{"id":35686,"date":"2022-07-07T11:23:16","date_gmt":"2022-07-07T03:23:16","guid":{"rendered":"https:\/\/swarma.org\/?p=35686"},"modified":"2022-07-07T11:23:16","modified_gmt":"2022-07-07T03:23:16","slug":"pnas%e9%80%9f%e9%80%92%ef%bc%9a%e5%9f%ba%e4%ba%8e%e7%9b%b8%e4%bc%bc%e5%8a%a0%e6%9d%83%e4%ba%a4%e5%8f%89%e5%ad%a6%e4%b9%a0%e7%9a%84%e6%b7%b1%e5%ba%a6%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c%e5%92%8c","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=35686","title":{"rendered":"PNAS\u901f\u9012\uff1a\u57fa\u4e8e\u76f8\u4f3c\u52a0\u6743\u4ea4\u53c9\u5b66\u4e60\u7684\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u548c\u5927\u8111\u5b66\u4e60"},"content":{"rendered":"
\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

\u7406\u89e3\u5927\u8111\u5728\u4e00\u751f\u7684\u8de8\u5ea6\u4e0b\u5982\u4f55\u5b66\u4e60\u4ecd\u7136\u662f\u957f\u671f\u7684\u6311\u6218\u3002\u5728\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\uff08ANNs\uff09\u4e2d\uff0c\u6574\u5408\u65b0\u4fe1\u606f\u7684\u901f\u5ea6\u8fc7\u5feb\u5e38\u5bfc\u81f4\u707e\u96be\u6027\u7684\u5e72\u6270\uff0c\u5373\u7a81\u7136\u5931\u53bb\u5148\u524d\u83b7\u5f97\u7684\u77e5\u8bc6\u3002\u4e92\u8865\u5b66\u4e60\u7cfb\u7edf\u7406\u8bba\uff08CLST\uff09\u8ba4\u4e3a\u65b0\u7684\u8bb0\u5fc6\u4f1a\u901a\u8fc7\u4e0e\u5df2\u6709\u77e5\u8bc6\u4ea4\u53c9\u5728\u4e00\u8d77\u7684\u65b9\u5f0f\u9010\u6e10\u6574\u5408\u8fdb\u65b0\u5927\u8111\u76ae\u5c42\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4e5f\u5c31\u5047\u5b9a\u4e86\u6bcf\u6b21\u5b66\u4e60\u65b0\u4e1c\u897f\u90fd\u9700\u8981\u4ea4\u53c9\u6240\u6709\u5df2\u77e5\u7684\u77e5\u8bc6\uff0c\u8fd9\u96be\u4ee5\u7f6e\u4fe1\uff0c\u56e0\u4e3a\u5b83\u4e0d\u4ec5\u8017\u65f6\u540c\u65f6\u9700\u8981\u5927\u91cf\u6570\u636e\u3002<\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/p>\n


