{"id":3070,"date":"2025-04-29T16:04:16","date_gmt":"2025-04-29T08:04:16","guid":{"rendered":"https:\/\/www.gnn.club\/?p=3070"},"modified":"2025-05-09T14:46:44","modified_gmt":"2025-05-09T06:46:44","slug":"coca%ef%bc%9a%e5%a4%9a%e6%a8%a1%e6%80%81%e9%a2%86%e5%9f%9f%e7%9a%84%e5%85%ad%e8%be%b9%e5%bd%a2%e6%88%98%e5%a3%ab","status":"publish","type":"post","link":"http:\/\/www.gnn.club\/?p=3070","title":{"rendered":"CoCa\uff1a\u591a\u6a21\u6001\u9886\u57df\u7684\u516d\u8fb9\u5f62\u6218\u58eb"},"content":{"rendered":"\n

1. \u7b80\u4ecb<\/h2>\n\n\n\n

\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u57fa\u7840\u6a21\u578b\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5177\u6709\u91cd\u8981\u610f\u4e49\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u6a21\u578b\u80fd\u591f\u5feb\u901f\u8fc1\u79fb\u81f3\u591a\u79cd\u4e0b\u6e38\u4efb\u52a1\u3002\u672c\u7814\u7a76\u63d0\u51fa\u5bf9\u6bd4-\u751f\u6210\u53cc\u6a21\u6001\u6a21\u578b\uff08Contrastive Captioner, CoCa\uff09\uff0c\u8be5\u6a21\u578b\u91c7\u7528\u6781\u7b80\u8bbe\u8ba1\uff0c\u901a\u8fc7\u8054\u5408\u4f18\u5316\u5bf9\u6bd4\u635f\u5931\u4e0e\u751f\u6210\u635f\u5931\uff0c\u5b9e\u73b0\u56fe\u50cf-\u6587\u672c\u7f16\u7801\u5668-\u89e3\u7801\u5668\u57fa\u7840\u6a21\u578b\u7684\u9884\u8bad\u7ec3\uff0c\u4ece\u800c\u517c\u5177CLIP\u7b49\u5bf9\u6bd4\u5f0f\u6a21\u578b\u4e0eSimVLM\u7b49\u751f\u6210\u5f0f\u6a21\u578b\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n

\u4e0e\u4f20\u7edf\u7f16\u7801\u5668-\u89e3\u7801\u5668\u67b6\u6784\uff08\u6240\u6709\u89e3\u7801\u5c42\u5747\u5173\u6ce8\u7f16\u7801\u5668\u8f93\u51fa\uff09\u4e0d\u540c\uff0cCoCa\u5728\u524d\u534a\u90e8\u5206\u89e3\u7801\u5c42\u4e2d\u7701\u7565\u8de8\u6a21\u6001\u6ce8\u610f\u529b\u673a\u5236\u4ee5\u751f\u6210\u5355\u6a21\u6001\u6587\u672c\u8868\u5f81\uff0c\u540e\u534a\u90e8\u5206\u89e3\u7801\u5c42\u5219\u4ee5\u7ea7\u8054\u65b9\u5f0f\u5bf9\u56fe\u50cf\u7f16\u7801\u5668\u8f93\u51fa\u8fdb\u884c\u8de8\u6a21\u6001\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u6700\u7ec8\u83b7\u5f97\u591a\u6a21\u6001\u56fe\u6587\u8868\u5f81\u3002\u8be5\u6a21\u578b\u5728\u5355\u6a21\u6001\u56fe\u50cf-\u6587\u672c\u5d4c\u5165\u4e4b\u95f4\u65bd\u52a0\u5bf9\u6bd4\u635f\u5931\uff0c\u540c\u65f6\u5728\u591a\u6a21\u6001\u89e3\u7801\u5668\u8f93\u51fa\u7aef\u91c7\u7528\u81ea\u56de\u5f52\u6587\u672c\u9884\u6d4b\u7684\u751f\u6210\u635f\u5931\u3002\u7531\u4e8e\u5171\u4eab\u8ba1\u7b97\u56fe\uff0c\u4e24\u4e2a\u8bad\u7ec3\u76ee\u6807\u80fd\u4ee5\u6781\u4f4e\u5f00\u9500\u9ad8\u6548\u5b9e\u73b0\u3002<\/p>\n\n\n\n

