{"id":4521,"date":"2025-11-14T17:25:37","date_gmt":"2025-11-14T09:25:37","guid":{"rendered":"https:\/\/www.ihub1.com\/?p=4521"},"modified":"2025-11-14T17:27:16","modified_gmt":"2025-11-14T09:27:16","slug":"lecun%e5%9c%a8meta%e7%9a%84%e6%9c%80%e5%90%8e%e4%b8%80%e7%af%87%e8%ae%ba%e6%96%87","status":"publish","type":"post","link":"https:\/\/www.ihub1.com\/index.php\/2025\/11\/14\/lecun%e5%9c%a8meta%e7%9a%84%e6%9c%80%e5%90%8e%e4%b8%80%e7%af%87%e8%ae%ba%e6%96%87\/","title":{"rendered":"LeCun\u5728Meta\u7684\u6700\u540e\u4e00\u7bc7\u8bba\u6587"},"content":{"rendered":"\t\t
\n\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

\u300aLeJEPA\uff1a\u65e0\u9700\u542f\u53d1\u5f0f\u7684\u53ef\u8bc1\u660e\u4e14\u53ef\u6269\u5c55\u7684\u81ea\u76d1\u7763\u5b66\u4e60\u300b<\/strong>\u3002<\/p>

” \u8fd9\u53ef\u80fd\u662f LeCun \u4ee5 Meta \u8eab\u4efd\u53d1\u8868\u7684\u6700\u540e\u4e00\u7bc7\u8bba\u6587\u4e86\u3002”<\/p>

<\/div><\/div>

\u6ca1\u9519\uff0c\u8fd9\u7bc7\u5e26 “Le” \u7684\u8bba\u6587\uff0c\u4ecb\u7ecd\u4e86\u4e00\u79cd\u81ea\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\uff0c\u4e8e 11 \u6708 11 \u65e5\u5728 arXiv \u63d0\u4ea4\uff0c\u662f LeCun \u7684\u6700\u65b0\u516c\u5f00\u6210\u679c\u3002<\/p>

\u4e5f\u662f\u5728\u8fd9\u4e00\u5929\uff0c\u4ed6\u79bb\u804c Meta \u7684\u6d88\u606f\u88ab\u66dd\u5149\u3002<\/p>

\u5982\u679c\u8bf4 LeCun \u5728 2013 \u5e74\u52a0\u5165\u5f00\u542f\u4e86 Meta AI \u7814\u7a76\u7684\u4e00\u4e2a\u65f6\u4ee3\uff0c\u90a3\u4e48LeJEPA<\/strong>\u5c31\u662f\u4ed6\u5728 Meta \u7684\u544a\u522b\u4e4b\u4f5c\u3002<\/p>

LeJEPA<\/strong>\u7a76\u7adf\u662f\u600e\u6837\u7684 ” \u6700\u540e\u4e00\u821e “\uff1f<\/p>

LeJEPA\uff1a\u57fa\u4e8e\u5404\u5411\u540c\u6027\u9ad8\u65af\u5d4c\u5165\u7684\u81ea\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5<\/p>

LeJEPA<\/strong>\u6838\u5fc3\u662f\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u5404\u5411\u540c\u6027\u9ad8\u65af\u5d4c\u5165<\/strong>\u7684\u81ea\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\uff0c\u901a\u8fc7\u5f15\u5165SIGReg<\/strong>\u6b63\u5219\u5316\uff0c\u6709\u6548\u89e3\u51b3\u4e86\u8868\u793a\u5d29\u6e83\u95ee\u9898\uff0c\u5e76\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>

\u5728\u4f20\u7edf\u7684 JEPA \u6846\u67b6\u4e2d\uff0c\u9884\u6d4b\u4efb\u52a1\u5e38\u9762\u4e34\u8868\u793a\u5d29\u6e83<\/strong>\u7684\u95ee\u9898\u3002<\/p>

\u8fd9\u610f\u5473\u7740\uff0c\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u53ef\u80fd\u5c06\u6240\u6709\u8f93\u5165\u6620\u5c04\u5230\u5355\u4e00\u7684\u70b9\u6216\u4f4e\u7ef4\u7a7a\u95f4\uff0c\u5bfc\u81f4\u5d4c\u5165\u7a7a\u95f4\u4e2d\u7684\u6837\u672c\u4e0d\u53ef\u533a\u5206\uff0c\u4ece\u800c\u65e0\u6cd5\u6709\u6548\u6355\u6349\u6837\u672c\u95f4\u7684\u8bed\u4e49\u5dee\u5f02\u3002<\/p>

\u9488\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u73b0\u6709\u65b9\u6cd5\u4f9d\u8d56\u542f\u53d1\u5f0f\u6280\u672f\uff0c\u5982\u505c\u6b62\u68af\u5ea6\u3001\u975e\u5bf9\u79f0\u89c6\u56fe\u751f\u6210\u548c\u6559\u5e08 – \u5b66\u751f\u7f51\u7edc\uff0c\u4f46\u8fd9\u4e9b\u65b9\u6cd5\u7531\u4e8e\u7f3a\u4e4f\u5bf9 JEPA \u57fa\u7840\u7406\u8bba\u7684\u63a2\u7d22\uff0c\u88ab\u89c6\u4e3a\u66ff\u4ee3\u65b9\u6848\u3002<\/p>

