{"id":60527,"date":"2021-07-14T20:12:37","date_gmt":"2021-07-14T11:12:37","guid":{"rendered":"https:\/\/smilegate.ai\/?p=60527"},"modified":"2021-07-20T18:54:24","modified_gmt":"2021-07-20T09:54:24","slug":"multimodal-few-shot-learner","status":"publish","type":"post","link":"https:\/\/smilegate.ai\/cn\/2021\/07\/14\/multimodal-few-shot-learner\/","title":{"rendered":"\u591a\u6a21\u6001\u5c0f\u6837\u672c\u5b66\u4e60\u5668"},"content":{"rendered":"
[\u524d\u7814\u7a76\u5c0f\u7ec4\u6c89\u5b8f\u524d]<\/p>\n\n\n\n
\u968f\u7740 Open AI \u7684 GPT-3 \u548c NAVER \u7684 Hyper CLOVA \u7b49\u8d85\u5de8\u578b\u8bed\u8a00\u6a21\u578b\u7684\u53d1\u5e03\uff0c\u6700\u8fd1\u4f7f\u7528\u5b83\u4eec\u7684\u5404\u79cd\u793a\u4f8b\u548c\u670d\u52a1\u4e0d\u65ad\u6d8c\u73b0\u3002\u6240\u6709\u8fd9\u4e9b\u8d85\u7ea7\u8bed\u8a00\u6a21\u578b\u90fd\u5177\u6709\u60ca\u4eba\u7684\u80fd\u529b\uff0c\u53ef\u4ee5\u901a\u8fc7\u5c55\u793a\u4e0e\u65b0\u4efb\u52a1\uff08= \u4efb\u52a1\uff09\u76f8\u5173\u7684\u7b80\u5355\u793a\u4f8b\u800c\u65e0\u9700\u68af\u5ea6\u66f4\u65b0\u6765\u5b66\u4e60\u5982\u4f55\u81ea\u884c\u6267\u884c\u65b0\u4efb\u52a1\u3002\u7136\u800c\uff0c\u867d\u7136\u8fd9\u4e9b\u8bed\u8a00\u6a21\u578b\u5f88\u597d\u5730\u6267\u884c\u4e86\u57fa\u4e8e\u6587\u672c\u7684\u4efb\u52a1\uff0c\u4f46\u5b83\u4eec\u5e76\u6ca1\u6709\u5b8c\u5168\u5c55\u793a\u5b83\u4eec\u5bf9\u6587\u672c\u4ee5\u5916\u7684\u4efb\u52a1\uff08\u4f8b\u5982 Visual Task\uff09\u7684\u80fd\u529b\u3002 <\/p>\n\n\n\n
\u6700\u8fd1\uff0cDeepMind \u901a\u8fc7\u4e00\u7bc7\u8bba\u6587\u516c\u5f00\u4e86\u4e00\u79cd\u540d\u4e3a Frozen \u7684\u65b9\u6cd5\uff0c\u53ef\u4ee5\u4e3a\u81ea\u56de\u5f52\u8bed\u8a00\u6a21\u578b\u6267\u884c Visual Tasks\u3002\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u6709\u6548\u7684\u65b9\u6cd5\u5c06\u5c11\u6837\u672c\u5b66\u4e60\u80fd\u529b\u8f6c\u79fb\u5230\u201c\u591a\u6a21\u6001\u201d\u3002\u901a\u8fc7\u5b66\u4e60\u89c6\u89c9\u7f16\u7801\u5668\u4f7f\u7528\u56fe\u50cf-\u5b57\u5e55\u5bf9\u6570\u636e\u5c06\u6bcf\u4e2a\u56fe\u50cf\u8868\u793a\u4e3a\u4e00\u7cfb\u5217\u8fde\u7eed\u5d4c\u5165\uff0c\u9884\u8bad\u7ec3\u7684\u3001\u51bb\u7ed3\u7684\u8bed\u8a00\u6a21\u578b\u53ef\u4ee5\u4f7f\u7528\u6b64\u524d\u7f00\u4fe1\u606f\u751f\u6210\u9002\u5f53\u7684\u5b57\u5e55\u3002
