![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-219.png\")
<\/figure>\n\n\n\n\u8bbe\u6bcf\u4e2a\u6a21\u5757\u7684\u8f93\u5165 \\(X \\in \\mathbb{R}^{n \\times d}\\)\uff0c n\u4e3a\u5e8f\u5217\u957f\u5ea6\uff0c d\u4e3a\u7279\u5f81\u7ef4\u5ea6\u3002\u6bcf\u4e2a\u6a21\u5757\u8868\u8fbe\u5982\u4e0b:
\\(Z=\\sigma(X U), \\quad \\tilde{Z}=s(Z), \\quad Y=\\tilde{Z} V\\)
\\(\\sigma\\) \u662fGELU\u7b49\u6fc0\u6d3b\u51fd\u6570\uff0c U \u548c V \u548cTransformer\u4e2d\u7684FFN\u7c7b\u4f3c\u90fd\u662f\u7ebf\u6027\u6620\u5c04\u3002\u4e3a\u4e86\u7b80\u6d01\u8868\u8fbe\u4e0a\u5f0f\u4e2d \u7701\u7565\u4e86shortcuts, normalizations \u548c biases\u3002
\u4e0a\u5f0f\u4e2d\u6700\u91cd\u8981\u7684\u662f\u80fd\u6355\u6349\u7a7a\u95f4\u4ea4\u4e92\u7684 \\(s(\\cdot)\\) \u3002\u5982\u679c\u4e0a\u5f0f\u53bb\u6389 \\(s(\\cdot)\\) \u90a3\u4e48\u5c06\u4e0d\u518d\u80fd\u8fdb\u884c\u7a7a\u95f4\u4ea4\u4e92\u548cFFN \u5e76\u65e0\u533a\u522b\u3002\u6587\u4e2d\u4f5c\u8005\u9009\u62e9\u540d\u4e3a Spatial Gating Unit (SGU) \u7684\u6a21\u5757\u4f5c\u4e3a \\(s(\\cdot)\\) \u6355\u6349\u7a7a\u95f4\u4f9d\u8d56\u3002\u53e6\u5916\uff0cgMLP\u5728NLP\u3001CV\u4efb\u52a1\u4e2d\u9075\u5faa\u4e0eBERT\u3001ViT\u4e00\u6837\u7684\u8f93\u5165\u8f93\u51fa\u89c4\u5219\u3002<\/p>\n\n\n\n
Spatial Gating Unit\uff1a<\/h2>\n\n\n\n
\u4e3a\u4e86\u80fd\u6709\u8de8token\u7684\u4ea4\u4e92\uff0c \\(s(\\cdot)\\) \u64cd\u4f5c\u987b\u5728\u7a7a\u95f4\u7ef4\u5ea6\u3002\u53ef\u4ee5\u7b80\u5355\u7684\u4f7f\u7528\u7ebf\u6027\u6620\u5c04\u8868\u793a\uff1a
\\(f_{W, b}(Z)=W Z+b\\)
\u5176\u4e2d \\(W \\in \\mathbb{R}^{n \\times n}\\) \u8868\u793a\u7a7a\u95f4\u4ea4\u4e92\u7684\u6620\u5c04\u53c2\u6570\u3002\u5728self-attention\u4e2d W \u662f\u901a\u8fc7 Z \u52a8\u6001\u8ba1\u7b97\u5f97\u5230\u7684\u3002 \u6b64\u6587\u5bf9\u4e0a\u5f0f\u4f7f\u7528gating\u64cd\u4f5c\u4ee5\u4fbf\u66f4\u597d\u7684\u8bad\u7ec3\uff0c\u5982\u4e0b\u6240\u793a\uff1a
\\(s(Z)=Z \\odot f_{W, b}(Z)\\)
\u4e3a\u4e86\u8bad\u7ec3\u66f4\u7a33\u5b9a\uff0c\u4f5c\u8005\u5c06 W \u548c b \u5206\u522b\u521d\u59cb\u5316\u4e3a\u63a5\u8fd1 0 \u4e0e 1 \u6765\u4fdd\u8bc1\u5728\u5f00\u59cb\u8bad\u7ec3\u65f6 \\(f_{W, b} \\approx 1\\) \u3001 \\(s(Z z) \\approx Z\\) \u4f7f\u5f97\u5728\u5f00\u59cb\u9636\u6bb5gMLP\u8fd1\u4f3c\u4e8eFFN\u5e76\u5728\u8bad\u7ec3\u4e2d\u9010\u6e10\u5b66\u4e60\u5230\u8de8token\u7684\u7a7a\u95f4\u4fe1\u606f\u3002