\u57fa\u4e8e\u4ee5\u4e0a\u80cc\u666f\uff0c\u7814\u7a76\u63d0\u51fa\u4e00\u79cd\u65b0\u7684 JEPA \u6846\u67b6\u2014\u2014\u6f5c\u5728\u6b27\u51e0\u91cc\u5f97 JEPA\uff08Latent-Euclidean Joint Embedding PredictiveArchitecture\uff0cLeJEPA\uff09<\/strong>\uff0c\u5176\u6838\u5fc3\u662f\u4f7f\u5d4c\u5165\u7a7a\u95f4\u9075\u5faa\u7279\u5b9a\u7684\u7edf\u8ba1\u5206\u5e03\uff0c\u4ece\u800c\u63d0\u5347\u6a21\u578b\u7684\u9884\u6d4b\u6027\u80fd\u3002<\/p>

\u5d4c\u5165\u5206\u5e03\u7684\u5f71\u54cd<\/p>

\u9996\u5148\uff0c\u7814\u7a76\u901a\u8fc7\u6700\u5c0f\u4e8c\u4e58\u56de\u5f52\uff08OLS\uff09\u5206\u6790\u4e86\u5d4c\u5165\u5206\u5e03\u5bf9\u504f\u5dee\u548c\u65b9\u5dee\u7684\u5f71\u54cd\u3002<\/p>

\u7ed3\u679c\u8868\u660e\uff0c\u7b49\u5411\u9ad8\u65af\u5206\u5e03\u80fd\u591f\u6700\u5c0f\u5316\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u504f\u5dee\u548c\u65b9\u5dee\u3002<\/p>

\u7279\u522b\u5730\uff0c\u5728\u603b\u65b9\u5dee\u76f8\u540c\u7684\u60c5\u51b5\u4e0b\uff0c\u975e\u7b49\u5411\u5206\u5e03\u4f1a\u5bfc\u81f4\u66f4\u9ad8\u7684\u504f\u5dee\u548c\u65b9\u5dee\uff0c\u800c\u7b49\u5411\u9ad8\u65af\u5206\u5e03\u5219\u80fd\u591f\u6709\u6548\u5730\u4fdd\u8bc1\u6700\u5c0f\u7684\u504f\u5dee\u548c\u65b9\u5dee\uff0c\u4ece\u800c\u63d0\u9ad8\u4e0b\u6e38\u4efb\u52a1\u7684\u7a33\u5b9a\u6027\u548c\u51c6\u786e\u6027\u3002<\/p>

\u901a\u8fc7\u5728\u975e\u7ebf\u6027\u63a2\u6d4b\u548c\u51e0\u4f55\u76f4\u89c9\u65b9\u9762\u7684\u5b9e\u9a8c\uff0c\u7814\u7a76\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86\u7b49\u5411\u9ad8\u65af\u5206\u5e03\u7684\u4f18\u8d8a\u6027\u3002<\/p>

\u5b9e\u9a8c\u8868\u660e\uff0c\u65e0\u8bba\u662f\u5728\u56de\u5f52\u4efb\u52a1\u8fd8\u662f\u5206\u7c7b\u4efb\u52a1\u4e2d\uff0c\u7b49\u5411\u9ad8\u65af\u5206\u5e03\u90fd\u80fd\u4fdd\u6301\u6700\u5c0f\u7684\u8bef\u5dee\uff0c\u800c\u975e\u7b49\u5411\u5206\u5e03\u5219\u8868\u73b0\u51fa\u8f83\u9ad8\u7684\u65b9\u5dee\u3002<\/p>

<\/div><\/div>

\u7814\u7a76\u8868\u660e\uff0c\u00a0\u5404\u5411\u540c\u6027\u9ad8\u65af\u5206\u5e03<\/strong>\u662f\u5d4c\u5165\u7a7a\u95f4\u7684\u6700\u4f73\u5206\u5e03\uff0c\u5b83\u53ef\u4ee5\u5728\u6ca1\u6709\u4efb\u52a1\u4fe1\u606f\u7684\u60c5\u51b5\u4e0b\uff0c\u4fdd\u8bc1\u6700\u5c0f\u5316\u504f\u5dee\u548c\u65b9\u5dee\uff0c\u4ece\u800c\u63d0\u9ad8\u4e0b\u6e38\u4efb\u52a1\u7684\u8868\u73b0\u3002<\/p>

SIGReg\uff1a\u5b9e\u73b0\u9ad8\u65af\u5206\u5e03\u7684\u6b63\u5219\u5316<\/p>

\u4e3a\u5b9e\u73b0\u4e0a\u8ff0\u5206\u5e03\u5339\u914d\uff0c\u7814\u7a76\u63d0\u51fa\u4e86\u8349\u56fe\u5316\u5404\u5411\u540c\u6027\u9ad8\u65af\u6b63\u5219\u5316\uff08Sketched Isotropic Gaussian Regularization\uff0cSIGReg\uff09<\/strong>\uff0c\u8fd9\u662f\u4e00\u79cd\u53ef\u5904\u7406\u3001\u53ef\u8bc1\u660e\u6b63\u786e\u7684\u6b63\u5219\u5316\u65b9\u6cd5\u3002<\/p>

SIGReg \u7684\u521b\u65b0\u70b9\u5728\u4e8e\uff1a<\/p>

\u5c06\u5206\u5e03\u5339\u914d\u95ee\u9898\u8f6c\u5316\u4e3a\u7edf\u8ba1\u5047\u8bbe\u68c0\u9a8c\uff0c\u901a\u8fc7\u96f6\u5047\u8bbe\u4e0e\u76ee\u6807\u5206\u5e03\u7684\u5339\u914d\u6765\u5b9e\u73b0<\/p>