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

\u8fd1\u65e5\u53d1\u8868\u4e8e PNAS \u7684\u6700\u65b0\u7814\u7a76\u5c55\u793a\u4e86\u6df1\u5ea6\u3001\u975e\u7ebf\u6027\u7684\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u53ef\u4ee5\u901a\u8fc7\u4ea4\u53c9\u4ec5\u4ec5\u4e00\u90e8\u5206\u7684\u8001\u77e5\u8bc6\u6765\u5b66\u4e60\u65b0\u4fe1\u606f\uff0c\u8fd9\u4e9b\u8001\u77e5\u8bc6\u548c\u65b0\u4fe1\u606f\u5728\u8868\u5f81\u4e0a\u6709\u7740\u5f88\u5927\u7684\u76f8\u4f3c\u5ea6\u3002\u5c31\u9760\u7740\u8fd9\u79cd\u76f8\u4f3c\u5ea6\u52a0\u6743\u4ea4\u53c9\u5b66\u4e60\uff08SWIL\uff09\u7684\u7b56\u7565\uff0c\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u80fd\u591f\u5728\u53d6\u5f97\u540c\u6837\u7cbe\u5ea6\u6c34\u5e73\u548c\u6700\u5c0f\u5e72\u6270\u7684\u60c5\u51b5\u4e0b\u5feb\u901f\u5b66\u4e60\u65b0\u4fe1\u606f\uff0c\u800c\u6bcf\u8f6e\u8bad\u7ec3\u53ea\u9700\u8981\u5f88\u5c11\u91cf\u7684\u8001\u77e5\u8bc6\uff08\u5feb\u901f\u4e14\u6570\u636e\u9ad8\u6548\uff09\u3002\u76f8\u4f3c\u5ea6\u52a0\u6743\u4ea4\u53c9\u5b66\u4e60\u7b56\u7565\u53ef\u4ee5\u5e94\u7528\u4e8e\u5404\u79cd\u6807\u51c6\u5206\u7c7b\u6570\u636e\u96c6\uff08Fashion-MNIST, CIFAR10, and CIFAR100\uff09\uff0c\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\uff0c\u4ee5\u53ca\u5e8f\u5217\u5b66\u4e60\u6846\u67b6\u4e4b\u4e2d\u3002\u7814\u7a76\u53d1\u73b0\uff0c\u6570\u636e\u6548\u7387\u548c\u5b66\u4e60\u65b0\u9879\u76ee\u7684\u901f\u5ea6\u4e0e\u5b58\u50a8\u5728\u7f51\u7edc\u4e2d\u7684\u975e\u91cd\u53e0\u7c7b\u7684\u6570\u91cf\u5927\u81f4\u6210\u6b63\u6bd4\u4f8b\uff0c\u8fd9\u53ef\u80fd\u6697\u793a\u7740\u4eba\u7c7b\u5927\u8111\u5728\u7f16\u7801\u5927\u91cf\u4e0d\u540c\u7c7b\u7684\u4e1c\u897f\u65f6\uff0c\u6570\u636e\u6548\u7387\u548c\u901f\u5ea6\u4f1a\u5927\u5927\u52a0\u5feb\u3002\u6700\u540e\uff0c\u8bba\u6587\u63d0\u51fa\u4e00\u4e2a\u5408\u9002\u7684\u7406\u8bba\u6a21\u578b\u6765\u89e3\u91ca\u76f8\u4f3c\u5ea6\u52a0\u6743\u4ea4\u53c9\u5b66\u4e60\u7b56\u7565\u5728\u5927\u8111\u4e2d\u7684\u5e94\u7528\u3002<\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/p>\n


<\/p>\n

\u7814\u7a76\u9886\u57df\uff1a\u4e92\u8865\u5b66\u4e60\u7cfb\u7edf\uff0c\u5b66\u4e60\uff0c\u795e\u7ecf\u7f51\u7edc\uff0c\u8bb0\u5fc6\u5de9\u56fa<\/span><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

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

\u6881\u680b\u680b<\/span> <\/span>| \u4f5c\u8005<\/strong><\/span><\/strong><\/span><\/span><\/strong><\/span><\/strong><\/p>\n

\u9093\u4e00\u96ea<\/span> <\/span>| <\/strong><\/span>\u7f16\u8f91<\/strong><\/p>\n


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


<\/p>\n

<\/p>\n

\n
\n
\n

\u8bba\u6587\u9898\u76ee\uff1aLearning in deep neural networks and brains with similarity-weighted interleaved learning<\/span><\/em><\/p>\n

\u8bba\u6587\u94fe\u63a5\uff1a<\/span>https:\/\/www.pnas.org\/doi\/10.1073\/pnas.2115229119<\/span><\/p>\n

