{"id":8574,"date":"2016-10-11T23:52:55","date_gmt":"2016-10-11T15:52:55","guid":{"rendered":"https:\/\/ihower.tw\/blog\/?p=8574"},"modified":"2024-04-14T00:55:47","modified_gmt":"2024-04-13T16:55:47","slug":"%e4%b8%80%e5%a4%a9%e6%90%9e%e6%87%82%e6%b7%b1%e5%ba%a6%e5%ad%b8%e7%bf%92-%e5%bf%83%e5%be%97%e7%ad%86%e8%a8%98","status":"publish","type":"post","link":"https:\/\/ihower.tw\/blog\/8574-%e4%b8%80%e5%a4%a9%e6%90%9e%e6%87%82%e6%b7%b1%e5%ba%a6%e5%ad%b8%e7%bf%92-%e5%bf%83%e5%be%97%e7%ad%86%e8%a8%98","title":{"rendered":"\u53f0\u7063\u8cc7\u6599\u79d1\u5b78\u611b\u597d\u8005\u5e74\u6703: \u4e00\u5929\u641e\u61c2\u6df1\u5ea6\u5b78\u7fd2 \u5fc3\u5f97\u7b46\u8a18"},"content":{"rendered":"<p>2016\/9\/24 \u53bb\u4e2d\u7814\u9662\u53c3\u52a0\u674e\u5b8f\u6bc5\u8001\u5e2b\u7684<a href=\"http:\/\/datasci.tw\/event\/deep_learning_one_day\/\">\u4e00\u5929\u641e\u61c2\u6df1\u5ea6\u5b78\u7fd2<\/a>\u8ab2\u7a0b\uff0c\u6536\u7a6b\u883b\u591a\u7684\uff0c\u6295\u5f71\u7247<a href=\"http:\/\/www.slideshare.net\/tw_dsconf\/ss-62245351\">\u5728\u9019\u88e1<\/a>\u3002\u5c0d\u6211\u4f86\u8aaa Part 2 \u6536\u7a6b\u6700\u591a\uff0cPart 3 \u662f\u6211\u7b2c\u4e00\u6b21\u958b\u59cb\u4e86\u89e3 CNN \u548c RNN\uff0c\u4e00\u6642\u7121\u6cd5\u5b8c\u5168\u7406\u89e3\u5176\u904b\u7b97\u904e\u7a0b\uff0c\u5927\u6982\u4e86\u89e3\u539f\u7406\u800c\u5df2\u3002\u4ee5\u4e0b\u662f\u4e00\u4e9b\u7b46\u8a18\uff1a<\/p>\n<p><!--more--><\/p>\n<h2 id=\"toc_0\">Part1: DNN \u7c21\u4ecb<\/h2>\n<p>\u56e0\u70ba\u4fee\u904e\u53f0\u5927<a href=\"http:\/\/www.csie.ntu.edu.tw\/~htlin\/mooc\/\">\u6797\u8ed2\u7530\u8001\u5e2b<\/a>\u7684\u6a5f\u5668\u5b78\u7fd2\u8ab2\u7684<a href=\"https:\/\/www.coursera.org\/account\/accomplishments\/certificate\/24W587BA37\">\u80cc<\/a><a href=\"https:\/\/www.coursera.org\/account\/accomplishments\/certificate\/LMCNF56QSS\">\u666f<\/a>\uff0c\u7c21\u4ecb\u7684\u5167\u5bb9\u8f15\u9b06\u807d\u3002\u4e3b\u5c31\u662f\u4ecb\u7d39\u6a5f\u5668\u5b78\u7fd2\u7684\u904e\u7a0b\u3001\u4ec0\u9ebc\u662f\u985e\u795e\u7d93\u7db2\u8def\u548c Gradient Descent\u3002\u4f46\u6709\u9ede\u9a5a\u8a1d\u7684\u662f\uff0c\u8001\u5e2b\u5b8c\u5168\u4e0d\u8457\u58a8\u6559\u5982\u4f55\u8a08\u7b97 Backpropagation\uff0c\u4ed6\u8aaa\u592a\u591a toolkits \u73fe\u6210\u5de5\u5177\u53ef\u4ee5\u8655\u7406\u4e86\uff0c\u53ef\u4ee5\u7576\u5b83\u662f\u500b\u9ed1\u7bb1\u5b50\uff0c\u54c8\u54c8\u3002<\/p>\n<p>\u53e6\u5916\u9084\u6709\u63a2\u8a0e\u4e86\u4e00\u4e0b\u70ba\u4ec0\u9ebc\u985e\u795e\u7d93\u7db2\u8def\u7684\u767c\u5c55\u662f Deep \u800c\u4e0d\u662f Fat\uff0c\u9019\u5169\u7a2e\u7684\u795e\u7d93\u5143\u53ef\u4ee5\u4e00\u6a23\u591a\u554a\u3002\u7d93\u904e\u5be6\u9a57 Fat+Short v.s. Thin + Tail \u53ef\u4ee5\u767c\u73fe\u5f8c\u8005\u7684\u7d50\u679c\u6bd4\u8f03\u597d\u3002\u8001\u5e2b\u6c92\u6709\u7d66\u8b49\u660e\uff0c\u800c\u662f\u7d66\u4e86\u4e00\u500b\u76f4\u89c0\u7684\u985e\u6bd4\uff1a\u5c31\u50cf Logic circuits (\u679c\u7136\u662f\u96fb\u6a5f\u7cfb\u7684\u8001\u5e2b)\uff0c\u7528\u591a\u5c64\u7684\u908f\u8f2f\u9598\u53ef\u4ee5\u6709\u6548\u7387\u7684\u7d44\u51fa\u8907\u96dc function\u3002<\/p>\n<h2 id=\"toc_1\">Part2: Recipe of Deep Learning<\/h2>\n<ul>\n<li>\u8001\u5e2b\u7528 <a href=\"https:\/\/keras.io\/\">Keras<\/a> \u4f5c\u73fe\u5834\u793a\u7bc4\uff0c\u9019\u662f\u4e00\u500b TensorFlow \u548c theano \u7684\u5305\u88dd\u51fd\u5f0f\u5eab\uff0cAPI \u66f4\u597d\u7528\u3002\u5982\u679c\u96fb\u8166\u6709 Nvidia \u986f\u793a\u5361\uff0c\u53ef\u4ee5\u7528 GPU \u5e73\u884c\u904b\u7b97\u52a0\u901f\uff0c\u8d85\u5febder~~~(\u7fa8\u6155)<\/li>\n<li>\u547c\u53eb Keras API \u5f88\u7c21\u55ae\uff0c\u4f46\u662f\u5982\u679c\u4e0d\u6703\u8abf\u53c3\u6578\uff0c\u99ac\u4e0a\u8b93\u4f60\u5f9e\u5165\u9580\u5230\u653e\u68c4\u3002<\/li>\n<li>\u6df1\u5ea6\u5b78\u7fd2\u7684 Hello World \u5c31\u662f\u62ff <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\">MNIST \u8cc7\u6599\u96c6<\/a>\u505a\u624b\u5beb\u8fa8\u8b58<\/li>\n<li>\u4e0d\u6703\u8abf\u53c3\u6578\u662f\u6e96\u78ba\u7387\u53ea\u6709 11%\uff0c\u6703\u7c21\u55ae\u8abf\u53c3\u6578\u99ac\u4e0a\u8b8a 80%\uff0c\u5b78\u6703\u5f8c\u53ef\u4ee5\u63d0\u5347\u5230 97%<\/li>\n<li>\u8abf\u53c3\u6578\u8981\u5148\u5224\u65b7\u662f\u4e0d\u662f\u5f9e training set \u5c31\u6c92\u8a13\u7df4\u597d\uff0c\u9084\u662f\u53ea\u6709 test set \u4e0d\u597d\u3002\n<ul>\n<li>\u52a0\u591a\u53c3\u6578\uff0c\u6548\u80fd\u8b8a\u721b\uff0c\u4e0d\u4e00\u5b9a\u662f overfitting\uff0c\u8981\u770b\u662f\u4e0d\u662f training set \u7d50\u679c\u5c31\u4e0d\u597d\u3002\u4f8b\u5982 dropout \u6280\u5de7\u662f\u91dd\u5c0d overfiiting \u554f\u984c\u8655\u7406\uff0c\u4f60\u4e0d\u80fd\u62ff\u9019\u62db\u53bb\u8655\u7406 training set \u554f\u984c\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h3 