\u63d0\u4f9b\u4e86\u4e00\u79cd\u6d4b\u8bd5\u65b9\u6cd5\uff0c\u4fdd\u8bc1\u5728\u591a GPU \u8bad\u7ec3\u65f6\u7684\u9ad8\u6548\u6027\uff0c\u5e76\u786e\u4fdd\u68af\u5ea6\u548c\u66f2\u7387\u6709\u754c<\/p>

\u89e3\u51b3\u4e86\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u7684\u7ef4\u5ea6\u707e\u96be\u95ee\u9898\u3002<\/p>

SIGReg \u901a\u8fc7\u5355\u53d8\u91cf\u65b9\u5411\u68c0\u9a8c\uff0c\u7ed3\u5408 Epps-Pulley \u6d4b\u8bd5\u6765\u5224\u65ad\u5d4c\u5165\u5206\u5e03\u4e0e\u76ee\u6807\u5206\u5e03\uff08\u7b49\u5411\u9ad8\u65af\u5206\u5e03\uff09\u7684\u5339\u914d\u7a0b\u5ea6\u3002<\/p>

\u5b83\u5c06\u5206\u5e03\u5339\u914d\u8f6c\u5316\u4e3a\u96f6\u5047\u8bbe\u4e0e\u5907\u62e9\u5047\u8bbe\u7684\u68c0\u9a8c\uff0c\u5e76\u901a\u8fc7\u7edf\u8ba1\u91cf\u5224\u65ad\u662f\u5426\u62d2\u7edd\u96f6\u5047\u8bbe\uff0c\u4ece\u800c\u786e\u8ba4\u5206\u5e03\u662f\u5426\u5339\u914d\u3002<\/p>

\u9ad8\u7ef4\u95ee\u9898\u7684\u89e3\u51b3<\/p>

SIGReg \u8fd8\u901a\u8fc7\u4e24\u6761\u673a\u5236\u89e3\u51b3\u4e86\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u7684\u8ba1\u7b97\u6311\u6218\uff1a<\/p>

\u5e73\u6ed1\u6027\uff1a\u5d4c\u5165\u51fd\u6570\u7684 Sobolev \u5e73\u6ed1\u6027\u4fdd\u8bc1\u4e86\u5728\u4ec5\u9700 O ( K ) \u4e2a\u65b9\u5411\u5207\u7247\u7684\u60c5\u51b5\u4e0b\u5373\u53ef\u6709\u6548\u7ea6\u675f\u6574\u4e2a\u7a7a\u95f4\uff0c\u8fdb\u884c\u6709\u6548\u7684\u7edf\u8ba1\u68c0\u9a8c\u3002<\/p>

SGD \u8fed\u4ee3\u7279\u6027\uff1a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u65b9\u5411\u7684\u91cd\u590d\u91c7\u6837\u7d2f\u79ef\u6548\u5e94\u4f7f\u5f97\u5373\u4f7f\u65b9\u5411\u6570\u91cf\u5f88\u5c11\uff08\u5982 M=16\uff09\uff0c\u4e5f\u80fd\u8fc5\u901f\u6536\u655b\u5230\u5404\u5411\u540c\u6027\u5206\u5e03\uff0c\u4f18\u4e8e\u56fa\u5b9a\u65b9\u5411\u96c6\u3002<\/p>

\u5728\u5b9e\u73b0\u65b9\u9762\uff0cLeJEPA \u7ed3\u5408\u4e86 SIGReg \u548c\u9884\u6d4b\u635f\u5931\u4e24\u90e8\u5206\uff0c\u901a\u8fc7 Epps-Pulley \u7edf\u8ba1\u91cf\u5b9e\u73b0\u5206\u5e03\u5339\u914d\uff0c\u5e76\u901a\u8fc7\u5c0f\u6279\u6b21\u8bad\u7ec3\u4fdd\u8bc1\u8ba1\u7b97\u6548\u7387\u548c\u7a33\u5b9a\u6027\u3002\u6700\u7ec8\u7684\u603b\u635f\u5931\u662f SIGReg \u635f\u5931\u548c\u9884\u6d4b\u635f\u5931\u7684\u52a0\u6743\u548c\u3002<\/p>

SIGReg \u635f\u5931\uff1a\u901a\u8fc7 Epps-Pulley \u7edf\u8ba1\u91cf\u8ba1\u7b97\uff0c\u786e\u4fdd\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u68af\u5ea6\u6709\u754c\uff0c\u5e76\u901a\u8fc7\u79ef\u5206\u8fd1\u4f3c\u63d0\u5347\u8ba1\u7b97\u6548\u7387\u3002\u5c0f\u6279\u6b21\u8bad\u7ec3\u5f15\u5165\u7684\u504f\u5dee\u5bf9\u8bad\u7ec3\u5f71\u54cd\u8f83\u5c0f\u3002<\/p>

\u9884\u6d4b\u635f\u5931\uff1a\u4e0e DINO \u65b9\u6cd5\u76f8\u4f3c\uff0c\u901a\u8fc7\u8ba1\u7b97\u6240\u6709\u89c6\u56fe\u9884\u6d4b\u5168\u5c40\u89c6\u56fe\u7684\u5dee\u5f02\u3002<\/p>

LeJEPA \u603b\u635f\u5931\uff1a\u662f SIGReg \u635f\u5931\u548c\u9884\u6d4b\u635f\u5931\u7684\u52a0\u6743\u548c\uff0c\u5176\u4e2d\u4e00\u4e2a\u8d85\u53c2\u6570 \u03bb \u7528\u4e8e\u5e73\u8861\u8fd9\u4e24\u90e8\u5206\u7684\u6743\u91cd\u3002<\/p>