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

\n
\n

 <\/span><\/p>\n<\/section>\n<\/h3>\n

\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u5728\u5b66\u4e60\u65b0\u4e1c\u897f\u65f6\uff0c\u4f1a\u8fc5\u901f\u5fd8\u8bb0\u4e4b\u524d\u5b66\u4e60\u7684\u4fe1\u606f\uff0c\u5fc5\u987b\u901a\u8fc7\u6df7\u5408\u65b0\u65e7\u6837\u672c\uff0c\u91cd\u65b0\u8bad\u7ec3\u3002\u7136\u800c\uff0c\u6df7\u5408\u6240\u6709\u65e7\u6837\u672c\u662f\u8017\u65f6\u7684\u4e5f\u662f\u4e0d\u5fc5\u8981\u7684\u3002\u90a3\u4e48\u6211\u4eec\u5e94\u8be5\u6309\u7167\u4ec0\u4e48\u6837\u7684\u7b56\u7565\u6765\u9009\u62e9\u65e7\u6837\u672c\u91cd\u65b0\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u5462\uff1f<\/span><\/section>\n

<\/section>\n
\u672c\u6587\u4f5c\u8005\u7ed9\u51fa\u4e86\u4e00\u4e2a\u76f8\u5bf9\u66f4\u597d\u7684\u7b56\u7565\uff1a\u53ea\u6df7\u5408\u4e0e\u65b0\u6837\u672c\u6709\u5f88\u5927\u76f8\u4f3c\u5ea6\u7684\u65e7\u6837\u672c\u5c31\u8db3\u591f\u4e86\u3002\u672c\u6587\u7814\u7a76\u8868\u660e\uff0c\u901a\u8fc7\u65b0\u65e7\u6837\u672c\u7684\u76f8\u4f3c\u5ea6\u52a0\u6743\u4ea4\u53c9\u8bad\u7ec3\uff0c\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u53ef\u4ee5\u5728\u4f7f\u7528\u66f4\u5c11\u7684\u6570\u636e\u7684\u540c\u65f6\uff0c\u5feb\u901f\u5b66\u4e60\u65b0\u6837\u672c\u800c\u4e0d\u5fd8\u8bb0\u65e7\u6837\u672c\u4e2d\u5b66\u4e60\u5f97\u5230\u7684\u77e5\u8bc6\u3002<\/span><\/section>\n
 <\/span><\/section>\n
<\/section>\n

\u56fe1. \u795e\u7ecf\u7f51\u7edc\u5728\u8bad\u7ec3\u65b0\u79cd\u7c7b\u7684\u56fe\u7247\u540e\u5bf9\u65e7\u79cd\u7c7b\u56fe\u7247\u7684\u53ec\u56de\u7387\uff0c\u51c6\u786e\u7387\u663e\u8457\u4e0b\u964d\uff0c\u4f46\u662f\u4e0d\u540c\u79cd\u7c7b\u7684\u65e7\u56fe\u7247\u4e0b\u964d\u7684\u7a0b\u5ea6\u4e0d\u540c\u3002\u800c\u5c06\u65b0\u65e7\u6df7\u5408\u8bad\u7ec3\uff0c\u5219\u53ec\u56de\u7387\uff0c\u51c6\u786e\u7387\u80fd\u591f\u4fdd\u6301\u7a33\u5b9a\u3002<\/span><\/p>\n

