<\/span>\u7684\u5e7f\u6cdb\u5e94\u7528\u6781\u5927\u5730\u6539\u53d8\u4e86\u4eba\u4eec\u7684\u751f\u6d3b\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u6a21\u578b\u5728\u901a\u7528\u5316\u80fd\u529b\u4e0a\u4ecd\u9762\u4e34\u7740\u53ef\u89e3\u91ca\u6027\u3001\u6570\u636e\u6548\u7387\u548cgrokking\u73b0\u8c61\u7b49\u6311\u6218\u3002\u8fd1\u671fMax Tegmark\u7ec4\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u635f\u5931\u51fd\u6570\u2014\u2014\u8c03\u548c\u635f\u5931<\/span>\uff08Harmonic Loss\uff09<\/span>\uff0c\u4f5c\u4e3a\u4f20\u7edf\u4ea4\u53c9\u71b5\u635f\u5931\u7684\u66ff\u4ee3\u65b9\u6848\uff0c\u53ef\u663e\u8457\u63d0\u9ad8\u5927\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u548c\u8bad\u7ec3\u6548\u7387\u3002<\/span><\/section>\n

<\/span><\/section>\n
\n
\u6587\u7ae0\u5171\u540c\u7b2c\u4e00\u4f5c\u8005\u5218\u5b50\u9e23\u5728\u96c6\u667a\u4ff1\u4e50\u90e8\u5c31\u795e\u7ecf\u6807\u5ea6\u5f8b\u505a\u8fc7\u76f8\u5173\u62a5\u544a\uff0c\u63ed\u793a\u5927\u6a21\u578b\u7684\u9884\u6d4b\u80fd\u529b\u968f\u7740\u66f4\u591a\u7684\u6570\u636e\u548c\u66f4\u5927\u7684\u6a21\u578b\u800c\u5e42\u5f8b\u63d0\u5347\u8fd9\u4e00\u73b0\u8c61\uff0c\u5176\u80cc\u540e\u7684\u7406\u8bba\u548c\u5bf9AI+Science\u5e26\u6765\u7684\u542f\u53d1\u3002<\/span><\/section>\n
\u56de\u770b\u5730\u5740\uff1ahttps:\/\/pattern.swarma.org\/study_group_issue\/446?play_start_time=425<\/span><\/section>\n
\u4ee5\u53ca\u4f5c\u4e3a24\u5e74\u7206\u706b\u7684\u5168\u65b0\u6df1\u5ea6\u5b66\u4e60\u7f51\u7edc\u7ed3\u6784KAN\u67b6\u6784\u7684\u63d0\u51fa\u8005\uff0c\u5218\u5b50\u9e23\u672c\u4eba\u5728\u96c6\u667a\u4ff1\u4e50\u90e8\u5bf9\u5176\u5de5\u4f5c\u8fdb\u884c\u4e86\u6df1\u5165\u89e3\u8bfb\uff0c\u611f\u5174\u8da3\u8005\u53ef\u52a0\u5165AI+Science\u8bfb\u4e66\u4f1a<\/a>\u3002<\/span><\/section>\n
\u56de\u770b\u5730\u5740\uff1ahttps:\/\/pattern.swarma.org\/study_group_issue\/668<\/span><\/section>\n<\/blockquote>\n

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

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

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

\u8c03\u548c\u635f\u5931\u7684\u539f\u7406\u4e0e\u4f18\u52bf<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