\u5b9e\u9a8c\u9a8c\u8bc1\u4e0e\u7ed3\u679c<\/p>

\u4e3a\u4e86\u9a8c\u8bc1 LeJEPA \u7684\u53ef\u9760\u6027\uff0c\u7814\u7a76\u5728\u591a\u4e2a\u5927\u578b\u67b6\u6784\u4e0a\u8fdb\u884c\u5b9e\u9a8c\uff0c\u5305\u62ec ViT\u3001ConvNeXt\u3001ResNet\u3001MaxViT \u548c Swin Transformer \u7b49\uff0c\u6a21\u578b\u89c4\u6a21\u63a5\u8fd1 10 \u4ebf\u53c2\u6570\u3002<\/p>

<\/div><\/div>

\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0cLeJEPA \u5728\u8fd9\u4e9b\u67b6\u6784\u4e0a\u8868\u73b0\u8d85\u8d8a\u73b0\u6709\u65b9\u6cd5\uff0c\u4e14\u4fdd\u6301\u4e86\u8bad\u7ec3\u7684\u7b80\u4fbf\u6027\u548c\u9c81\u68d2\u6027\u3002<\/p>

\u7279\u522b\u5730\uff0c\u5728\u9886\u57df\u7279\u5b9a\u7684\u6570\u636e\u96c6\uff08\u5982 Galaxy10\u3001Food101\uff09\u4e0a\uff0cLeJEPA \u5728\u76f4\u63a5\u5728\u76ee\u6807\u6570\u636e\u4e0a\u9884\u8bad\u7ec3\u65f6\u8d85\u8d8a\u4e86\u57fa\u4e8e DINOv2 \u7684\u8fc1\u79fb\u5b66\u4e60\u65b9\u6cd5\u3002<\/p>

<\/div><\/div>

\u603b\u7684\u6765\u8bf4\uff0cLeJEPA \u5ef6\u7eed\u4e86\u4e4b\u524d JEPA \u7684\u63a2\u7d22\uff0c\u91cd\u65b0\u786e\u7acb\u4e86\u81ea\u76d1\u7763\u5b66\u4e60\u4f5c\u4e3a AI \u7814\u7a76\u7684\u6838\u5fc3\u65b9\u6cd5\u3002<\/p>

LeJEPA \u901a\u8fc7\u63d0\u4f9b\u4e00\u4e2a\u7b80\u5355\u4e14\u7406\u8bba\u4e0a\u6709\u652f\u6301\u7684\u6846\u67b6\uff0c\u4f7f\u5f97\u4ece\u6570\u636e\u4e2d\u5b66\u4e60\u8868\u793a\u53d8\u5f97\u66f4\u52a0\u9ad8\u6548\uff0c\u5e76\u5728\u591a\u4e2a\u4efb\u52a1\u4e2d\u5c55\u793a\u4e86\u4f18\u8d8a\u7684\u6027\u80fd\u3002<\/p>

JEPA \u4e16\u754c\u6a21\u578b<\/p>

\u81ea LeCun \u4e8e 2022 \u5e74\u5728\u300aA Path Towards Autonomous Machine Intelligence\u300b\u4e2d\u9996\u6b21\u63d0\u51fa JEPA \u4ee5\u6765\uff0c\u57fa\u4e8e JEPA \u7684\u67b6\u6784\u5df2\u7ecf\u53d1\u5c55\u4e86\u6574\u6574\u4e09\u5e74\u3002<\/p>

<\/div><\/div>

JEPA\uff08Joint-Embedding Predictive Architecture\uff09\u662f\u4e00\u79cd\u81ea\u76d1\u7763\u5b66\u4e60\u6846\u67b6\uff0c\u65e8\u5728\u901a\u8fc7\u57fa\u4e8e\u5d4c\u5165\u7a7a\u95f4\u7684\u8054\u5408\u9884\u6d4b\u65b9\u6cd5\u6765\u63d0\u5347\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\u548c\u63a8\u7406\u80fd\u529b\u3002<\/p>

\u4e0e\u751f\u6210\u5f0f\u6a21\u578b\u4e0d\u540c\uff0c\u5b83\u5e76\u4e0d\u80fd\u7b80\u5355\u5730\u7528\u4e8e\u4ece x \u9884\u6d4b y\uff0c\u5b83\u4ec5\u6355\u6349 x \u548c y \u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u800c\u65e0\u9700\u663e\u5f0f\u751f\u6210 y \u7684\u9884\u6d4b\u3002<\/p>

<\/div><\/div>

\u6b64\u5916\uff0c\u4e3a\u4e86\u5e94\u5bf9\u957f\u671f\u89c4\u5212\u95ee\u9898\uff0cJEPA \u8fd8\u53ef\u4ee5\u8fdb\u4e00\u6b65\u901a\u8fc7\u5206\u5c42\u67b6\u6784\uff08\u5373 H-JEPA\uff09\u6765\u589e\u5f3a\u5176\u62bd\u8c61\u80fd\u529b\u3002<\/p>

\u5728 H-JEPA \u4e2d\uff0c\u4f4e\u5c42\u8868\u793a\u5904\u7406\u77ed\u671f\u9884\u6d4b\u4efb\u52a1\uff0c\u800c\u9ad8\u5c42\u8868\u793a\u5219\u7528\u4e8e\u957f\u671f\u9884\u6d4b\u3002<\/p>