CoCa\u7684\u6a21\u578b\u8bbe\u8ba1\u5176\u5b9e\u662fALBEF\u7684\u5ef6\u5c55\u6027\u5de5\u4f5c\uff0c\u53ea\u662f\u6a21\u578b\u66f4\u5927\uff0c\u6570\u636e\u96c6\u8d85\u5927\uff0c\u628a\u6027\u80fd\u5806\u7684\u8d85\u7ea7\u597d\u3002\u5b9e\u9a8c\u8868\u660e\uff0cCoCa\u5728\u96f6\u6837\u672c\u8fc1\u79fb\u6216\u6781\u5c11\u91cf\u4efb\u52a1\u9002\u914d\u6761\u4ef6\u4e0b\uff0c\u4e8e\u89c6\u89c9\u8bc6\u522b\uff08ImageNet\u3001Kinetics400\/600\/700\u3001Moments-in-Time\uff09\u3001\u8de8\u6a21\u6001\u68c0\u7d22\uff08MSCOCO\u3001Flickr30K\u3001MSR-VTT\uff09\u3001\u591a\u6a21\u6001\u7406\u89e3\uff08VQA\u3001SNLI-VE\u3001NLVR2\uff09\u53ca\u56fe\u50cf\u63cf\u8ff0\u751f\u6210\uff08MSCOCO\u3001NoCaps\uff09\u7b49\u5e7f\u6cdb\u4e0b\u6e38\u4efb\u52a1\u4e2d\u8fbe\u5230\u6700\u5148\u8fdb\u6027\u80fd\u3002\u7279\u522b\u5728ImageNet\u5206\u7c7b\u4efb\u52a1\u4e2d\uff0cCoCa\u53d6\u5f9786.3%\u7684\u96f6\u6837\u672cTop-1\u51c6\u786e\u7387\uff0c\u51bb\u7ed3\u7f16\u7801\u5668\u5fae\u8c03\u5206\u7c7b\u5934\u65f6\u8fbe90.6%\uff0c\u5b8c\u6574\u5fae\u8c03\u540e\u66f4\u4ee591.0%\u7684Top-1\u51c6\u786e\u7387\u5237\u65b0\u5f53\u524d\u6700\u4f73\u8bb0\u5f55\u3002<\/p>\n\n\n\n

\u8bba\u6587\u4e2d\u7ed9\u5230\u4e86\u975e\u5e38\u76f4\u89c2\u7684\u4e00\u5f20\u56fe\u3002\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u5982\u56fe4\u6240\u793a\uff0c\u5728\u8fd9\u4e2a\u8718\u86db\u7f51\u56fe\u4e2d\uff0c\u6bcf\u4e2a\u8fb9\u8868\u5f81\u4e00\u4e2a\u6570\u636e\u96c6\uff08\u4e5f\u53ef\u4ee5\u770b\u4f5c\u4e00\u4e2a\u4efb\u52a1\uff09\u3002\u9ec4\u8272\u9634\u5f71\u662fCoCa\u88ab\u63d0\u51fa\u4e4b\u524d\uff0c\u6bcf\u4e2a\u6570\u636e\u96c6\u4e0a\u4e0d\u540c\u6a21\u578b\u80fd\u8fbe\u5230\u7684\u6700\u597d\u7684\u6027\u80fd\uff08SOTA\uff09\uff0c\u7d2b\u8272\u5219\u662fCoCa\u80fd\u5230\u8fbe\u7684\u6027\u80fd\uff0c\u7b80\u5355\u8bf4\u5c31\u662f\u540a\u6253\u56fe\u4e2d\u6240\u6709\u4efb\u52a1\u7684\u6240\u6709SOTA\u6a21\u578b\u3002<\/mark><\/p>\n\n\n\n

\u4e0d\u8fc7\uff0c\u6084\u6084\u7684\u8bf4\uff0c\u6211\u4e2a\u4eba\u611f\u89c9\u8fd9\u770b\u4f3c\u6050\u6016\u4e14\u9707\u64bc\u7684\u6027\u80fd\u5b9e\u9645\u4e0a\u6ca1\u6709\u592a\u591a\u8d21\u732e\uff0c\u4e3b\u8981\u662f\u8d85\u5927\u8bad\u7ec3\u96c6\u652f\u6301\u7684...<\/p>\n\n\n\n