 <\/span><\/section>\n
\u6b63\u5982\u56fe1\u5c55\u793a\uff1a\u5148\u75288\u7c7b\u56fe\u7247\u9884\u8bad\u7ec3\u53ea\u6709\u4e00\u5c42\u9690\u85cf\u5c42\u7684\u7ebf\u6027\u795e\u7ecf\u7f51\u7edc\uff0c\u4e4b\u540e\u9884\u8bad\u7ec3\u7f51\u7edc\u5728\u4e24\u79cd\u60c5\u51b5\u4e0b\u5b66\u4e60\u65b0\u7c7b<\/span>\uff08Boots\uff0c\u9774\u5b50\uff09<\/span>\u7684\u8868\u73b0\u3002FoL<\/span>\uff08focused learning\uff09<\/span>\uff1a\u53ea\u63d0\u4f9b\u65b0\u7c7b\u7684\u56fe\u7247\uff1bFIL<\/span>\uff08focused all classes<\/span>\uff09<\/span>\uff1a\u63d0\u4f9b\u6240\u6709\u7c7b\u522b\u7684\u56fe\u7247\u3002\u7b2c\u4e00\u4e8c\u5217\u56fe\u7247\u8868\u793a\uff0c\u5728\u4e24\u79cd\u8bad\u7ec3\u7b56\u7565\u4e0b\uff0c\u5206\u522b\u5728\u65b0\u7c7b\u548c\u8bad\u7ec3\u8fc7\u7684\u65e7\u7c7b\u4e0b\uff0c\u53ec\u56de\u7387<\/span>\uff08Recall\uff09<\/span>\u968f\u7740\u8bad\u7ec3\u8fed\u4ee3<\/span>\uff08epoch\uff09<\/span>\u6b21\u6570\u7684\u53d8\u5316\u60c5\u51b5\uff1b\u7b2c\u4e09\u5217\u56fe\u7247\u8868\u793a\uff0c\u5728\u4e24\u79cd\u8bad\u7ec3\u7b56\u7565\u4e0b\uff0c\u6240\u6709\u7c7b\u7684\u51c6\u786e\u7387<\/span>\uff08accuracy\uff09<\/span>\u968f\u7740\u8bad\u7ec3\u8fed\u4ee3\u6b21\u6570\u7684\u53d8\u5316\uff1b\u7b2c\u56db\u5217\u8868\u793a\uff0c\u4e24\u79cd\u8bad\u7ec3\u7b56\u7565\u4e0b\uff0c\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\u968f\u7740\u8bad\u7ec3\u8fed\u4ee3\u6b21\u6570\u7684\u53d8\u5316\u3002FoL\u4f1a\u4f7f\u5f97\u5b66\u8fc7\u7684\u65e7\u7c7b\u56fe\u7247\u7684\u4fe1\u606f\u88ab\u9057\u5fd8\uff0c\u53ec\u56de\u7387\u3001\u51c6\u786e\u7387\u6307\u6807\u663e\u8457\u4e0b\u964d\uff0c\u800c\u6df7\u5408\u6240\u6709\u65e7\u7c7b\u56fe\u7247\u548c\u65b0\u7c7b\u56fe\u7247\uff0c\u8bad\u7ec3\u6548\u679c\u5219\u4fdd\u6301\u7a33\u5b9a\u3002<\/span><\/section>\n
 <\/span><\/section>\n
<\/section>\n
\u56fe2. 6\u79cd\u8bad\u7ec3\u7b56\u7565\u4e0b\uff0c\u795e\u7ecf\u7f51\u7edc\u7684\u8bad\u7ec3\u8868\u73b0\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u56fe2\u8868\u660e\uff0c\u57288\u7c7b\u56fe\u7247\u4e0a\u9884\u8bad\u7ec3\u597d\u7684\u4e00\u5c42\u9690\u85cf\u5c42\u7684\u7ebf\u6027\u795e\u7ecf\u7f51\u7edc\u5b66\u4e60\u65b0\u7c7b<\/span>\uff08Boots \u56fe\u7247\uff09<\/span>\uff0c\u57286\u79cd\u4e0d\u540c\u7684\u8bad\u7ec3\u7b56\u7565\u4e0b\uff0c\u7f51\u7edc\u8868\u73b0\u968f\u8fed\u4ee3\u6b21\u6570\u7684\u53d8\u5316\u3002A\u56fe\u8868\u793a\u516d\u79cd\u4e0d\u540c\u662f\u8bad\u7ec3\u7b56\u7565\uff1aFoL\uff1bFIL\uff1bPIL <\/span>(partial interleaved learning) <\/span>\uff1a9\u7c7b\u56fe\u7247\u7684\u5b50\u96c6\u53c2\u4e0e\u6bcf\u6b21\u8fed\u4ee3\u8bad\u7ec3<\/span>\uff08n=350 images per epoch\uff0c39 images\/class\uff09<\/span>\uff1bSWIL<\/span>\uff08\u76f8\u4f3c\u5ea6\u52a0\u6743\u4ea4\u53c9\u5b66\u4e60\uff09<\/span>\uff1a\u6839\u636e\u8bad\u7ec3\u8fc7\u76848\u7c7b\u56fe\u7247\u548c\u65b0\u7c7b<\/span>\uff08Boots\uff09<\/span>\u7684\u76f8\u4f3c\u5ea6\uff0c\u51b3\u5b9a\u6bcf\u7c7b\u5728\u8bad\u7ec3\u65b0\u7c7b\u8fc7\u7a0b\u4e2d\u7684\u56fe\u7247\u6570\u91cf\uff0c\u8d8a\u76f8\u4f3c\u5219\u6570\u91cf\u8d8a\u591a\uff1bEqWIL<\/span>\uff08Equally Weighted Interleaved Learning\uff09<\/span>\uff1a\u65b0\u7c7b\u7684\u56fe\u7247\u6570\u91cf\u548cSWIL\u4fdd\u6301\u4e00\u81f4\uff0c\u5176\u5b838\u7c7b\u7684\u56fe\u7247\u548c\u65b0\u7c7b\u56fe\u7247\u76f8\u4f3c\u5ea6\u6743\u91cd\u8bbe\u7f6e\u4e3a\u4e00\u81f4\uff0c\u5219\u8bad\u7ec3\u56fe\u7247\u4e3a75 images\/class\u3002B\u56fe\u8868\u793a\uff0c6\u4e2a\u8bad\u7ec3\u7b56\u7565\u4e0b\uff0c\u53ec\u56de\u7387\u5728\u65b0\u7c7b\u548c\u76f8\u4f3c\u5ea6\u9ad8\u7684\u65e7\u7c7b<\/span>\uff08sneaker\uff0cSandals\uff09<\/span>\u4ee5\u53ca\u76f8\u4f3c\u5ea6\u4e0d\u9ad8\u7684\u5176\u5b836\u7c7b\u4e0a\u968f\u7740\u8fed\u4ee3\u6b21\u6570\u7684\u53d8\u5316\uff1b\u6574\u4f53\u51c6\u786e\u7387\uff0c\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\u8bad\u7ec3\u8fc7\u7a0b\u7684\u53d8\u5316\u3002\u53ef\u89c1\uff0c\u6587\u7ae0\u4f5c\u8005\u63d0\u51fa\u7684SWIL\u8bad\u7ec3\u7b56\u7565\u80fd\u591f\u5f88\u597d\u7684\u7f13\u548c\u795e\u7ecf\u7f51\u7edc\u9057\u5fd8\u65e7\u4fe1\u606f\u7684\u95ee\u9898\uff0c\u540c\u65f6\u63d0\u9ad8\u8bad\u7ec3\u7684\u65f6\u95f4\u6548\u7387\u3002<\/span><\/section>\n
 <\/span><\/section>\n
<\/section>\n
\u56fe3. \u5404\u7c7b\u56fe\u7247\u7684\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u7ed3\u679c\u4ee5\u53ca\u805a\u7c7b\u7ed3\u679c<\/span><\/section>\n