\u4f5c\u8005\u8fdb\u4e00\u6b65\u53d1\u73b0\u5c06 Z \u4ece\u901a\u9053\u7ef4\u5ea6\u5206\u5272\u6210\u4e24\u90e8\u5206 \\(\\left(Z_1, Z_2\\right)\\) \u8fdb\u884cgating\u64cd\u4f5c\u66f4\u6709\u7528\uff0c\u5982\u4e0b\u6240\u793a\uff1a
\\(
s(Z)=Z_1 \\odot f_{W, b}\\left(Z_2\\right)
\\)
\u53e6\u5916\u51fd\u6570 \\(f_{W, b}\\)\u7684\u8f93\u5165\u901a\u5e38\u9700\u8981normalizel\u4ee5\u6b64\u63d0\u5347\u6a21\u578b\u7684\u7a33\u5b9a\u6027\u3002<\/strong><\/p>\n\n\n\n\u4e00\u4e9b\u601d\u8003\uff1a\u8fd9\u91cc\u7684SpatialGatingUnit\u91cc\u9762\u7528\u5230\u4e86\u4e00\u4e2a\u901a\u9053split\uff0c\u7136\u540e\u518d\u5c06\u5206\u5272\u540e\u7684\u4e24\u90e8\u5206\u505a\u4e58\u6cd5\uff0c\u8ba9\u6211\u60f3\u5230\u4e86NAFnet\u4e2d\u7684simplegate\uff0c\u8fd9\u4e2a\u7684\u4f5c\u7528\u4e00\u662f\u51cf\u5c11\u8ba1\u7b97\u91cf\uff08\u76f8\u6bd4\u4e8eGELU\uff09\u3001\u53e6\u5916\u5f15\u5165\u95e8\u63a7\u673a\u5236\uff0c\u5728\u901a\u9053\u7ef4\u5ea6\u8fdb\u884c\u901a\u9053\u4ea4\u7ec7\uff0c\u5bf9\u4e8e\u6a21\u578b\u7684\u6548\u679c\u8868\u73b0\u5f88\u597d\u3002<\/p>\n\n\n\n
\u4f5c\u8005\u8fdb\u4e00\u6b65\u5206\u6790\u4e86SGU\u4e0e\u73b0\u6709\u7684\u4e00\u4e9b\u64cd\u4f5c\u7684\u76f8\u4f3c\u4e4b\u5904\uff1a\u9996\u5148\u662fGated Linear Units (GLU) \u4e0e SGU\u7684\u533a\u522b\u5728\u4e8eSGU\u5bf9spatial dimension\u800cGLU\u5bf9channel dimension; \u5176\u6b21SGU\u548c
Squeeze-and-Excite (SE) \u4e00\u6837\u4f7f\u7528hadamard-product\uff0c\u53ea\u662fSGU\u5e76\u6ca1\u6709\u8de8\u901a\u9053\u7684\u6620\u5c04\u6765\u4fdd \u8bc1\u6392\u5217\u4e0d\u53d8\u6027\uff1bSGU\u7684\u7a7a\u95f4\u6620\u5c04\u53ef\u4ee5\u770b\u4f5cdepthwise convolution\u4e0d\u8fc7SGU\u53ea\u5b66\u4e60\u8de8\u901a\u9053\u53ea\u662f\uff0c \u5e76\u6ca1\u6709\u8de8\u901a\u9053\u8fc7\u6ee4\u5668\uff1bSGU\u5b66\u4e60\u7684\u662f\u4e8c\u9636\u7a7a\u95f4\u4ea4\u4e92 \\(z_i z_j\\) \uff0c self-attention\u5b66\u4e60\u7684\u662f\u4e09\u9636\u4ea4\u4e92 \\(q_i k_j v_k\\) \uff0c SGU\u7684\u590d\u6742\u5ea6\u4e3a \\(n^2 e \/ 2\\) \u800cself-attention\u7684\u590d\u6742\u5ea6\u4e3a \\(2 n^2 d_{\\text {\u3002 }}\\)<\/p>\n\n\n\n
\u5b9e\u9a8c\uff1a<\/h2>\n\n\n\n
1\u3001Image Classification<\/p>\n\n\n\n
\u6b64\u6587\u9996\u5148\u5c06gMLP\u5e94\u7528\u4e8e\u56fe\u7247\u5206\u7c7b\uff0c\u4f7f\u7528ImageNet\u6570\u636e\u96c6\u800c\u4e14\u4e0d\u4f7f\u7528\u989d\u5916\u6570\u636e\u3002\u4e0b\u8868\u9996\u5148\u5c55\u793a\u4e86gMLP\u7528\u4e8e\u56fe\u7247\u5206\u7c7b\u7684\u53c2\u6570\uff0cgMLP\u548cViT\/B16\u4e00\u6837\u4f7f\u7528 16\u00d716 \u4e2apatch\uff0c\u540c\u65f6\u91c7\u7528\u548cDeiT\u76f8\u4f3c\u7684\u6b63\u5219\u5316\u65b9\u6cd5\u9632\u6b62\u8fc7\u62df\u5408\u3002<\/p>\n\n\n\n![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-220.png\")