\u8fd9\u79cd\u5206\u5c42\u7ed3\u6784\u4f7f\u5f97\u6a21\u578b\u5728\u8fdb\u884c\u957f\u671f\u89c4\u5212\u65f6\uff0c\u53ef\u4ee5\u5728\u4e0d\u540c\u7684\u62bd\u8c61\u5c42\u6b21\u4e0a\u8fdb\u884c\u64cd\u4f5c\uff0c\u4ece\u800c\u63d0\u9ad8\u53ef\u9884\u6d4b\u6027\u548c\u51cf\u5c11\u4fe1\u606f\u635f\u5931\u3002<\/p>

<\/div><\/div>

\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cJEPA \u67b6\u6784\u901a\u5e38\u4e0e\u4e16\u754c\u6a21\u578b<\/strong>\u5bc6\u5207\u76f8\u5173\uff0c\u53ea\u4e0d\u8fc7\u5176\u4e0e\u4e00\u822c\u610f\u4e49\u4e0a\u7684\u4e16\u754c\u6a21\u578b\u4ecd\u6709\u533a\u522b\u3002<\/p>

\u4f20\u7edf\u7684\u4e16\u754c\u6a21\u578b\u4e00\u822c\u6307\u80fd\u591f\u6a21\u62df\u73af\u5883\u6216\u7cfb\u7edf\u7684\u6a21\u578b\uff0c\u5176\u4e3b\u8981\u76ee\u7684\u662f\u901a\u8fc7\u9884\u6d4b\u672a\u6765\u72b6\u6001\u6765\u5b9e\u73b0\u957f\u671f\u89c4\u5212\u548c\u51b3\u7b56\uff08\u5982\u5f3a\u5316\u5b66\u4e60\uff09\u3002<\/p>

\u800c JEPA \u5219\u662f\u4e00\u79cd\u901a\u8fc7\u8054\u5408\u5d4c\u5165\u7a7a\u95f4\u5b66\u4e60\u72b6\u6001\u4e0e\u52a8\u4f5c\u8f6c\u79fb\u7684\u67b6\u6784\uff0c\u91cd\u70b9\u5728\u4e8e\u7ed3\u5408\u8868\u793a\u5b66\u4e60\u548c\u81ea\u76d1\u7763\u5b66\u4e60\u6765\u5b8c\u6210\u9884\u6d4b\u548c\u89c4\u5212\u4efb\u52a1\u3002<\/p>

<\/div><\/div>

\u5728 JEPA \u4e2d\uff0c\u4e16\u754c\u6a21\u578b\u7684\u76ee\u7684\u662f\u9884\u6d4b\u4e16\u754c\u72b6\u6001\u7684\u672a\u6765\u8868\u73b0\u5f62\u5f0f\u3002<\/p>

\u5177\u4f53\u800c\u8a00\uff0cJEPA \u901a\u8fc7\u5b66\u4e60\u72b6\u6001\u4e0e\u52a8\u4f5c\u7684\u8f6c\u79fb\u6765\u8bad\u7ec3\u4e16\u754c\u6a21\u578b\uff0c\u5176\u6838\u5fc3\u5728\u4e8e\u4ece\u5f53\u524d\u72b6\u6001\u7684\u8868\u793a\u63a8\u65ad\u672a\u6765\u72b6\u6001\u7684\u8868\u793a\uff0c\u8fd9\u4e00\u8fc7\u7a0b\u662f\u5728\u8054\u5408\u5d4c\u5165\u7a7a\u95f4\u4e2d\u5b8c\u6210\u7684\u3002\u8be5\u7a7a\u95f4\u901a\u8fc7\u6700\u5c0f\u5316\u9884\u6d4b\u8bef\u5dee\u6765\u5b66\u4e60\u72b6\u6001\u8868\u793a\u4e0e\u52a8\u4f5c\u4e4b\u95f4\u7684\u5173\u7cfb\u3002<\/p>

\u5c3d\u7ba1\u6700\u521d\u7684 JEPA \u8bba\u6587\u5448\u73b0\u51fa\u4e00\u79cd\u5bf9\u751f\u6210\u5f0f AI \u7684\u53cd\u601d\uff0c\u63cf\u8ff0\u4e86\u4eba\u5de5\u667a\u80fd\u672a\u6765\u7684\u613f\u666f\uff0c\u5e76\u6307\u51fa\u8fd9\u4e00\u613f\u666f\u53ef\u80fd\u9700\u8981\u6570\u5341\u5e74\u7684\u65f6\u95f4\u624d\u80fd\u5b9e\u73b0\u3002<\/p>

\u4f46\u81ea 2022 \u5e74\u590f\u5b63\u53d1\u5e03\u4ee5\u6765\uff0c\u5728 LeCun \u7684\u63a8\u52a8\u4e0b\uff0cJEPA \u67b6\u6784\u5df2\u7ecf\u53d6\u5f97\u4e86\u4e00\u4e9b\u663e\u8457\u8fdb\u5c55\u3002<\/p>

I-JEPA: Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture<\/strong><\/p>

<\/div><\/div>

\u4e0e\u5176\u4ed6\u56fe\u50cf SSL \u65b9\u6cd5\u76f8\u6bd4\uff0cI-JEPA \u5145\u5206\u5229\u7528\u4e86 Transformer \u67b6\u6784\u7684\u7075\u6d3b\u6027\u3002<\/p>