\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2205.01917<\/a><\/p>\n\n\n\n

2. \u6a21\u578b\u4e0e\u635f\u5931<\/h2>\n\n\n\n

\u8fd9\u5f20\u56fe\u5c55\u793a\u4e86Contrastive Captioners (CoCa)\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u6846\u67b6\u53ca\u5176\u5e94\u7528\u573a\u666f\uff0c\u4ee5\u53ca\u8bad\u7ec3\u6240\u7528\u7684\u635f\u5931\u51fd\u6570\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

CoCA\u7684\u6a21\u578b\u67b6\u6784\u4e0eALBEF\u4e0d\u80fd\u8bf4\u4e00\u6a21\u4e00\u6837\u5427\uff0c\u53ef\u4ee5\u8bf4\u6beb\u65e0\u5dee\u522b\uff0c\u4f46\u5728\u635f\u5931\u51fd\u6570\u7684\u9009\u62e9\u4e0a\u662f\u6709\u5dee\u5f02\u7684\u3002\u5177\u4f53\u6765\u8bf4CoCA\u7684\u635f\u5931\u51fd\u6570\u7531\u4e24\u90e8\u5206\u7ec4\u6210\uff0c\u901a\u8fc7\u52a0\u6743\u8054\u5408\u4f18\u5316\uff1a<\/p>\n\n\n\n

1\uff0e\u5bf9\u6bd4\u635f\u5931\uff08Contrastive Loss\uff09<\/strong><\/p>\n\n\n\n

    \n
  • \u76ee\u6807\uff1a\u62c9\u8fd1\u5339\u914d\u7684\u56fe\u50cf\uff0d\u6587\u672c\u5bf9\u7684\u7279\u5f81\u8ddd\u79bb\uff0c\u63a8\u5f00\u4e0d\u5339\u914d\u5bf9\u3002<\/li>\n\n\n\n
  • \u5b9e\u73b0\u65b9\u5f0f\uff1a<\/li>\n\n\n\n
  • \u8ba1\u7b97\u56fe\u50cf\u7279\u5f81 $v_i$ \u548c\u6587\u672c\u7279\u5f81 $t_j$ \u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u77e9\u9635\u3002<\/li>\n\n\n\n
  • \u91c7\u7528\u5bf9\u79f0\u7684InfoNCE\u635f\u5931\uff08\u540cCLIP\uff09\uff1a<\/li>\n<\/ul>\n\n\n\n

    $$
    \\mathcal{L}{\\text {cont }}=-\\frac{1}{2 N}\\left(\\sum{i=1}^N \\log \\frac{e^{s\\left(v_i, t_i\\right) \/ \\tau}}{\\sum_{j=1}^N e^{s\\left(v_i, t_j\\right) \/ \\tau}}+\\sum_{j=1}^N \\log \\frac{e^{s\\left(v_j, t_j\\right) \/ \\tau}}{\\sum_{i=1}^N e^{s\\left(v_i, t_j\\right) \/ \\tau}}\\right)
    $$<\/p>\n\n\n\n

    \uff0d\u5176\u4e2d $s(\\cdot)$ \u4e3a\u76f8\u4f3c\u5ea6\u51fd\u6570\uff0c$\\tau$ \u4e3a\u6e29\u5ea6\u7cfb\u6570\uff0c$N$ \u4e3abatch size\u3002<\/p>\n\n\n\n

    2\uff0e\u751f\u6210\u635f\u5931\uff08Captioning Loss\uff09<\/strong><\/p>\n\n\n\n

    CoCA\u7684\u751f\u6210\u635f\u5931\u91c7\u7528\u6807\u51c6\u7684\u81ea\u56de\u5f52\u4ea4\u53c9\u6458\u635f\u5931\uff0c\u5177\u4f53\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n

    $$
    \\mathcal{L}{\\text {cap }}=-\\sum{t=1}^T \\sum_{w \\in \\mathcal{V}} y_{t, w} \\log P\\left(w \\mid w_{1: t-1}, v\\right)
    $$<\/p>\n\n\n\n