<\/p>\n

\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u635f\u5931\u51fd\u6570\u8c03\u548c\u635f\u5931<\/strong>\uff0c\u7528\u4e8e\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u548c\u5927\u8bed\u8a00\u6a21\u578b\u3002\u4f20\u7edf\u7684\u635f\u5931\u8ba1\u7b97\u4f7f\u7528\u7684\u4ea4\u53c9\u71b5\u635f\u5931\uff0c\u6838\u5fc3\u8ba1\u7b97\u7684\u662f\u5185\u79ef\uff0c\u5e76\u8fdb\u884cSoftmax\u5f52\u4e00\u5316\uff1b\u800c\u8c03\u548c\u635f\u5931\u51fd\u6570\u8ba1\u7b97\u5219\u4f7f\u7528\u6b27\u51e0\u91cc\u5f97\u8ddd\u79bb\u66ff\u4ee3\u5185\u79ef\uff0c\u5e76\u901a\u8fc7Harmonic Softmax\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u4f5c\u4e3a\u5411\u91cf\u76f8\u4f3c\u6027\u5ea6\u91cf<\/strong>\u3002\u4ece\u800c\u63d0\u9ad8\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u3001\u6570\u636e\u5229\u7528\u6548\u7387\uff0c\u5e76\u51cf\u5c11\u5ef6\u8fdf\u6cdb\u5316<\/span>\uff08grokking\uff09<\/span>\u73b0\u8c61\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u8c03\u548c\u635f\u5931\u5177\u6709\u4e24\u5927\u6838\u5fc3\u6570\u5b66\u7279\u6027\uff1a\u5c3a\u5ea6\u4e0d\u53d8\u6027<\/strong>\u548c\u6709\u9650\u6536\u655b\u70b9<\/strong>\u3002Harmonic Loss \u5177\u6709\u5c3a\u5ea6\u4e0d\u53d8\u6027<\/strong>\uff0c\u5373\u5f53\u8f93\u5165\u5411\u91cf\u6574\u4f53\u7f29\u653e\u65f6\uff0c\u6a21\u578b\u7684\u635f\u5931\u503c\u548c\u4f18\u5316\u65b9\u5411\u4fdd\u6301\u4e0d\u53d8\uff0c\u4ece\u800c\u63d0\u9ad8\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u3002\u76f8\u6bd4\u4e8e\u4f20\u7edf\u4ea4\u53c9\u71b5\u635f\u5931\u9700\u8981\u5c06 logits \u63a8\u5411\u65e0\u7a77\u5927\u624d\u80fd\u8fbe\u5230\u786e\u5b9a\u6027\u5206\u7c7b\uff0cHarmonic Loss \u4f9d\u8d56\u4e8e\u6709\u9650\u6536\u655b\u70b9<\/strong>\uff0c\u4f7f\u5f97\u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u80fd\u591f\u66f4\u5feb\u6536\u655b\uff0c\u5e76\u83b7\u5f97\u66f4\u52a0\u7d27\u51d1\u548c\u53ef\u89e3\u91ca\u7684\u7279\u5f81\u8868\u793a\u3002<\/span><\/section>\n

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

\u56fe 1. \u4ea4\u53c9\u71b5\u635f\u5931\u4e0e\u8c03\u548c\u635f\u5931\u3002(a) \u5b9a\u4e49\uff1a\u4ea4\u53c9\u71b5\u635f\u5931\u5229\u7528\u5185\u79ef\u4f5c\u4e3a\u76f8\u4f3c\u6027\u5ea6\u91cf\uff0c\u800c\u8c03\u548c\u635f\u5931\u4f7f\u7528\u6b27\u51e0\u91cc\u5f97\u8ddd\u79bb\u3002(b) \u73a9\u5177\u6a21\u578b1\uff1a\u67092\u4e2a\u7684\u70b9\uff08\u7c7b\uff09\u3002\u8c03\u548c\u635f\u8017\u548cL2\u8303\u6570\u5728\u8c03\u548c\u635f\u8017\u4e0b\u6536\u655b\u5f97\u66f4\u5feb\u3002(c) \u73a9\u5177\u6a21\u578b2\uff1a\u67095\u4e2a\u70b9\uff08\u7c7b\uff09\u3002\u8c03\u548c\u635f\u5931\u53ef\u4ee5\u6311\u51fa\u4e2d\u95f4\u7684\u7ea2\u70b9\uff0c\u800c\u4ea4\u53c9\u71b5\u635f\u5931\u4e0d\u80fd\uff0c\u56e0\u4e3a\u7ea2\u70b9\u4e0e\u5176\u4ed6\u70b9\u4e0d\u80fd\u7ebf\u6027\u5206\u79bb\u3002\u52a0\u6743\u77e9\u9635\u5728\u8c03\u548c\u635f\u5931\u4e0b\u6bd4\u5728\u4ea4\u53c9\u71b5\u635f\u5931\u4e0b\u66f4\u6613\u4e8e\u89e3\u91ca\u3002<\/span><\/p>\n