<\/span><\/section>\n
\u90a3\u4e48\uff0c\u5404\u4e2a\u79cd\u7c7b\u7684\u56fe\u7247\u7684\u76f8\u4f3c\u5ea6\u6307\u6807\u7684\u8ba1\u7b97\u4e5f\u662f\u6211\u4eec\u9700\u8981\u5173\u6ce8\u7684\u91cd\u70b9\u4e4b\u4e00\uff0c\u6982\u62ec\u6765\u8bf4\uff0c\u6587\u7ae0\u4f5c\u8005\u4f7f\u7528\u516c\u5f00\u7684\u65b9\u6cd5\u5728\u7279\u5f81\u5c42\u9762\u8ba1\u7b97\u76f8\u4f3c\u5ea6\u3002\u7b80\u5355\u5730\u8bf4\uff0c\u4f5c\u8005\u8ba1\u7b97\u4e86\u76ee\u6807\u9690\u85cf\u5c42\u4e2d\u5df2\u6709\u7c7b\u6837\u672c\u548c\u65b0\u7c7b\u6837\u672c\u7684\u6fc0\u6d3b\u5411\u91cf\u5e73\u5747\u503c\u4e4b\u95f4\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u3002\u56fe3\u5c55\u793a\u4e86\u4f5c\u8005\u8ba1\u7b97\u5f97\u5230\u7684\u5404\u79cd\u7c7b\u56fe\u7247\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u77e9\u9635A\u56fe\uff1bB\u56fe\uff1a\u4f9d\u636eA\u56fe\u7684\u805a\u7c7b\u7ed3\u679c\uff1bC\u56fe\uff1a\u7ecf\u8fc7FIL\u7b56\u7565\u8bad\u7ec3\u5b66\u4e60\u201cBoots\u201d\u7c7b\u56fe\u7247\u540e\u7684\u6df7\u6dc6\u77e9\u9635 \u3002\u5bf9\u89d2\u7ebf\u503c\u88ab\u5220\u9664\uff0c\u4ee5\u63d0\u9ad8\u7f29\u653e\u7684\u6e05\u6670\u5ea6\u3002<\/span><\/section>\n
 <\/span><\/section>\n
\u6700\u8fd1\u4ee5\u5927\u8111\u4e3a\u7075\u611f\u7684\u7f13\u89e3\u707e\u96be\u6027\u5e72\u6270\u7684\u65b9\u6cd5\u53ef\u4ee5\u5206\u4e3a\uff1a1.\u57fa\u4e8e\u6b63\u5219\u5316\u7684\u65b9\u6cd5<\/span>\uff08regularization-based \uff09<\/span>2.\u57fa\u4e8e\u751f\u6210\u91cd\u653e\u7684\u65b9\u6cd5<\/span>\uff08generative-replay\u2013based methods\uff09<\/span>\u3002\u57fa\u4e8e\u6b63\u5219\u5316\u7684\u65b9\u6cd5\u6709\uff1aEWC<\/span>\uff08Elastic Weight Consolidation\uff09<\/span>\uff1bLearning without Forgetting\uff1bSynaptic Intelligence\uff1b\u8fd9\u4e9b\u901a\u5e38\u5305\u62ec\u6d4b\u91cf\u6bcf\u4e2a\u53c2\u6570\u7684\u91cd\u8981\u6027\u548c\u6dfb\u52a0\u4e00\u79cd\u6b63\u5219\u5316\u9879\uff0c\u60e9\u7f5a\u7f51\u7edc\u4e2d\u6700\u76f8\u5173\u7684\u53c2\u6570\u6216\u6620\u5c04\u51fd\u6570\u7684\u53d8\u5316\u3002\u5f53\u9700\u8981\u589e\u91cf\u5730\u5b66\u4e60\u8bb8\u591a\u65b0\u7c7b\u65f6\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u901a\u5e38\u4f1a\u53d7\u5230\u5f71\u54cd\u3002\u57fa\u4e8e\u751f\u6210\u91cd\u653e\u7684\u65b9\u6cd5\uff0c\u901a\u5e38\u5305\u542b\u6df1\u5ea6\u751f\u6210\u548c\u4efb\u52a1\u89e3\u51b3\u7f51\u7edc\u3002\u5728\u91cd\u5b66\u4e60\u7684\u8fc7\u7a0b\u4e2d\uff0c\u65b0\u7c7b\u6837\u672c\u4e0e\u751f\u6210\u7684\u4f2a\u6570\u636e\u6df7\u5408\u5728\u4e00\u8d77<\/span>\uff08\u6355\u6349\u5148\u524d\u5b66\u4e60\u7684\u4ee3\u8868\u6027\u7edf\u8ba1\u4fe1\u606f\uff09<\/span>\u3002\u751f\u6210\u91cd\u653e\u65b9\u6cd5\u4e0d\u7528\u518d\u8bbf\u95ee\u65e7\u6837\u672c\uff0c\u7136\u800c\uff0c\u95ee\u9898\u8f6c\u5411\u4e86\u6539\u8fdb\u751f\u6210\u5668<\/span>\uff08\u4e00\u4e2a\u91cd\u8981\u548c\u56f0\u96be\u7684\u95ee\u9898\uff09<\/span>\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u4f5c\u8005\u8ba1\u7b97\u4e86\u65b0\u7c7b\u7684\u76ee\u6807\u5c42\u7684\u5e73\u5747\u6fc0\u6d3b\u4e0e\u65e7\u7c7b\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u3002\u56e0\u4e3a\u4ed6\u4eec\u5173\u6ce8\u7684\u662f\u7c7b\u5c42\u6b21\u4e0a\u7684\u5b66\u4e60\u52a8\u529b\uff0c\u800c\u4e0d\u662f\u8de8\u7c7b\u7684\u5355\u4e2a\u5c5e\u6027\uff0c\u6240\u4ee5\u7814\u7a76\u5355\u4e2a\u7684\u7279\u5f81\u6620\u5c04\u7684\u5b66\u4e60\u52a8\u529b\u4e5f\u4f1a\u5f88\u6709\u8da3\u3002\u76f8\u4f3c\u5ea6\u4e5f\u53ef\u4ee5\u5728\u4e0d\u540c\u5c42\u6b21\u4e0a\u8ba1\u7b97\uff1a\u50cf\u7d20\u7ea7\u6216\u529f\u80fd\u6027\u3002<\/span><\/section>\n
 <\/span><\/section>\n