      \n
    • $T$ \uff1a\u751f\u6210\u6587\u672c\u7684\u6700\u5927\u957f\u5ea6\u3002<\/li>\n\n\n\n
    • $\\mathcal{V}$ \uff1a\u8bcd\u8868\u96c6\u5408\u3002<\/li>\n\n\n\n
    • $y_{t, w}$ \uff1a\u7b2c $t$ \u65f6\u523b\u8bcd $w$ \u7684one\uff0dhot\u6807\u7b7e\u3002<\/li>\n\n\n\n
    • $P\\left(w \\mid w_{1: t-1}, v\\right)$ \uff1a\u57fa\u4e8e\u56fe\u50cf\u7279\u5f81 $v$ \u548c\u5386\u53f2\u8bcd $w_{1: t-1}$ \u7684\u9884\u6d4b\u6982\u7387\u3002<\/li>\n<\/ul>\n\n\n\n

      CoCA\u907f\u514d\u4f7f\u7528ITM\uff08Image-Text Matching\uff09\u548cMLM\uff08Masked Language Modeling\uff09\u7684\u4e00\u4e2a\u91cd\u8981\u539f\u56e0\u5728\u4e8e\u591a\u6b21\u524d\u5411\u8ba1\u7b97\u5e26\u6765\u7684\u6548\u7387\u74f6\u9888<\/strong>\u3002ITM\u9700\u8981\u989d\u5916\u6784\u9020\u8d1f\u6837\u672c\u5bf9\u5e76\u8fdb\u884c\u4e8c\u5206\u7c7b\u5224\u65ad\uff0c\u6bcf\u5bf9\u56fe\u6587\u6570\u636e\u9700\u72ec\u7acb\u524d\u5411\u8ba1\u7b97\u76f8\u4f3c\u5ea6\uff0c\u663e\u8457\u589e\u52a0\u8ba1\u7b97\u5f00\u9500\uff1b\u800cMLM\u9700\u8981\u5bf9\u6587\u672c\u968f\u673a\u63a9\u7801\u5e76\u9010\u4f4d\u7f6e\u9884\u6d4b\uff0c\u9700\u5bf9\u540c\u4e00\u6587\u672c\u8fdb\u884c\u591a\u6b21\u90e8\u5206\u524d\u5411\u4f20\u64ad\uff08\u63a9\u7801\u4f4d\u7f6e\u53d8\u5316\u65f6\uff09\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cCoCA\u7684\u5bf9\u6bd4\u635f\u5931\uff08InfoNCE\uff09\u4ec5\u9700\u4e00\u6b21\u524d\u5411\u8ba1\u7b97\u5373\u53ef\u83b7\u53d6\u6574\u4e2abatch\u7684\u56fe\u6587\u76f8\u4f3c\u5ea6\u77e9\u9635\uff0c\u751f\u6210\u635f\u5931\uff08Captioning\uff09\u867d\u9700\u81ea\u56de\u5f52\u89e3\u7801\uff0c\u4f46\u4ec5\u9700\u5355\u5411\u5e8f\u5217\u5efa\u6a21\uff0c\u907f\u514d\u4e86MLM\u7684\u968f\u673a\u63a9\u7801\u5e26\u6765\u7684\u8ba1\u7b97\u788e\u7247\u5316\u3002\u8fd9\u79cd\u8bbe\u8ba1\u5c06\u591a\u4efb\u52a1\u8ba1\u7b97\u590d\u6742\u5ea6\u4eceO(N\u00b2 + T\u00b7L)<\/strong>\uff08ITM\u7684\u8d1f\u6837\u672c\u5bf9+MLM\u7684\u63a9\u7801\u4f4d\u7f6e\uff09\u964d\u81f3O(N + T)<\/strong>\uff08N\u4e3abatch size\uff0cT\u4e3a\u751f\u6210\u5e8f\u5217\u957f\u5ea6\uff09\uff0c\u66f4\u9002\u5408\u7aef\u5230\u7aef\u7684\u5927\u89c4\u6a21\u8bad\u7ec3\u3002<\/p>\n\n\n\n