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

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

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

\u5b9e\u9a8c\u65b9\u6cd5\u4e0e\u7ed3\u679c<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

<\/p>\n

\u7814\u7a76\u9996\u5148\u5229\u7528\u7b97\u6cd5\u6570\u636e\u96c6\u8fdb\u884c\u6d4b\u8bd5\uff0c\u5171\u4f7f\u7528 5 \u4e2a\u4efb\u52a1\uff1a\u4e0a\u4e0b\u6587\u5b66\u4e60<\/span>\uff08In-Context Learning\uff09<\/span>\uff0c\u6a21\u5757\u5316\u52a0\u6cd5<\/span>\uff08Modular Addition\uff09<\/span>\uff0c\u7b49\u4ef7\u7c7b\u5224\u5b9a<\/span>\uff08Equivalence Classes\uff09<\/span>\uff0c\u5bb6\u8c31\u63a8\u7406<\/span>\uff08Genealogy Learning\uff09<\/span>\uff0c \u6392\u5217\u7ec4\u5408<\/span>\uff08Permutation Composition\uff09<\/span>\uff0c\u6bd4\u8f83\u6807\u51c6 MLP\u3001\u6807\u51c6 Transformer\u3001Harmonic MLP \u548c Harmonic Transformer \u6a21\u578b\u7684\u8868\u73b0\u3002\u8c03\u548c\u635f\u5931\u5728\u6240\u6709\u4efb\u52a1\u4e0a\u5747\u5b66\u4e60\u5230\u4e86\u66f4\u7d27\u51d1\u7684\u4f4e\u7ef4\u7ed3\u6784\uff0c\u5e76\u4e14\u6570\u636e\u5229\u7528\u6548\u7387\u66f4\u9ad8\uff0c\u9700\u8981\u66f4\u5c11\u7684\u6570\u636e\u5373\u53ef\u8fbe\u5230\u76f8\u540c\u7684\u6cdb\u5316\u6027\u80fd\u3002\u8c03\u548c\u635f\u5931\u7684\u5f15\u5165\u8fd8\u51cf\u5c11\u4e86\u6a21\u578b Grokking\u73b0\u8c61<\/strong>\uff0c\u4f8b\u5982\uff0c\u5728\u6a21\u5757\u5316\u52a0\u6cd5\u4efb\u52a1\u4e2d\uff0c\u6807\u51c6 MLP \u9700\u8981 10\u2075 \u8f6e\u8bad\u7ec3\u624d\u80fd\u6cdb\u5316\uff0c\u800c Harmonic MLP \u4ec5\u9700 10\u00b3 \u8f6e\u8bad\u7ec3\u3002<\/span><\/section>\n

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

\u56fe 2. \u7efc\u5408\u6d4b\u8bd5\u4e2d\u5d4c\u5165\u7684\u524d\u4e24\u4e2a\u4e3b\u6210\u5206\u7684\u53ef\u89c6\u5316\u3002\u6bcf\u4e2a\u5b50\u56fe\u7684\u6807\u9898\u663e\u793a\u4e86\u524d\u4e24\u4e2a\u4e3b\u6210\u5206\u6240\u89e3\u91ca\u7684\u65b9\u5dee\u3002\u6bcf\u884c\u5bf9\u5e94\u4e00\u4e2a\u6570\u636e\u96c6\u548c\u4e00\u4e2a\u6a21\u578b\u7684\u7ec4\u5408\uff0c\u800c\u6bcf\u5217\u4ee3\u8868\u4f7f\u7528\u4e0d\u540c\u968f\u673a\u79cd\u5b50\u8fdb\u884c\u7684\u4e0d\u540c\u8bad\u7ec3\u8fd0\u884c\u7684\u5d4c\u5165\u3002\u8fde\u7eed\u7684\u4e24\u884c\u5c5e\u4e8e\u540c\u4e00\u4e2a\u6570\u636e\u96c6\uff0c\u6a21\u578b\u6309\u4ee5\u4e0b\u987a\u5e8f\u6392\u5217\uff1a{\u6807\u51c6\u591a\u5c42\u611f\u77e5\u673a\uff0c\u8c03\u548c\u591a\u5c42\u611f\u77e5\u673a}\u3002\u6570\u636e\u96c6\u7684\u987a\u5e8f\u5982\u4e0b\uff1a{\u4e0a\u4e0b\u6587\u5b66\u4e60\uff0c\u8c31\u7cfb\u5b66\u4e60\uff0c\u7b49\u4ef7\u7c7b\uff0c\u6a21\u5757\u5316\u52a0\u6cd5\u548c\u7f6e\u6362\u7fa4}\u3002X \u8f74\u548c Y \u8f74\u7684\u8de8\u5ea6\u76f8\u7b49\u3002<\/span><\/p>\n