<\/p>\n\n\n\n
\u56fe[1]\u5c55\u793a\u4e86\u5305\u62ec\u672c\u6587\u4ecb\u7ecd\u7684\u89c6\u89c9\u7f16\u7801\u5668\u5728\u5185\u7684\u6574\u4e2a\u7cfb\u7edf\u7ed3\u6784\u3002\u56fe\u4e2d\u4e24\u4e2aLanguage Model\u7684\u53c2\u6570\u662f\u56fa\u5b9a\u548c\u51bb\u7ed3\u7684\u3002 Vision Encoder \u4f7f\u7528\u5de6\u4fa7\u7684\u56fe\u50cf\u7f16\u7801\u4fe1\u606f\u548c\u53f3\u4fa7\u5b57\u5e55\u7684\u4e00\u4e9b\u7f16\u7801\u4fe1\u606f\u8fdb\u884c\u5b66\u4e60\uff0c\u4ee5\u4fbf\u8bed\u8a00\u6a21\u578b\uff08Self Attention Layers\uff09\u53ef\u4ee5\u751f\u6210\u5b57\u5e55\u7684\u5176\u4f59\u90e8\u5206\u3002\u8bed\u8a00\u6a21\u578b\u7684\u6743\u91cd\u4fdd\u6301\u56fa\u5b9a\uff0c\u4f46\u68af\u5ea6\u901a\u8fc7\u5b83\u4f20\u64ad\u56de\u6765\u4ee5\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u56fe\u50cf\u7f16\u7801\u5668\u3002<\/p>\n\n\n\n
\u4f7f\u7528\u4e0a\u8ff0\u7ed3\u6784\u8bad\u7ec3\u7684 Frozen \u5728\u672c\u6587\u4e2d\u79f0\u4e3a\u591a\u6a21\u6001\u5c0f\u6837\u672c\u5b66\u4e60\u5668\u3002 Frozen \u662f\u5728\u5355\u4e2a\u56fe\u50cf\u5b57\u5e55\u5bf9\u4e0a\u8bad\u7ec3\u7684\uff0c\u4f46\u662f\u4e00\u65e6\u8bad\u7ec3\uff0c\u5b83\u4e5f\u53ef\u4ee5\u6709\u6548\u5730\u54cd\u5e94\u4e0e\u591a\u4e2a\u56fe\u50cf\u548c\u6587\u672c\u5bf9\u9f50\u7684\u96c6\u5408\u3002\u6b64\u5916\uff0c\u5373\u4f7f\u5728\u672a\u7ecf\u8bad\u7ec3\u7684\u591a\u6a21\u6001\u4efb\u52a1\uff08\u4f8b\u5982\u89c6\u89c9\u95ee\u7b54 (VQA)\uff09\u4e2d\uff0c\u5b83\u4e5f\u5177\u6709\u901a\u8fc7\u5229\u7528\u9884\u8bad\u7ec3\u7684\u8bed\u8a00\u6a21\u578b\u6765\u5b66\u4e60\u65b0\u4efb\u52a1\u7684\u60ca\u4eba\u80fd\u529b\u3002\u5728\u672c\u6587\u4e2d\uff0c\u4e3a\u4e86\u63d0\u9ad8frozen\u5728\u591a\u6a21\u6001\u4efb\u52a1\u4e2d\u7684\u5c11\u955c\u5934\u5b66\u4e60\u6027\u80fd\uff0c\u5728\u6bd4\u8f83\u591a\u4e2a\u4efb\u52a1\u7684\u955c\u5934\u7684\u540c\u65f6\u8fdb\u884c\u4e86\u5b9e\u9a8c\u3002 <\/p>\n\n\n\n