<\/p>\n

\n
\n
\n

\u8bba\u6587 Abstract<\/strong><\/span><\/p>\n<\/section>\n<\/section>\n<\/section>\n

Understanding how the brain learns throughout a lifetime remains a long-standing challenge.In artificial neural networks (ANNs), incorporating novel information too rapidly results in catastrophic interference, i.e., abrupt loss of previously acquired knowledge. Complementary Learning Systems Theory (CLST) suggests that new memories can be gradually integrated into the neocortex by interleaving new memories with existing knowledge. This approach, however, has been assumed to require interleaving all existing knowledge every time something new is learned, which is implausible because it is timeconsuming and requires a large amount of data. We show that deep, nonlinear ANNs can learn new information by interleaving only a subset of old items that share substantial representational similarity with the new information. By using such similarityweighted interleaved learning (SWIL), ANNs can learn new information rapidly with a similar accuracy level and minimal interference, while using a much smaller number of old items presented per epoch (fast and data-efficient). SWIL is shown to work with various standard classification datasets (Fashion-MNIST, CIFAR10, and CIFAR100), deep neural network architectures, and in sequential learning frameworks. We show that data efficiency and speedup in learning new items are increased roughly proportionally to the number of nonoverlapping classes stored in the network, which implies an enormous possible speedup in human brains, which encode a high number of separate categories. Finally, we propose a theoretical model of how SWIL might be implemented in the brain.<\/span><\/section>\n

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

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

\u795e\u7ecf\u52a8\u529b\u5b66\u6a21\u578b\u8bfb\u4e66\u4f1a<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n