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

\u56fe 3.\uff08a\uff09\u4e3b\u6210\u5206\u51fd\u6570\u7684\u7d2f\u79ef\u89e3\u91ca\u65b9\u5dee\uff0820 \u4e2a\u79cd\u5b50\u7684\u4e2d\u4f4d\u6570\uff09\u3002\u8c03\u548c\u8868\u793a\u6bd4\u6807\u51c6\u8868\u793a\u66f4\u7d27\u51d1\u3002\uff08b\uff09\u8bad\u7ec3\u6bd4\u4f8b\u51fd\u6570\u7684\u6d4b\u8bd5\u51c6\u786e\u7387\u3002\u8c03\u548c\u6a21\u578b\u5728\u6570\u636e\u66f4\u5c11\u7684\u60c5\u51b5\u4e0b\u6bd4\u6807\u51c6\u6a21\u578b\u66f4\u5feb\u5730\u6cdb\u5316\u3002\uff08c\uff09\u6d4b\u8bd5\u51c6\u786e\u7387\uff1e0.9 \u6240\u9700\u7684\u8f6e\u6b21\u4e0e\u8bad\u7ec3\u51c6\u786e\u7387\uff1e0.9 \u6240\u9700\u7684\u8f6e\u6b21\uff08\u8fde\u7eed 20 \u6b21\uff09\u3002y = x \u7ebf\u8868\u793a\u6ca1\u6709grokking\uff0c\u5373\u8bad\u7ec3\u51c6\u786e\u7387\u548c\u6d4b\u8bd5\u51c6\u786e\u7387\u540c\u65f6\u63d0\u9ad8\u3002\u9760\u8fd1 y \u8f74\u7684\u70b9\u8868\u793a\u638c\u63e1\u7a0b\u5ea6\u66f4\u9ad8\u3002\u7ed8\u5236\u4e86 20 \u4e2a\u4e0d\u540c\u968f\u673a\u79cd\u5b50\u7684\u7ed3\u679c\uff0c\u672a\u80fd\u8fbe\u5230 90%\u51c6\u786e\u7387\u7684\u6d4b\u8bd5\u88ab\u7701\u7565\u3002<\/span><\/p>\n