\u4e0b\u56fe[3]\u663e\u793a\u4e86\u5c06\u4e0a\u8ff0\u5b9e\u9a8c\u5f97\u5230\u7684\u6700\u4f18\u8bbe\u7f6e\u5e94\u7528\u4e8eFrozen\u6a21\u578b\u540e\u7684\u6d4b\u8bd5\u6570\u636e\u7684\u6d4b\u8bd5\u7ed3\u679c\u3002\u5982\u679c\u4f60\u770b\u56fe\u7684\u7b2c\u4e00\u884c\uff0c\u5982\u679c\u4f60\u628a\u4e24\u5f20\u56fe\u7247\u548c\u4e00\u4e2a\u6807\u9898\u4f5c\u4e3a\u4f8b\u5b50\uff0c\u6700\u540e\u63d0\u4f9b\u4e00\u4e2a\u56fe\u7247\u548c\u6587\u5b57\u4fe1\u606f\u201c\u8fd9\u4e2a\u4eba\u662f\u8c01\u201d\u4f5c\u4e3a\u6697\u793a\u8bcd\uff0c\u6a21\u578b\u6700\u7ec8\u751f\u6210\u4e86\u201c\u60ca\u6050\u201d\u6050\u6016\u201d\u8868\u60c5\u7b26\u53f7\u3002\u7b2c\u4e8c\u884c\u4e5f\u7ed9\u51fa\u4e86\u4e00\u4e2a\u7c7b\u4f3c\u683c\u5f0f\u7684\u4f8b\u5b50\uff0c\u53ea\u80fd\u5f97\u5230\u201c\u98de\u673a\u201d\u7684\u4fe1\u606f\u4f5c\u4e3a\u56fe\u50cf\u4fe1\u606f\uff0c\u4f46\u662f\u8bed\u8a00\u6a21\u578b\u4e2d\u4e5f\u5305\u542b\u4e86\u4e8b\u5b9e\u77e5\u8bc6\uff0c\u6240\u4ee5\u6700\u7ec8\u7684\u6a21\u578b\u662f\u4eba\u7684\u201c\u83b1\u7279\u5144\u5f1f\u201d\u8c01\u53d1\u660e\u4e86\u98de\u673a\u6211\u4eec\u53ef\u4ee5\u4e3a\u60a8\u63d0\u4f9b\u7b54\u6848\u3002 Frozen \u7684\u5438\u5f15\u529b\u5728\u4e8e\u5b83\u53ef\u4ee5\u901a\u8fc7\u7ed3\u5408\u8bed\u8a00\u6a21\u578b\u7684\u89c6\u89c9\u4fe1\u606f\u548c\u4e8b\u5b9e\u77e5\u8bc6\u4fe1\u606f\u6765\u6267\u884c\u6b64\u7c7b\u4efb\u52a1\u3002
<\/p>\n\n\n\n
\u5f00\u53d1 Frozen \u7684\u76ee\u6807\u4e0d\u662f\u5728\u7279\u5b9a\u4efb\u52a1\u4e2d\u6700\u5927\u5316\u6027\u80fd\uff0c\u56e0\u6b64\u5b83\u5177\u6709\u6267\u884c\u4efb\u52a1\u6240\u5fc5\u9700\u7684\u80fd\u529b\uff0c\u4f46\u4e0e SOTA \u5728\u51e0\u6b21\u5b66\u4e60\u7684\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u4ecd\u7136\u5b58\u5728\u5dee\u5f02\u3002\u7136\u800c\uff0c\u5728\u6ca1\u6709\u67e5\u770b\u57fa\u51c6\u63d0\u4f9b\u7684\u6240\u6709\u5fc5\u8981\u8bad\u7ec3\u793a\u4f8b\u7684\u60c5\u51b5\u4e0b\uff0c\u5404\u79cd\u4efb\u52a1\u7684\u6027\u80fd\u8fdc\u8fdc\u8d85\u8fc7\u57fa\u7ebf\u3002\u6b64\u5916\uff0c\u5982\u56fe [3] \u6240\u793a\uff0cFrozen \u7ecf\u5e38\u4ea7\u751f\u6709\u5438\u5f15\u529b\u7684\u8f93\u51fa\uff0c\u53ef\u4ee5\u5c06\u5176\u89c6\u4e3a\u4e00\u4e2a\u7cfb\u7edf\uff0c\u7528\u4e8e\u5bf9\u56fe\u50cf\u8fdb\u884c\u771f\u6b63\u5f00\u653e\u548c\u4e0d\u53d7\u7ea6\u675f\u7684\u8bed\u8a00\u89e3\u91ca\u3002
<\/p>\n\n\n\n
\u5230\u76ee\u524d\u4e3a\u6b62\uff0c\u6211\u4eec\u5df2\u7ecf\u4ecb\u7ecd\u4e86\u5982\u4f55\u4f7f\u7528 Frozen \u5c06\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u8f6c\u6362\u4e3a\u591a\u6a21\u6001\u5c0f\u6837\u672c\u5b66\u4e60\u7cfb\u7edf\u3002\u66f4\u591a\u7ec6\u8282\u8bf7\u53c2\u8003\u8bba\u6587\u3002<\/p>\n\n\n\n
<\/p>\n\n\n\n
\u53c2\u8003<\/strong><\/p>\n\n\n\n [1] Tsimpoukelli, M., Menick, J., Cabi, S., Eslami, S.M., Vinyals, O., & Hill, F. (2021)\u3002\u4f7f\u7528\u51bb\u7ed3\u8bed\u8a00\u6a21\u578b\u7684\u591a\u6a21\u6001\u5c0f\u6837\u672c\u5b66\u4e60\u3002 arXiv \u9884\u5370\u672c arXiv:2106.13884<\/em>.<\/p>\n\n\n\n [2] https:\/\/www.youtube.com\/watch?v=FYA_jwPpXi0<\/p>\n\n\n\n <\/p>\n [\uc120\ud589\uc5f0\uad6c\ud300 \uc2ec\ud64d\ub9e4] Open AI\uc758 GPT-3, NAVER\uc758 Hyper CLOVA \uc640 \uac19\uc740 \ucd08 \uac70\ub300 \uc5b8\uc5b4 \ubaa8\ub378\uc774 \uacf5\uac1c\ub418\uba74\uc11c \ucd5c\uadfc\uc5d0\ub294 \uc774\ub97c \ud65c\uc6a9\ud55c \ub2e4\uc591\ud55c \uc0ac\ub840 \ubc0f \uc11c\ube44\uc2a4\ub4e4\ub3c4 \ub9ce\uc774 \uc3df\uc544\uc838 \ub098\uc624\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\uc640 \uac19\uc740 \ucd08 \uac70\ub300 \uc5b8\uc5b4 \ubaa8\ub378\ub4e4\uc740 \ubaa8\ub450 gradient \uc5c5\ub370\uc774\ud2b8 \uc5c6\uc774\ub3c4 \uc0c8\ub85c\uc6b4 Task(=\uc791\uc5c5)\uc640 \uad00\ub828\ub41c \uac04\ub2e8\ud55c \uc608\uc2dc\ub4e4\uc744 \uc81c\uc2dc\ud574\uc8fc\ub294 \uac83 \ub9cc\uc73c\ub85c\ub3c4 \uc0c8\ub85c\uc6b4 Task\ub97c \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc744 \uc2a4\uc2a4\ub85c \ubc30\uc6b8 \uc218 \uc788\ub2e4\ub294 \ub180\ub77c\uc6b4 \ub2a5\ub825\uc744 \uac16\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \uc774\ub7ec\ud55c \uc5b8\uc5b4…<\/p>\n