      \u6b63\u662f\u7531\u4e8e\u751f\u6210\u635f\u5931\uff08Captioning\uff09\u9700\u8981\u81ea\u56de\u5f52\u5efa\u6a21\uff0cCoCA\u5728\u6587\u672c\u7aef\u7684\u7279\u5f81\u63d0\u53d6\u91c7\u7528\u4e86Transformer\u7684Decoder\u7ed3\u6784\u800c\u975eEncoder\u3002<\/strong> \u8fd9\u4e00\u8bbe\u8ba1\u7684\u5173\u952e\u533a\u522b\u5728\u4e8eDecoder\u4f7f\u7528\u4e86Masked Multi-Head Attention\uff08MMHA\uff09<\/strong>\uff0c\u800c\u975eEncoder\u7684\u6807\u51c6Multi-Head Attention\uff08MHA\uff09\u3002MMHA\u901a\u8fc7\u56e0\u679c\u63a9\u7801\uff08Causal Mask\uff09\u5f3a\u5236\u6bcf\u4e2a\u4f4d\u7f6e\u4ec5\u80fd\u5173\u6ce8\u5f53\u524d\u4f4d\u7f6e\u53ca\u4e4b\u524d\u7684\u5e8f\u5217\u4fe1\u606f\uff0c\u4ece\u800c\u4e25\u683c\u9075\u5faa\u81ea\u56de\u5f52\u751f\u6210\u7684\u65f6\u5e8f\u4f9d\u8d56\uff08\u5373\"\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u4ec5\u4f9d\u8d56\u5386\u53f2\u8bcd\"\uff09\u3002\u82e5\u91c7\u7528Encoder\u7684MHA\uff08\u65e0\u63a9\u7801\u5168\u5c40\u6ce8\u610f\u529b\uff09\uff0c\u751f\u6210\u4efb\u52a1\u4f1a\u56e0\u672a\u6765\u4fe1\u606f\u6cc4\u9732\u5bfc\u81f4\u8bad\u7ec3\u4e0e\u63a8\u7406\u4e0d\u4e00\u81f4\uff1b\u800c\u5bf9\u6bd4\u5b66\u4e60\u5206\u652f\u5728\u590d\u7528\u540c\u4e00Decoder\u65f6\uff0c\u53ef\u4e34\u65f6\u5173\u95ed\u56e0\u679c\u63a9\u7801\uff0c\u8f6c\u4e3a\u5168\u5c40\u6ce8\u610f\u529b\u4ee5\u63d0\u53d6\u5b8c\u6574\u6587\u672c\u8868\u5f81\u3002\u8fd9\u79cd\u7075\u6d3b\u5207\u6362\u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u65e2\u6ee1\u8db3\u4e86\u751f\u6210\u4efb\u52a1\u7684\u81ea\u56de\u5f52\u7ea6\u675f\uff0c\u53c8\u907f\u514d\u4e86\u4e3a\u5bf9\u6bd4\u5b66\u4e60\u5355\u72ec\u8bbe\u8ba1\u6587\u672c\u7f16\u7801\u5668\u7684\u5197\u4f59\uff0c\u5b9e\u73b0\u4e86\u53c2\u6570\u6548\u7387\u4e0e\u4efb\u52a1\u6027\u80fd\u7684\u5e73\u8861\u3002<\/p>\n\n\n\n

      \u5728\u6587\u672c\u7aef\u4f7f\u7528Decoder\u800c\u975eEncoder\u53ef\u4ee5\u8bf4\u662fCoCa\u548cALBEF\u5728\u6a21\u578b\u8bbe\u8ba1\u4e0a\u7684\u552f\u4e00\u533a\u522b\u4e86\u3002<\/p>\n\n\n\n

      \u5f53\u6a21\u578b\u9884\u8bad\u7ec3\u597d\u4ee5\u540e\uff0c\u53ef\u4ee5\u9009\u7528\u6a21\u578b\u7684\u4e0d\u540c\u90e8\u5206\u8fdb\u884c\u4e0d\u540c\u7684\u4efb\u52a1\u5904\u7406\u3002\u4f8b\u5982\uff0c\u9009\u7528single-encoder models\u53ef\u4ee5\u8fdb\u884c\u56fe\u50cf\u5904\u7406\uff1b\u9009\u7528dual-encoder models\u53ef\u4ee5\u8fdb\u884c\u56fe\u6587\u68c0\u7d22\u4efb\u52a1\uff1b\u9009\u7528encoder-decoder models\u53ef\u4ee5\u8fdb\u884c\u751f\u6210\u7c7b\u4efb\u52a1\u3002\u5982\u4e0b\u56fe\uff1a<\/p>\n\n\n\n