<\/span><\/section>\n
\u5728MNIST\u7684\u5b9e\u9a8c\u4e2d\uff0c\u4f7f\u7528\u8c03\u548c\u635f\u5931\u7684\u7b80\u5355\u795e\u7ecf\u7f51\u7edc\u80fd\u591f\u5b66\u5230\u66f4\u52a0\u53ef\u89e3\u91ca\u7684\u7279\u5f81\uff0c\u663e\u793a\u51fa\u4e0e\u7c7b\u522b\u4e2d\u5fc3\u5bf9\u9f50\u7684\u6743\u91cd\u5206\u5e03\uff0c\u8fd9\u8868\u660e\u8c03\u548c\u635f\u5931\u5728\u89c6\u89c9\u4efb\u52a1\u4e2d\u540c\u6837\u6709\u6548\u3002\u800c\u5728\u5927\u8bed\u8a00\u6a21\u578bGPT-2\u7684\u5b9e\u9a8c\u4e2d\uff0c\u7814\u7a76\u8868\u660e\uff0c\u8c03\u548c\u635f\u5931\u8bad\u7ec3\u7684GPT-2 \u8bad\u7ec3\u901f\u5ea6\u66f4\u5feb\uff0c\u5177\u6709\u66f4\u597d\u7684\u53ef\u89e3\u91ca\u6027\uff0c\u5e76\u4e14\u6027\u80fd\u4e0d\u900a\u4e8e\u6807\u51c6 GPT-2\u3002<\/strong>\u5728\u53ef\u89e3\u91ca\u4e0a\uff0c\u7814\u7a76\u91c7\u7528 \u201cpresent-past\u201d<\/span>\uff08\u52a8\u8bcd\u65f6\u6001\u8f6c\u6362\uff09<\/span>\u4efb\u52a1\u8fdb\u884c\u6d4b\u8bd5\uff0c\u8c03\u548c\u635f\u5931\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\u5728\u751f\u6210\u7684\u8868\u793a\u4e0a\u66f4\u4f18\uff0c\u751f\u6210\u7684\u8bcd\u5411\u91cf\u5448\u73b0\u66f4\u89c4\u5219\u7684\u5e73\u884c\u56db\u8fb9\u5f62\u7ed3\u6784\uff0c\u800c\u6807\u51c6 GPT-2 \u751f\u6210\u7684\u8bcd\u5411\u91cf\u5448\u73b0\u6241\u5e73\u7684\u3001\u5f62\u72b6\u4e0d\u89c4\u5219\u7684\u7ed3\u6784\uff0c\u6697\u793a\u8c03\u548c\u635f\u5931\u80fd\u6355\u6349\u5230\u66f4\u6df1\u5c42\u6b21\u7684\u8bed\u8a00\u6a21\u5f0f\u3002<\/span><\/section>\n

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

\u56fe 4. GPT2 \u5b9e\u9a8c\uff0c\u57fa\u4e8e OpenWebText \u8bad\u7ec3\u4e86 10000 \u6b65\u3002\uff08\u5de6\u4e0a\u89d2\uff09\u635f\u5931\u66f2\u7ebf\u3002Harmonic GPT \u8fbe\u5230\u7684\u635f\u5931\u7565\u4f4e\u4e8e\u6807\u51c6 GPT\u3002\uff08\u53f3\u4e0a\u89d2\uff09\u9488\u5bf9\u5341\u4e8c\u4e2a\u51fd\u6570\u5411\u91cf\u4efb\u52a1\u7684\u5e73\u884c\u56db\u8fb9\u5f62\u635f\u5931\u7684\u7d2f\u79ef\u5206\u5e03\u51fd\u6570\u3002Harmonic GPT \u4e00\u76f4\u8868\u73b0\u51fa\u66f4\u4f4e\u7684\u5e73\u884c\u56db\u8fb9\u5f62\u635f\u5931\uff08\u5373\u66f4\u597d\u7684\u5e73\u884c\u56db\u8fb9\u5f62\uff09\u3002\uff08\u5e95\u90e8\uff09\u6309\u8d28\u91cf\u4ece\u5de6\u5230\u53f3\u964d\u5e8f\u6392\u5217\u7684\u5e73\u884c\u56db\u8fb9\u5f62\uff08\u7b2c\u4e00\u548c\u7b2c\u4e8c\u4e3b\u6210\u5206\uff09\u3002Harmonic GPT \u503e\u5411\u4e8e\u751f\u6210\u66f4\u201c\u77e9\u5f62\u201d\u7684\u5e73\u884c\u56db\u8fb9\u5f62\uff0c\u800c\u6807\u51c6 GPT \u5219\u751f\u6210\u6241\u5e73\u7684\u201c\u5e73\u884c\u56db\u8fb9\u5f62\u201d\u3002<\/span><\/p>\n


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

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

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

\u7ed3\u8bba\u4e0e\u5c55\u671b<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