\u5728 I-JEPA \u4e2d\uff0c\u4e0a\u4e0b\u6587\u7f16\u7801\u5668\u662f\u4e00\u4e2a ViT\uff0c\u5b83\u4ec5\u5904\u7406\u53ef\u89c1\u7684\u4e0a\u4e0b\u6587\u5757\u3002<\/p>

\u9884\u6d4b\u5668\u63a5\u6536\u4e0a\u4e0b\u6587\u7f16\u7801\u5668\u7684\u8f93\u51fa\uff0c\u5e76\u6839\u636e\u4f4d\u7f6e\u6807\u8bb0\uff08\u4ee5\u989c\u8272\u663e\u793a\uff09\u9884\u6d4b\u7279\u5b9a\u4f4d\u7f6e\u7684\u76ee\u6807\u5757\u8868\u793a\u3002<\/p>

\u76ee\u6807\u8868\u793a\u5bf9\u5e94\u4e8e\u76ee\u6807\u7f16\u7801\u5668\u7684\u8f93\u51fa\uff0c\u6743\u91cd\u5728\u6bcf\u6b21\u8fed\u4ee3\u4e2d\u901a\u8fc7\u4e0a\u4e0b\u6587\u7f16\u7801\u5668\u6743\u91cd\u7684\u6307\u6570\u6ed1\u52a8\u5e73\u5747\u8fdb\u884c\u66f4\u65b0\u3002<\/p>

V-JEPA: Revisiting Feature Prediction for Learning Visual Representations from Video<\/strong><\/p>

<\/div><\/div>

V-JEPA \u662f I-JEPA \u5728\u89c6\u9891\u9886\u57df\u7684\u6269\u5c55\uff0c\u5b83\u901a\u8fc7\u5c06\u89c6\u9891\u89c6\u4e3a 3D \u56fe\u50cf\u6765\u5b9e\u73b0\u8fd9\u4e00\u70b9\u3002<\/p>

\u8bad\u7ec3\u8fc7\u7a0b\u57fa\u4e8e\u4e00\u4e2a\u5305\u542b T \u5e27\u7684\u89c6\u9891\u7247\u6bb5\uff0c\u7a7a\u95f4\u5206\u8fa8\u7387\u4e3a H \u00d7 W\uff0c\u5e76\u5c06\u5176\u5c55\u5e73\u6210\u4e00\u4e2a L \u4e2a token \u7684\u5e8f\u5217\u3002<\/p>

\u9996\u5148\u901a\u8fc7\u4ece\u89c6\u9891\u7247\u6bb5\u4e2d\u53bb\u9664\u4e00\u4e9b tokens \u6765\u83b7\u5f97 x-encoder \u7684\u8f93\u5165\u3002<\/p>

\u63a5\u7740\uff0cx-encoder \u5904\u7406\u88ab\u5c4f\u853d\u7684\u89c6\u9891\u5e8f\u5217\uff0c\u5e76\u4e3a\u6bcf\u4e2a\u8f93\u5165 token \u8f93\u51fa\u4e00\u4e2a\u5d4c\u5165\u5411\u91cf\u3002<\/p>

\u7136\u540e\uff0cx-encoder \u7684\u8f93\u51fa\u4e0e\u4e00\u7ec4\u53ef\u5b66\u4e60\u7684 mask tokens \u8fdb\u884c\u62fc\u63a5\uff0c\u8fd9\u4e9b mask tokens \u5305\u542b\u4e86\u88ab\u5c4f\u853d\u7684\u65f6\u7a7a\u8865\u4e01\u7684\u4f4d\u7f6e\u4fe1\u606f\u5d4c\u5165\u3002<\/p>

\u9884\u6d4b\u7f51\u7edc\u5904\u7406\u62fc\u63a5\u540e\u7684 token \u5e8f\u5217\uff0c\u5e76\u4e3a\u6bcf\u4e2a mask token \u8f93\u51fa\u4e00\u4e2a\u5d4c\u5165\u5411\u91cf\u3002<\/p>

\u6700\u540e\uff0c\u9884\u6d4b\u7f51\u7edc\u7684\u8f93\u51fa\u901a\u8fc7 L1 \u635f\u5931\u56de\u5f52\u5230\u9884\u6d4b\u76ee\u6807\u3002\u9884\u6d4b\u76ee\u6807\u5bf9\u5e94\u4e8e y-encoder \u7684\u8f93\u51fa\u3002<\/p>

\u4eca\u5e74 7 \u6708\uff0cLeCun \u56e2\u961f\u53c8\u8fdb\u4e00\u6b65\u53d1\u5e03\u4e86V-JEPA 2<\/strong>\u3002<\/p>

<\/div><\/div>

V -JEPA 2 \u57fa\u4e8e V-JEPA\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u52a8\u4f5c\u9884\u6d4b\u548c\u4e16\u754c\u5efa\u6a21\u80fd\u529b\uff0c\u4f7f\u673a\u5668\u4eba\u80fd\u591f\u4e0e\u964c\u751f\u7269\u4f53\u548c\u73af\u5883\u8fdb\u884c\u4ea4\u4e92\uff0c\u4ece\u800c\u5b8c\u6210\u4efb\u52a1\u3002<\/p>

MC-JEPA: A Joint-Embedding Predictive Architecture for Self-Supervised Learning of Motion and Content Features<\/strong><\/p>

<\/div><\/div>

MC-JEPA \u662f JEPA \u7684\u6269\u5c55\uff0c\u4f7f\u5176\u80fd\u591f\u5305\u542b\u8fd0\u52a8\u4fe1\u606f\u3002<\/p>