id=\"toc_2\">\u4ee5\u4e0b\u8655\u7406 training data \u6e96\u78ba\u7387\u554f\u984c<\/h3>\n<ul>\n<li>\u9996\u5148\u662f loss function \u9078\u64c7\n<ul>\n<li>\u5728 softmaxoutput layer\uff0c\u53ef\u4ee5\u9078\u64c7\u7528 cross entropy \u4f86\u8655\u7406\u5206\u985e\u554f\u984c<\/li>\n<li>\u56e0\u70ba entropy \u7684\u5fae\u5206\u5c71\u5761\u6bd4\u8f03\u9661\u5ced<\/li>\n<li>\u4e0d\u8981\u7528 mse \u5c31\u5dee\u5f88\u591a\uff0c\u99ac\u4e0a\u6e96\u78ba\u7387\u5f9e 11% \u5230 8X%<\/li>\n<\/ul>\n<\/li>\n<li>Mini-batch\n<ul>\n<li>\u6709\u5169\u500b\u53c3\u6578 batch size \u8ddf\u8fed\u4ee3\u6b21\u6578<\/li>\n<li>batch size \u8abf\u5c0f\u6703\u6bd4\u8f03\u5feb\uff0c\u4f46\u662f\u5e73\u884c\u904b\u7b97\u4e5f\u53ef\u4ee5\u8b93 batch \u5f88\u5927\u4e5f\u53ef\u4ee5\u5f88\u5feb<\/li>\n<li>\u4f46\u662f mini-batch \u8b93\u7d50\u679c\u4e5f\u6bd4\u8f03\u597d\uff0c\u56e0\u70ba\u96a8\u6a5f\u6027\u907f\u514d\u4e86 local optimal \u554f\u984c<\/li>\n<\/ul>\n<\/li>\n<li>New activation function\n<ul>\n<li>\u589e\u52a0\u66f4\u591a\u5c64\u4f46\u662f\u7d50\u679c\u537b\u8b8a\u721b: Vanishing gradient problem<\/li>\n<li>\u9019\u662f sigmoid \u7684 activation function \u554f\u984c\uff0c\u9760 input layer \u90a3\u908a\u7684\u5fae\u5206\u503c\u592a\u5c0f\u4e86<\/li>\n<li>2006 \u6642\u7528 RBM pre-training\uff0c\u73fe\u5728\u4e0d\u6d41\u884c<\/li>\n<li>2015 \u6642\u63db activation function\uff0c\u6539\u7528 ReLU \u6703\u8b93\u6bcf\u5c64\u8b8a\u7626!<\/li>\n<li>\u53ef\u4ee5\u4e0d\u7528\u64d4\u5fc3\u9700\u8981\u50cf sigmoid \u9650\u5236\u6578\u5b57\u8b8a\u6210 0~1\uff0c\u56e0\u70ba DNN \u672c\u8eab\u5c31\u53ef\u4ee5\u8655\u7406\u6578\u5b57\u7bc4\u570d\u5f88\u5927\u5f88\u5c0f<\/li>\n<li>\u5230\u9019\u908a\uff0c\u6e96\u78ba\u7387\u8b8a 96% \u4e86!<\/li>\n<li>Maxout \u53ef\u8b93\u6bcf\u500b activation function \u90fd\u4e0d\u4e00\u6a23\uff0c\u8b8a\u6210 learnable\u3002ReLU \u662f Maxout \u7684\u4e00\u7a2e\u7279\u4f8b\u3002<\/li>\n<\/ul>\n<\/li>\n<li>Adaptive Learning Rate\n<ul>\n<li>Learning Rate \u5982\u679c\u56fa\u5b9a\uff0c\u4e00\u958b\u59cb\u6703\u8d70\u592a\u6162\uff0c\u6216\u6700\u5f8c\u8d70\u592a\u5feb\u3002<\/li>\n<li>\u61c9\u8a72\u8981\u96a8\u8457\u6642\u9593\uff0c\u8b93 lr \u8d8a\u4f86\u8d8a\u5c0f\u3002\u5fae\u5206\u8d8a\u5c0f\uff0clr \u5927\u4e00\u9ede\u3002<\/li>\n<li>Adagrad \u65b9\u6cd5\u548c\u5176\u4ed6\u4e00\u6253\u7684\u65b9\u6cd5\u53ef\u4ee5\u9078\u7528<\/li>\n<\/ul>\n<\/li>\n<li>Momentum\n<ul>\n<li>\u9664\u4e86 local optimal \u554f\u984c\uff0c\u5fae\u5206\u8d70\u5230 saddle point \u4e5f\u6703\u505c\u4e0b\u4f86\uff0c\u56e0\u70ba\u5fae\u5206\u975e\u5e38\u5c0f\uff0c\u8655\u5728\u9ad8\u539f\u7684\u5730\u65b9&#8230;&#8230; \u4e5f\u6703\u505c\u4e0b\u4f86\u3002<\/li>\n<li>\u89e3\u6c7a\u7684\u60f3\u6cd5\u662f\u5c31\u50cf\u7269\u7406\u7279\u6027\u6709\u6163\u6027\uff0c\u6709\u67d0\u7a2e\u6163\u6027\u5728\uff0c\u6240\u4ee5\u5373\u4f7f\u5fae\u5206=0\uff0c\u9084\u662f\u7e7c\u7e8c\u8abf\u6574\u53c3\u6578\uff0c\u53ef\u4ee5\u5728\u5e73\u539f\u4e5f\u7e7c\u7e8c\u8d70\uff0c\u9003\u51fa local optim\uff0c\u751a\u81f3\u722c\u904e\u5c0f\u5c71\u5761<\/li>\n<li>optimizer \u628a SGD \u6539\u6210\u7528 Adam \u6cd5\uff0c\u53ef\u4ee5\u8b93\u53c3\u6578\u66f4\u65b0\u901f\u5ea6\u66f4\u5feb\uff0c\u6700\u5f8c\u7d50\u679c\u597d\u4e00\u9ede \u8b8a 97%<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h3 id=\"toc_3\">\u4ee5\u4e0b\u8655\u7406 overfitting \u554f\u984c<\/h3>\n<ul>\n<li>\u842c\u80fd\u62db\u6578: \u589e\u52a0\u66f4\u591a training data\n<ul>\n<li>\u6c92\u6709\u600e\u9ebc\u8fa6? \u81ea\u5df1\u589e\u52a0: \u4f8b\u5982\u589e\u52a0\u566a\u97f3\uff0c\u6216\u662f\u4f4d\u79fb\u8cc7\u6599\u7b49\u7b49<\/li>\n<\/ul>\n<\/li>\n<li>Early Stopping\n<ul>\n<li>\u900f\u904e validation test \u7576 lost \u589e\u52a0\u6642\uff0c\u5c31\u505c\u6b62 training<\/li>\n<\/ul>\n<\/li>\n<li>Regularization\n<ul>\n<li>Weight decay is one kind of regularization<\/li>\n<li>\u6bcf\u6b21 weight \u4e58 0.99\uff0c\u8b93\u4ed6\u840e\u7e2e<\/li>\n<\/ul>\n<\/li>\n<li>Dropout\n<ul>\n<li>\u9019\u662f deep learning \u4e2d\u7279\u6709\u7684 regularization \u65b9\u5f0f<\/li>\n<li>training \u6642\u6bcf\u6b21 mini-batch \u96a8\u6a5f\u4e1f\u68c4\u795e\u7d93\u5143! \u8b93\u4ed6\u8b8a\u7626\u9577<\/li>\n<li>\u6ce8\u610f testing \u6642\u4e0d\u8981 dropout<\/li>\n<li>\u539f\u7406: Dropout is a kind of ensemble.<\/li>\n<li>\u6ce8\u610f: \u52a0\u4e86 dropout\uff0c\u4f60\u7684 training set \u7d50\u679c\u5176\u5be6\u6703\u8b8a\u5dee\uff0c\u4f46\u662f test set \u6703\u8b8a\u597d<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h2 id=\"toc_4\">Part3: Variants of Neural Networks \u5404\u7a2e\u4e0d\u540c\u7684\u7d50\u69cb<\/h2>\n<h1 id=\"toc_5\"><\/h1>\n<h3 id=\"toc_6\">Convolutional Neural Network (CNN)<\/h3>\n<ul>\n<li>\u9069\u5408\u7528\u5728\u5f71\u50cf\u4e0a<\/li>\n<li>\u56e0\u70ba fullly-connected netowkring \u5982\u679c\u7528\u5728\u5f71\u50cf\u8fa8\u8b58\u4e0a\uff0c\u6703\u5c0e\u81f4\u53c3\u6578\u904e\u591a(\u56e0\u70ba\u50cf\u7d20\u5f88\u591a)\uff0c\u5c0e\u81f4 over-fitting<\/li>\n<li>CNN \u91dd\u5c0d\u5f71\u50cf\u8fa8\u8b58\u7684\u7279\u6027\uff0c\u7279\u5225\u8a2d\u8a08\u904e\uff0c\u4f86\u6e1b\u5c11\u53c3\u6578:<\/li>\n<li>Convolution: \u5b78\u51fa filter \u6bd4\u5c0d\u539f\u59cb\u5716\u7247\uff0c\u7522\u751f\u51fa feature map (\u4e5f\u7576\u6210image)<\/li>\n<li>Max Pooling: \u5c07 feature map \u7e2e\u5c0f<\/li>\n<li>Flatten: \u5c07\u6bcf\u500b\u50cf\u7d20\u7684 channels (\u6709\u591a\u5c11\u500bfilters) \u5c55\u958b\u6210 fully connected feedforward network<\/li>\n<li>\u4f46\u662f CNN \u4e0d\u6703\u8655\u7406\u65cb\u8f49<\/li>\n<li>AlphaGo \u4e5f\u7528\u4e86 CNN\uff0c\u4f46\u662f\u6c92\u6709\u7528 Max Pooling (\u6240\u4ee5\u4e0d\u540c\u554f\u984c\u9700\u8981\u4e0d\u540cmodel)<\/li>\n<li>\u9078\u64c7\u8a2d\u8a08 feature engineering \u6216\u662f \u9078\u64c7\u8a2d\u8a08 DNN \u7d50\u69cb\uff0c\u5176\u5be6\u662f\u540c\u4e00\u56de\u4e8b\u3002\u56e0\u70ba\u5982\u679c\u4eba\u53ef\u4ee5\u8a2d\u8a08\u51fa\u4e0d\u932f\u7684 feature\uff0c\u90a3\u9ebc DNN \u5e6b\u52a9\u5c31\u4e0d\u5927\u4e86\u3002<\/li>\n<li>\u4f8b\u5982\u5f71\u50cf\u8fa8\u8b58\u7528 DNN \u53ef\u4ee5\u6539\u9032\u5f88\u5927\uff0c\u4f46\u662f\u6587\u5b57\u76f8\u95dc\u5c31\u6539\u9032\u4e0d\u5927\uff0c\u56e0\u70ba\u672c\u4f86\u7684\u6587\u5b57\u8655\u7406 feature engineering \u5c31\u5f88\u4e0d\u932f\u4e86\u3002<\/li>\n<\/ul>\n<h3 id=\"toc_7\">Recurrent Neural Network (RNN)<\/h3>\n<ul>\n<li>\u89e3\u6c7a Slot Filling \u554f\u984c\uff0c\u8655\u7406\u6587\u5b57\u627e\u95dc\u9375\u5b57\uff0c\u4f46\u662f\u6709\u8a18\u61b6\u529b<\/li>\n<li>\u9069\u7528\u65bc\u8655\u7406\u6642\u9593\u3001\u7a7a\u9593\u5e8f\u5217\u4e0a\u6709\u5f37\u95dc\u806f\u7684\u8a0a\u606f<\/li>\n<li>\u6709\u500b\u6301\u7e8c\u8a18\u61b6\u7684\u53c3\u6578\uff0c\u6bcf\u6b21\u904b\u7b97\u6703\u8003\u616e\u9032\u4f86\uff0c\u901a\u5e38\u5c31\u662f\u7528 Long Short-term Memory (LSTM)\uff0c\u5176\u4ed6\u9084\u6709 GRU (\u7c21\u55ae\u7248)\u3001 SimpleRNN (\u66f4\u7c21\u55ae\u7248?)<\/li>\n<li>RNN \u4e0d\u597d train\uff0c\u53ef\u80fd\u7a81\u7136 loss \u7a81\u9ad8\uff0c\u9019\u4e0d\u662f bug&#8230;..