<\/p>\n

\u672c\u6587\u901a\u8fc7\u5f15\u5165\u8c03\u548c\u635f\u5931\uff0c\u4e3a\u795e\u7ecf\u7f51\u7edc\u548c\u5927\u578b\u8bed\u8a00\u6a21\u578b\u63d0\u4f9b\u4e86\u4e00\u79cd\u65b0\u7684\u8bad\u7ec3\u673a\u5236\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u548c\u8bad\u7ec3\u6548\u7387\u3002\u672a\u6765\u7684\u7814\u7a76\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63a2\u7d22\u8c03\u548c\u635f\u5931\u5728\u66f4\u5927\u89c4\u6a21\u6a21\u578b\u4e2d\u7684\u9002\u7528\u6027\u53ca\u5176\u5728\u5404\u7c7b\u5e94\u7528\u573a\u666f\u4e2d\u7684\u6f5c\u529b\uff0c\u4e3a\u6784\u5efa\u66f4\u9c81\u68d2\u548c\u9ad8\u6548\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u94fa\u5e73\u9053\u8def\u3002<\/span><\/section>\n
<\/span><\/section>\n
<\/section>\n

<\/p>\n

\u5f6d\u6668<\/span> <\/span>| \u7f16\u8bd1<\/span><\/strong><\/span><\/strong><\/p>\n

<\/p>\n

<\/p>\n

\n
\n
\n
\n

\u5927\u6a21\u578b2.0\u8bfb\u4e66\u4f1a\u542f\u52a8<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

<\/section>\n

o1\u6a21\u578b\u4ee3\u8868\u5927\u8bed\u8a00\u6a21\u578b\u878d\u5408\u5b66\u4e60\u4e0e\u63a8\u7406\u7684\u65b0\u8303\u5f0f\u3002\u96c6\u667a\u4ff1\u4e50\u90e8\u8054\u5408\u5317\u4eac\u5e08\u8303\u5927\u5b66\u7cfb\u7edf\u79d1\u5b66\u5b66\u9662\u6559\u6388\u5f20\u6c5f\u3001Google DeepMind\u7814\u7a76\u79d1\u5b66\u5bb6\u51af\u7199\u680b\u3001\u963f\u91cc\u5df4\u5df4\u5f3a\u5316\u5b66\u4e60\u7814\u7a76\u5458\u738b\u7ef4\u57d9\u548c\u4e2d\u79d1\u9662\u4fe1\u5de5\u6240\u5f20\u6770\u5171\u540c\u53d1\u8d77\u300c\u5927\u6a21\u578bII\uff1a\u878d\u5408\u5b66\u4e60\u4e0e\u63a8\u7406\u7684\u5927\u6a21\u578b\u65b0\u8303\u5f0f \u300d\u8bfb\u4e66\u4f1a<\/a>\uff0c\u672c\u6b21\u8bfb\u4e66\u4f1a\u5c06\u5173\u6ce8\u5927\u6a21\u578b\u63a8\u7406\u8303\u5f0f\u7684\u6f14\u8fdb\u3001\u57fa\u4e8e\u641c\u7d22\u4e0e\u8499\u7279\u5361\u6d1b\u6811\u7684\u63a8\u7406\u4f18\u5316\u3001\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684\u5927\u6a21\u578b\u4f18\u5316\u3001\u601d\u7ef4\u94fe\u65b9\u6cd5\u4e0e\u5185\u5316\u673a\u5236\u3001\u81ea\u6211\u6539\u8fdb\u4e0e\u63a8\u7406\u9a8c\u8bc1\u3002\u5e0c\u671b\u901a\u8fc7\u8bfb\u4e66\u4f1a\u63a2\u7d22o1\u5177\u4f53\u5b9e\u73b0\u7684\u6280\u672f\u8def\u5f84\uff0c\u5e2e\u52a9\u6211\u4eec\u66f4\u597d\u7684\u7406\u89e3\u673a\u5668\u63a8\u7406\u548c\u4eba\u5de5\u667a\u80fd\u7684\u672c\u8d28\u3002<\/span><\/p>\n


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

\u4ece2024\u5e7411\u670830\u65e5\u5f00\u59cb\uff0c\u9884\u8ba1\u6bcf\u5468\u516d\u8fdb\u884c\u4e00\u6b21\uff0c\u6301\u7eed\u65f6\u95f4\u9884\u8ba1 6-8 \u5468\u5de6\u53f3\u3002\u6b22\u8fce\u611f\u5174\u8da3\u7684\u670b\u53cb\u62a5\u540d\u53c2\u52a0\uff0c\u6fc0\u53d1\u66f4\u591a\u7684\u601d\u7ef4\u706b\u82b1\uff01<\/span><\/p>\n