<\/figure>\n\n\n\n\u4e0b\u8868\u4e2dgMLP\u4e0ebaselines\u5728ImageNet\u4e0a\u7684\u7ed3\u679c\u8868\u793agMLP\u53d6\u5f97\u4e86\u4e0e\u89c6\u89c9Transformer\u76f8\u5f53\u7684\u7ed3\u679c\uff0c\u540c\u65f6\u4e0e\u5176\u5b83MLP\u89c6\u89c9\u6a21\u578b\u76f8\u6bd4\uff0cgMLP\u53d6\u5f97\u4e86\u51c6\u786e\u7387\u3001\u901f\u5ea6\u6743\u8861\u4e0b\u6700\u597d\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-221.png\")
<\/figure>\n\n\n\nMasked Language Modeling with BERT:<\/h3>\n\n\n\n
\u6b64\u6587\u540c\u65f6\u5c06gMLP\u5e94\u7528\u4e8emasked language modeling\uff08MLM\uff09\u4efb\u52a1\uff0c\u5bf9\u4e8e\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u4efb\u52a1\uff0c\u6a21\u578b\u7684\u8f93\u5165\u8f93\u51fa\u89c4\u5219\u90fd\u4fdd\u6301\u4e0eBERT\u4e00\u81f4\u3002<\/p>\n\n\n\n
\u4f5c\u8005\u89c2\u5bdf\u5230\u5728MLM\u4efb\u52a1\u6700\u540e\u5b66\u4e60\u5230\u7684\u7a7a\u95f4\u6620\u5c04\u77e9\u9635\u603b\u662fToeplitz-like matrics\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u6240\u4ee5\u4f5c\u8005\u8ba4\u4e3agMLP\u662f\u80fd\u4ece\u6570\u636e\u4e2d\u5b66\u4e60\u5230\u5e73\u79fb\u4e0d\u53d8\u6027\u7684\u6982\u5ff5\u7684\uff0c\u8fd9\u4f7f\u5f97gMLP\u5b9e\u8d28\u8d77\u5230\u4e86\u5377\u79ef\u6838\u662f\u6574\u4e2a\u5e8f\u5217\u957f\u5ea6\u76841-d\u5377\u79ef\u7684\u4f5c\u7528\u3002\u5728\u63a5\u4e0b\u6765\u7684MLM\u5b9e\u9a8c\u4e2d\uff0c\u4f5c\u8005\u521d\u59cb W \u4e3aToeplitz matrix\u3002<\/p>\n\n\n\n![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-222.png\")
<\/figure>\n\n\n\nAblation: The Importance of Gating in gMLP for BERT\u2019s Pretraining\uff1a\u4e0b\u8868\u5c55\u793a\u4e86gMLP\u7684\u5404\u79cd\u53d8\u4f53\u4e0eTransoformer\u6a21\u578b\u3001MLP-Mixer\u7684\u6bd4\u8f83\uff0c\u53ef\u4ee5\u770b\u5230gMLP\u5728\u4e0eTransformer\u76f8\u540c\u6a21\u578b\u5927\u5c0f\u7684\u60c5\u51b5\u4e0b\u80fd\u8fbe\u5230\u4e0eTransformer\u76f8\u5f53\u7684\u6548\u679c\u3002\u540c\u65f6gating\u64cd\u4f5c\u5bf9\u4e8e\u7a7a\u95f4\u6620\u5c04\u5341\u5206\u6709\u7528\u3002\u540c\u65f6\u4e0b\u56fe\u8fd8\u53ef\u89c6\u5316\u4e86\u6a21\u578b\u5b66\u4e60\u5230\u7684\u7a7a\u95f4\u6620\u5c04\u53c2\u6570\u3002<\/p>\n\n\n\n![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-223.png\")