      \"\"<\/figure>\n\n\n\n

      \u5b9e\u9a8c<\/h2>\n\n\n\n

      CoCa\u603b\u7684\u6765\u8bf4\u662f\u4e00\u4e2a\u5356\u6027\u80fd\u7684\u5de5\u4f5c\uff0c\u5176\u6548\u679c\u5728\u7b80\u4ecb\u4e2d\u7684\u8718\u86db\u7f51\u56fe4\u4e2d\u5df2\u7ecf\u4ecb\u7ecd\u8fc7\u4e86\u3002\u9664\u6b64\u4e4b\u5916\uff0c\u9884\u8bad\u7ec3\u540e\u7684CoCa\u5728\u5355\u6a21\u6001\u4e0a\u7684\u6027\u80fd\u4e5f\u975e\u5e38\u5938\u5f20\uff0c\u4f8b\u5982\u5728ImageNet\u4e0a\u53ef\u4ee5\u8fbe\u5230\u6050\u6016\u768491.0\u7684\u6027\u80fd\u3002<\/p>\n\n\n

      \n
      \"\"<\/figure><\/div>\n\n\n

      \u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cCoCa\u8fd8\u53ef\u4ee5\u7528\u4e8e\u89c6\u9891\u52a8\u4f5c\u8bc6\u522b\u3002\u9996\u5148\uff0c\u83b7\u53d6\u89c6\u9891\u7684\u591a\u4e2a\u5e27\uff0c\u5e76\u5c06\u6bcf\u4e2a\u5e27\u5206\u522b\u8f93\u5165\u5230\u5171\u4eab\u56fe\u50cf\u7f16\u7801\u5668\u4e2d\uff0c\u5982\u56fe 3 \u6240\u793a\u3002\u5bf9\u4e8e\u51bb\u7ed3\u7279\u5f81\u8bc4\u4f30\u6216\u5fae\u8c03\uff0c\u6211\u4eec\u5728\u7a7a\u95f4\u548c\u65f6\u95f4\u7279\u5f81token\u7684\u57fa\u7840\u4e0a\u5b66\u4e60\u4e00\u4e2a\u989d\u5916\u7684attention pooling\uff0c\u5e76\u4f7f\u7528softmax \u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8be5\u6c60\u5316\u5668\u53ea\u6709\u4e00\u4e2a\u67e5\u8be2token\uff0c\u56e0\u6b64\u5bf9\u6240\u6709\u7a7a\u95f4\u548c\u65f6\u95f4\u6807\u8bb0\u8fdb\u884c\u6c60\u5316\u8ba1\u7b97\u7684\u5f00\u9500\u5e76\u4e0d\u5927\u3002\u5bf9\u4e8e\u96f6\u6837\u672c\u89c6\u9891\u6587\u672c\u68c0\u7d22\uff0c\u6211\u4eec\u4f7f\u7528\u4e00\u79cd\u66f4\u7b80\u5355\u7684\u65b9\u6cd5\uff0c\u5373\u8ba1\u7b97\u89c6\u9891 16 \u5e27\uff08\u5e27\u662f\u4ece\u89c6\u9891\u4e2d\u5747\u5300\u91c7\u6837\u7684\uff09\u7684\u5e73\u5747\u5d4c\u5165\u3002\u5728\u8ba1\u7b97\u68c0\u7d22\u6307\u6807\u65f6\uff0c\u6211\u4eec\u8fd8\u5c06\u6bcf\u4e2a\u89c6\u9891\u7684\u5b57\u5e55\u7f16\u7801\u4e3a\u76ee\u6807\u5d4c\u5165\u3002<\/p>\n\n\n