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

\u968f\u7740\u7535\u751f\u7406\u5b66\u3001\u7f51\u7edc\u5efa\u6a21\u3001\u673a\u5668\u5b66\u4e60\u3001\u7edf\u8ba1\u7269\u7406\u3001\u7c7b\u8111\u8ba1\u7b97\u7b49\u591a\u79cd\u6280\u672f\u65b9\u6cd5\u7684\u53d1\u5c55\uff0c\u6211\u4eec\u5bf9\u5927\u8111\u795e\u7ecf\u5143\u76f8\u4e92\u4f5c\u7528\u673a\u7406\u4e0e\u8fde\u63a5\u673a\u5236\uff0c\u5bf9\u610f\u8bc6\u3001\u8bed\u8a00\u3001\u60c5\u7eea\u3001\u8bb0\u5fc6\u3001\u793e\u4ea4\u7b49\u529f\u80fd\u7684\u8ba4\u8bc6\u9010\u6e10\u6df1\u5165\uff0c\u5927\u8111\u590d\u6742\u7cfb\u7edf\u7684\u8c1c\u5e95\u6b63\u5728\u88ab\u63ed\u5f00\u3002\u4e3a\u4e86\u4fc3\u8fdb\u795e\u7ecf\u79d1\u5b66\u3001\u7cfb\u7edf\u79d1\u5b66\u3001\u8ba1\u7b97\u673a\u79d1\u5b66\u7b49\u9886\u57df\u7814\u7a76\u8005\u7684\u4ea4\u6d41\u5408\u4f5c\uff0c\u6211\u4eec\u53d1\u8d77\u4e86\u3010<\/span>\u795e\u7ecf\u52a8\u529b\u5b66\u6a21\u578b\u8bfb\u4e66\u4f1a<\/a>\u3011\u3002<\/span><\/p>\n


<\/p>\n

\u96c6\u667a\u4ff1\u4e50\u90e8\u8bfb\u4e66\u4f1a\u662f\u9762\u5411\u5e7f\u5927\u79d1\u7814\u5de5\u4f5c\u8005\u7684\u7cfb\u5217\u8bba\u6587\u7814\u8bfb\u6d3b\u52a8\uff0c\u5176\u76ee\u7684\u662f\u5171\u540c\u6df1\u5165\u5b66\u4e60\u63a2\u8ba8\u67d0\u4e2a\u79d1\u5b66\u8bae\u9898\uff0c\u6fc0\u53d1\u79d1\u7814\u7075\u611f\uff0c\u4fc3\u8fdb\u79d1\u7814\u5408\u4f5c\u3002\u3010\u795e\u7ecf\u52a8\u529b\u5b66\u6a21\u578b\u8bfb\u4e66\u4f1a<\/a>\u3011\u7531\u96c6\u667a\u4ff1\u4e50\u90e8\u548c\u5929\u6865\u8111\u79d1\u5b66\u7814\u7a76\u9662\u8054\u5408\u53d1\u8d77\uff0c\u5df2\u4e8e3\u670819\u65e5\u5f00\u59cb\uff0c\u6bcf\u5468\u516d\u4e0b\u534814:00-16:00\uff08\u6216\u6bcf\u5468\u4e94\u665a\u4e0a19:00-21:00\uff0c\u6839\u636e\u5b9e\u9645\u60c5\u51b5\u8c03\u6574\uff09\u8fdb\u884c\uff0c\u9884\u8ba1\u6301\u7eed10-12\u5468\u3002\u671f\u95f4\u5c06\u56f4\u7ed5\u795e\u7ecf\u7f51\u7edc\u591a\u5c3a\u5ea6\u5efa\u6a21\u53ca\u5176\u5728\u8111\u75be\u75c5\u3001\u8111\u8ba4\u77e5\u65b9\u9762\u7684\u5e94\u7528\u8fdb\u884c\u7814\u8ba8\u3002<\/span><\/p>\n


<\/p>\n

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


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

\u8be6\u60c5\u8bf7\u89c1\uff1a<\/span>
<\/p>\n

\u795e\u7ecf\u52a8\u529b\u5b66\u6a21\u578b\u8bfb\u4e66\u4f1a\u542f\u52a8\uff1a\u6574\u5408\u8ba1\u7b97\u795e\u7ecf\u79d1\u5b66\u7684\u591a\u5b66\u79d1\u65b9\u6cd5<\/a><\/strong><\/p>\n


<\/section>\n

<\/section>\n
\n

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