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

<\/span><\/span><\/a><\/p>\n


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

\u8be6\u60c5\u8bf7\u89c1\uff1a\u5927\u6a21\u578b2.0\u8bfb\u4e66\u4f1a\uff1a\u878d\u5408\u5b66\u4e60\u4e0e\u63a8\u7406\u7684\u5927\u6a21\u578b\u65b0\u8303\u5f0f\uff01<\/a><\/span><\/span><\/p>\n

\n

<\/p>\n

<\/section>\n<\/section>\n
\u63a8\u8350\u9605\u8bfb<\/span><\/strong><\/strong><\/strong><\/section>\n
<\/strong><\/section>\n
1. <\/strong>\u5927\u8bed\u8a00\u6a21\u578b\u4f1a\u63a8\u7406\u5417\uff1f<\/a><\/strong><\/span><\/strong><\/section>\n
2. <\/strong>\u5927\u6a21\u578b\u9884\u6d4b\u95ee\u9898\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u7528\u4e0d\u540c\u80fd\u529b\u7684LLMs\u534f\u4f5c\u63d0\u9ad8\u63a8\u7406\u6548\u7387<\/a><\/strong><\/span><\/strong><\/section>\n
3. <\/strong>\u56e0\u679c\u63a8\u7406\u4e0e\u5927\u8bed\u8a00\u6a21\u578b\uff1a\u5f00\u8f9f\u56e0\u679c\u5173\u7cfb\u7684\u65b0\u524d\u6cbf<\/a><\/strong><\/span><\/strong><\/section>\n
<\/strong><\/section>\n
4. <\/strong><\/span>\u5f20\u6c5f\uff1a\u7b2c\u4e09\u4ee3\u4eba\u5de5\u667a\u80fd\u6280\u672f\u57fa\u7840\u2014\u2014\u4ece\u53ef\u5fae\u5206\u7f16\u7a0b\u5230\u56e0\u679c\u63a8\u7406 | \u96c6\u667a\u5b66\u56ed\u5168\u65b0\u8bfe\u7a0b<\/a><\/strong><\/strong><\/span><\/section>\n
<\/strong>5. <\/strong>\u89e3\u9501\u96c6\u667a\u5168\u7ad9\u5185\u5bb9\uff0c\u5f00\u542f\u65b0\u5e74\u5b66\u4e60\u8ba1\u5212<\/a><\/strong><\/section>\n

6. <\/strong>\u52a0\u5165\u96c6\u667a\uff0c\u4e00\u8d77\u590d\u6742\uff01<\/strong><\/span><\/a><\/p>\n

<\/p>\n


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

\u70b9\u51fb\u201c\u9605\u8bfb\u539f\u6587\u201d\uff0c\u62a5\u540d\u8bfb\u4e66\u4f1a<\/span><\/section>\n<\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

\u6458\u8981 \u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u635f\u5931\u51fd\u6570\u8c03\u548c\u635f\u5931\uff08Harmonic Loss\uff09\uff0c\u4f5c\u4e3a\u6807\u51c6\u4ea4\u53c9\u71b5\u635f\u5931\uff08cross-entropy loss\uff09\u7684\u66ff\u4ee3\u65b9\u6848\uff0c\u7528\u4e8e\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u548c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u3002\u8c03\u548c\u635f\u5931\u7531\u4e8e\u5176\u5c3a\u5ea6\u4e0d\u53d8\u6027\u548c\u6709\u9650\u6536\u655b\u70b9\u7684\u8bbe\u8ba1\uff0c\u4f7f\u5f97\u6a21\u578b\u5177\u6709\u66f4\u5f3a\u7684\u53ef\u89e3\u91ca\u6027\u5e76\u5b9e\u73b0\u66f4\u5feb\u7684\u6536\u655b\uff0c\u5176\u4e2d\u6709\u9650\u6536\u655b\u70b9…<\/p>\n","protected":false},"author":1,"featured_media":56774,"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\/56784"}],"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"}],"author":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=56784"}],"version-history":[{"count":0,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/56784\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/media\/56774"}],"wp:attachment":[{"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=56784"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=56784"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=56784"},{"taxonomy":"special","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fspecial&post=56784"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}