      \n
      \"\"<\/figure><\/div>\n\n
      \n
      \"\"<\/figure><\/div>\n\n\n

      \u89c6\u9891\u4efb\u52a1\u7684\u6548\u679c\u5982\u4e0a\u8868\u6240\u793a\uff0c\u540c\u6837\u662f\u8fbe\u5230\u4e86SOTA\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

      \u9664\u6b64\u4e4b\u5916\uff0cCoCa\u7684\u4f5c\u8005\u56e2\u961f\u8fd8\u5c1d\u8bd5\u4e86\u5176\u5728\u5176\u4ed6\u5e38\u89c1\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\uff0c\u5305\u62eczero-shot\u7b49\u80fd\u529b\uff0c\u6d88\u878d\u5b9e\u9a8c\u7b49\u3002\u539f\u6587\u4e2d\u6709\u5927\u91cf\u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u8be6\u89c1\u539f\u8bba\u6587\u3002.<\/p>\n\n\n\n

      \u7ed3\u8bba<\/h2>\n\n\n\n

      \u603b\u7684\u6765\u8bf4\uff0cCoCa\u7684\u5356\u70b9\u9664\u4e86\u5f3a\u5927\u7684\u6027\u80fd\u4ee5\u5916\uff0c\u4e5f\u8ba9\u5927\u5bb6\u5173\u6ce8\u5230\u635f\u5931\u51fd\u6570\u9009\u62e9\u4e0e\u8ba1\u7b97\u6548\u7387\u7684\u95ee\u9898\u3002\u5176\u820d\u5f03ITM\uff08\u9700\u8d1f\u6837\u672c\u5bf9\u91cd\u590d\u8ba1\u7b97\uff09\u548cMLM\uff08\u9700\u968f\u673a\u63a9\u7801\u591a\u6b65\u9884\u6d4b\uff09\u7b49\u4f20\u7edf\u591a\u6a21\u6001\u635f\u5931\uff0c\u4ec5\u4fdd\u7559\u5bf9\u6bd4\u635f\u5931\uff08\u5355\u6b21\u524d\u5411\u8ba1\u7b97\u76f8\u4f3c\u5ea6\u77e9\u9635\uff09<\/strong>\u548c\u751f\u6210\u635f\u5931\uff08\u81ea\u56de\u5f52\u89e3\u7801\u4f46\u65e0\u9700\u63a9\u7801\u6270\u52a8\uff09<\/strong>\uff0c\u53ef\u4ee5\u907f\u514d\u6a21\u578b\u5728\u8ba1\u7b97\u635f\u5931\u65f6\u9700\u8981\u591a\u6b21\u524d\u5411\u8ba1\u7b97\u3002\u56e0\u6b64\uff0cCoCa\u53ef\u4ee5\u5728\u8d85\u5927\u7684\u8bad\u7ec3\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u6b64\u5916\uff0c\u5176\u81ea\u8eab\u7ed3\u6784\u4e5f\u975e\u5e38\u7b80\u5355\u660e\u4e86\uff0c\u9002\u5408\u505a\u7f29\u653e\uff0c\u56e0\u6b64CoCa\u56e2\u961f\u4e5f\u5728\u8bba\u6587\u4e2d\u5c06\u5176\u505a\u5230\u4e86\u67091.2B\u53c2\u6570\u91cf\u7684\u5927\u6a21\u578b\uff0c\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u5927\u6570\u636e\uff0c\u5927\u6a21\u578b\u7684\u91cd\u8981\u6027\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

      1. \u7b80\u4ecb \u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u57fa\u7840\u6a21\u578b\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5177\u6709\u91cd\u8981\u610f\u4e49\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u6a21\u578b\u80fd\u591f\u5feb\u901f\u8fc1\u79fb\u81f3\u591a\u79cd\u4e0b\u6e38\u4efb\u52a1\u3002\u672c\u7814\u7a76\u63d0 […]<\/p>\n","protected":false},"author":1,"featured_media":3071,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[29,18],"tags":[],"class_list":["post-3070","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-29","category-18"],"_links":{"self":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3070","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3070"}],"version-history":[{"count":9,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3070\/revisions"}],"predecessor-version":[{"id":3086,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3070\/revisions\/3086"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/media\/3071"}],"wp:attachment":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3070"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3070"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3070"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}