\u67b6\u6784\u901a\u8fc7\u5728\u89c6\u9891\u4e2d\u4f7f\u7528\u5149\u6d41\u4f30\u8ba1\u5b66\u4e60\u8fd0\u52a8\uff0c\u5e76\u901a\u8fc7\u56fe\u50cf\u89c6\u89d2\u7684\u8054\u5408\u5d4c\u5165\u5b66\u4e60\u5185\u5bb9\uff0c\u4ee5\u591a\u4efb\u52a1\u65b9\u5f0f\u5171\u4eab\u7f16\u7801\u5668\u3002\u5176\u5149\u6d41\u4f30\u8ba1\u67b6\u6784\u57fa\u4e8e PWC-Net\u3002<\/p>

\u5149\u6d41\u4f30\u8ba1\u91c7\u7528\u7c97\u5230\u7cbe\u7684\u65b9\u5f0f\uff0c\u5148\u4f30\u7b97\u521d\u6b65\u7684\u5149\u6d41\uff0c\u518d\u9010\u5c42\u7cbe\u7ec6\u5316\uff0c\u6700\u7ec8\u901a\u8fc7\u56fe\u50cf\u91cd\u5efa\u635f\u5931\u6765\u4f18\u5316\u3002<\/p>

\u6574\u4e2a\u8fc7\u7a0b\u4e2d\uff0c\u524d\u5411 – \u53cd\u5411\u5149\u6d41\u7684\u4e00\u81f4\u6027\u901a\u8fc7\u5faa\u73af\u4e00\u81f4\u6027\u635f\u5931\u5f97\u5230\u5f3a\u5316\u3002\u4e3a\u4e86\u907f\u514d\u8bad\u7ec3\u4e0d\u7a33\u5b9a\uff0c\u6a21\u578b\u5728\u6bcf\u4e2a\u7279\u5f81\u5c42\u4f7f\u7528\u65b9\u5dee – \u534f\u65b9\u5dee\u6b63\u5219\u5316\u3002<\/p>

\u9664\u4e0a\u8ff0\u5de5\u4f5c\u4ee5\u5916\uff0cLeCun \u53ca\u5176\u56e2\u961f\u8fd8\u8fdb\u4e00\u6b65\u56f4\u7ed5 JEPA \u53d1\u8868\u4e86JEPAs<\/strong>\u3001LLM-JEAP<\/strong>\u7b49\u5de5\u4f5c\u3002<\/p>

\u603b\u7684\u6765\u8bf4\uff0c\u5c3d\u7ba1 JEPA \u67b6\u6784\u76f8\u8f83\u4e8e\u4f20\u7edf\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\u63d0\u4f9b\u4e86\u4e00\u6761\u65b0\u8def\u5f84\uff0c\u4f46\u5176\u81ea\u53d1\u5e03\u4ee5\u6765\u5c31\u6bc1\u8a89\u53c2\u534a\u3002<\/p>

\u6bd4\u5982\uff0c\u6709\u4eba\u6279\u8bc4 JEPA \u67b6\u6784\u8fc7\u4e8e\u62bd\u8c61\uff0c\u96be\u4ee5\u5e94\u7528\u5230\u6700\u524d\u6cbf\u7684\u4e3b\u6d41\u6a21\u578b\u4e2d\uff0c\u5176\u4e0e\u751f\u6210\u5f0f AI \u7684\u80cc\u79bb\u4e5f\u8ba9\u4e0d\u5c11\u7f51\u53cb\u5c06 Meta \u7684\u5931\u5229\u5f52\u548e\u4e8e LeCun \u7684\u5b66\u672f\u53d6\u5411\u3002<\/p>

\u751a\u81f3\u6709 Meta AI \u5458\u5de5\u8868\u793a\uff1aFair \u5c1a\u672a\u8bc1\u660e\u81ea\u5df1\u662f DeepMind \u7b49\u7814\u7a76\u673a\u6784\u7684\u771f\u6b63\u7ade\u4e89\u5bf9\u624b\u3002<\/p>

\u800c\u6700\u65b0\u53d1\u5e03\u7684 LeJEPA \u4e5f\u5f7b\u5e95\u5728 LeCun \u548c Meta \u4e4b\u95f4\u5212\u4e0a\u4e86\u53e5\u53f7\u3002<\/p>

<\/div><\/div>

\u4e0d\u8fc7\uff0cLeJEPA \u867d\u7136\u53ef\u80fd LeCun \u5728 Meta \u7684\u7814\u7a76\u7ec8\u70b9\uff0c\u4f46\u80af\u5b9a\u4e0d\u662f JEPA \u53d1\u5c55\u7684\u7ec8\u7ed3\u3002<\/p>

\u636e\u6d88\u606f\u79f0\uff0cLeCun \u5df2\u7ecf\u5728\u7b79\u96c6\u8d44\u91d1\uff0c\u521b\u529e\u521d\u521b\u516c\u53f8\uff0c\u7ee7\u7eed\u63a8\u8fdb\u4ed6\u5728\u4e16\u754c\u6a21\u578b\u65b9\u9762\u7684\u5de5\u4f5c\u3002<\/p>

\u800c\u5728 AI \u70ed\u8fa3\u6eda\u70eb\u4e4b\u4e0b\uff0c\u8fd9\u4f4d 65 \u5c81\u56fe\u7075\u5956\u5f97\u4e3b\u3001AI \u4e09\u5de8\u5934\u4e4b\u4e00\u7684\u6559\u7236\u7ea7\u4eba\u7269\uff0c\u521b\u4e1a\u53c8\u5c82\u4f1a\u7f3a\u8d44\u6e90\uff1f<\/p>