<\/figure>\n\n\n\nCase Study: The Behavior of gMLP as Model Size Increases\uff1a\u4e0b\u8868\u4e0e\u4e0b\u56fe\u5c55\u793a\u4e86gMLP\u968f\u7740\u6a21\u578b\u589e\u5927\u9010\u6e10\u80fd\u6709\u4e0eTransformer\u76f8\u5f53\u7684\u6548\u679c\uff0c\u53ef\u89c1Transformer\u7684\u6548\u679c\u5e94\u8be5\u4e3b\u8981\u662f\u4f9d\u8d56\u4e8e\u6a21\u578b\u5c3a\u5bf8\u800c\u975eself-attention\u3002<\/p>\n\n\n\n![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-224.png\")
<\/figure>\n\n\n\n![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-225.png\")
<\/figure>\n\n\n\n- Ablation: The Usefulness of Tiny Attention in BERT\u2019s Finetuning\uff1a\u4ece\u4e0a\u9762\u7684Case Study\u53ef\u4ee5\u53d1\u73b0gMLP\u5bf9\u4e8e\u9700\u8981\u8de8\u53e5\u5b50\u8fde\u63a5\u7684finetuing\u4efb\u52a1\u53ef\u80fd\u4e0d\u53caTransformer\uff0c\u6240\u4ee5\u4f5c\u8005\u63d0\u51fa\u4e86gMLP\u7684\u589e\u5f3a\u7248aMLP\u3002aMLP\u76f8\u8f83\u4e8egMLP\u4ec5\u589e\u52a0\u4e86\u4e00\u4e2a\u5355\u593464\u7684self-attention\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/li><\/ul>\n\n\n\n
![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-226.png\")
<\/figure>\n\n\n\n\u4ece\u4e0b\u56fe\u7ed3\u679c\u53ef\u4ee5\u53d1\u73b0aMLP\u76f8\u8f83\u4e8egMLP\u6781\u5927\u63d0\u5347\u4e86\u6548\u679c\u5e76\u5728\u6240\u6709task\u8d85\u8fc7\u4e86Transformer\u3002<\/p>\n\n\n\n![\"\"](\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-227.png\")
<\/figure>\n","protected":false},"excerpt":{"rendered":"MLP-Mixer\u7684\u589e\u5f3a\u7248\uff0c\u5e26gating\u7684MLP\u3002\u6709\u4e24\u4e2a\u7248\u672c\uff0c\u5206\u522b\u662fgMLP\u548caMLP\u3002Pay-Attent … \u7ee7\u7eed\u9605\u8bfbVision MLP –Pay Attention to MLPs<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[30,4,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/7981"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=7981"}],"version-history":[{"count":28,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/7981\/revisions"}],"predecessor-version":[{"id":9069,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/7981\/revisions\/9069"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=7981"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=7981"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=7981"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}