\u552f\u4e00\u7684\u95ee\u9898\uff0c\u53ea\u53ef\u80fd\u662f LeCun \u4e00\u76f4\u4ee5\u6765\u7ed9\u4eba\u8fc7\u4e8e ” \u5b66\u672f ” \u7684\u98ce\u683c\u548c\u5370\u8c61\uff0c\u62c5\u5fe7\u4ed6\u5728\u4e25\u683c\u8ba1\u7b97 ROI \u7684\u5546\u4e1a\u4e16\u754c\u91cc\uff0c\u518d\u6b21\u906d\u9047 Meta \u751f\u6daf\u672b\u671f\u91cc\u7684\u5206\u6b67\u548c\u56f0\u5883\u3002<\/p>

One more thing<\/p>

\u867d\u7136 LeCun \u4e0e Meta \u7684 ” \u5206\u624b ” \u8bf4\u4e0d\u4e0a\u4f53\u9762\uff0c\u4f46\u5728 Meta \u7684\u8fd9\u5341\u591a\u5e74\u4e5f\u8bb8\u662f LeCun \u751f\u6daf\u4e2d\u6700\u987a\u7684\u65f6\u671f\u2014\u2014<\/p>

\u6df1\u5ea6\u5b66\u4e60\u7206\u53d1\u3001\u52a0\u5165\u7ebd\u5927\u3001\u5efa\u7acb FAIR\u3001\u529e\u9876\u4f1a\uff08ICLR\uff09\u3001\u62ff\u56fe\u7075\u5956\u3001\u51fa\u81ea\u4f20\uff08\u300a\u79d1\u5b66\u4e4b\u8def\u300b\uff09\u3001\u53d1 paper\uff0c\u5e26\u5b66\u751f\u3002<\/p>

<\/div><\/div>

\u81ea 2013 \u5e74\u52a0\u5165 Meta \u4ee5\u6765\uff0cLeCun \u7684\u4e2a\u4eba\u88ab\u5f15\u98d9\u5347\uff0c\u7ea6\u4e3a406919<\/strong>\u6b21\uff0c\u5360\u603b\u6570\u768493%<\/strong>\u3002<\/p>

\u867d\u7136\u5149\u4e09\u5de8\u5934\u5408\u5199\u7684\u300a\u6df1\u5ea6\u5b66\u4e60\u300b\u7efc\u8ff0\u5c31\u8d21\u732e\u4e8610 \u4e07<\/strong>\u591a\u6b21\uff0c\u4f46\u4e5f\u8db3\u89c1 LeCun \u5b66\u672f\u5f71\u54cd\u529b\u7684\u98d9\u5347\u3002<\/p>

<\/div><\/div>

\u8fd9\u6bb5\u9ec4\u91d1\u65f6\u671f\u4e0d\u4ec5\u89c1\u8bc1\u4e86 LeCun \u4e2a\u4eba\u4e8b\u4e1a\u7684\u5dc5\u5cf0\uff0c\u66f4\u63a8\u52a8\u4e86\u6574\u4e2a AI \u9886\u57df\u7684\u53d1\u5c55\u3002<\/p>

\u6c5f\u6e56\u4e0a\u81f3\u4eca\u6d41\u4f20\u7740\u624e\u514b\u4f2f\u683c\u5f53\u5e74\u73b0\u8eabNIPS<\/strong>\u65f6\u7684\u8f70\u52a8\uff0c\u5e76\u4e14\u90fd\u5728\u8ba4\u4e3a\u4ed6\u4f1a\u5e26\u8d70\u4e00\u7968 AI \u5929\u624d\u2026\u2026<\/p>

<\/div><\/div>

\u25b3<\/strong>\u624e\u514b\u4f2f\u683c\u4e0e\u7ea6\u4e66\u4e9a \u00b7 \u672c\u5409\u5965\u5728 NIPS 2013<\/p>

\u4f46\u6700\u540e\u624e\u514b\u4f2f\u683c\u53ea\u662f\u8bf7\u6765\u4e86 LeCun\uff0c\u4e00\u4eba\u80dc\u8fc7\u5343\u519b\u4e07\u9a6c\u3002<\/p>

LeCun \u6210\u5c31\u4e86 FacebookMeta \u7684 AI \u8f6c\u578b\u4e4b\u540d\uff0c\u6210\u4e3a\u4e86\u91d1\u5b57\u62db\u724c\uff1bMeta \u4e5f\u4e00\u5ea6\u7ed9\u8fc7 LeCun \u8db3\u591f\u81ea\u7531\u7684\u7814\u7a76\u5f85\u9047\u548c\u73af\u5883\u3002<\/p>

<\/div><\/div>

\u53ea\u662f\u6700\u540e\u7ed3\u5c40\u524d\u7684\u6ce2\u6f9c\uff0c\u4ee4\u4eba\u550f\u5618\uff5e<\/p>

\u597d\u5728\uff0c65 \u5c81\u7684 Yann LeCun\uff0c\u4f9d\u65e7\u662f\u95ef\u7684\u5e74\u7eaa\uff5e\uff5e<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"

henry \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n","protected":false},"author":1,"featured_media":4524,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4521","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/4521","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/comments?post=4521"}],"version-history":[{"count":7,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/4521\/revisions"}],"predecessor-version":[{"id":4530,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/4521\/revisions\/4530"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media\/4524"}],"wp:attachment":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media?parent=4521"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/categories?post=4521"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/tags?post=4521"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}