\n<ul>\n<li>\u6709\u4e00\u4e9b\u89e3\u6cd5\uff0c\u4f8b\u5982\u7528 LSTM \u9632\u6b62 gradient vanishing<\/li>\n<\/ul>\n<\/li>\n<li>\u5404\u7a2e\u61c9\u7528\u4ecb\u7d39:\n<ul>\n<li>\u4e00\u5c0d\u591a: \u89e3\u6790\u6587\u5b57\uff0c\u5f9e\u6587\u5b57 \u7522\u751f \u5206\u985e<\/li>\n<li>\u591a\u5c0d\u591a: \u8a9e\u97f3\u8fa8\u8b58\uff0c\u5f9e \u8a9e\u97f3 \u7522\u751f \u6587\u5b57<\/li>\n<li>\u8a9e\u8a00\u7ffb\u8b6f<\/li>\n<li>\u8f38\u5165\u5f71\u50cf\uff0c\u8f38\u51fa\u5167\u5bb9\u76f8\u95dc\u7684\u6587\u5b57 (Video Caption Generation)<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h2 id=\"toc_8\">Part4: Next wave \u65b0\u6f6e\u6d41<\/h2>\n<ul>\n<li>Supervised learning\n<ul>\n<li>\u65b0\u7d50\u69cb: Ultra Deep Network !! \u4f46\u662f\u9700\u8981\u7279\u5225\u8a2d\u8a08 shortcut \u907f\u514d over-fitting \u8b93 model \u58de\u6389<\/li>\n<li>Attention-based model: \u6839\u64da\u554f\u984c\uff0c\u53bb\u95b1\u8b80\u6587\u5b57\u3001\u5f71\u50cf\u3001\u8072\u97f3\u627e\u5230\u95dc\u9375(attention)\u7684\u5730\u65b9\u56de\u7b54\u554f\u984c<\/li>\n<\/ul>\n<\/li>\n<li>Reinforcement learning\n<ul>\n<li>\u73a9\u904a\u6232\u7b49\u5f88\u591a\u61c9\u7528<\/li>\n<li>\u539f\u7406\u6642\u9593\u4e0d\u5920\u8b1b\uff0c\u8acb\u770b\u53c3\u8003\u8cc7\u6599<\/li>\n<\/ul>\n<\/li>\n<li>Unsupervised learning\n<ul>\n<li>\u6a21\u4eff\u756b\u98a8<\/li>\n<li>Pixel Recurrent Neural Networks \u81ea\u52d5\u88dc\u5716<\/li>\n<li>auto-encoder: \u8a13\u7df4\u4e00\u500b\u6a21\u578b\uff0c\u8b93\u8f38\u5165\u548c\u8f38\u51fa\u4e00\u6a23\uff0c\u4f46\u6545\u610f\u8b93\u904e\u7a0b\u4e2d dim \u964d\u4f4e\uff0c\u8b93DNN\u8b8a\u7a84\u3002\u5f9e\u4e2d\u5f97\u5230 encoder \u548c decoder\u3002<\/li>\n<li>\u6a5f\u5668\u7522\u751f\u5716\u7247\uff0c\u4f8b\u5982\u6a5f\u5668\u81ea\u52d5<a href=\"https:\/\/github.com\/mattya\/chainer-DCGAN\">\u756b\u6f2b\u756b<\/a><\/li>\n<li>\u6a5f\u5668\u95b1\u8b80: Generating Word Vector\/Embedding is unsupervised \u627e\u51fa\u8a5e\u5f59\u6db5\u7fa9\u900f\u904e \u627e\u51fa\u610f\u7fa9\u76f8\u8fd1\u7684\u8a5e\u5f59\u3001\u627e\u51fa\u5c0d\u6bd4\u63a8\u8ad6<\/li>\n<li>\u6a5f\u5668\u807d\u8072\u97f3 (WaveNet DeepMind)[https:\/\/deepmind.com\/blog\/wavenet-generative-model-raw-audio\/]<\/li>\n<li>\u81ea\u52d5\u8aaa\u8a71\u548c\u7522\u751f\u97f3\u6a02<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h2 id=\"toc_9\">\u7d50\u8a9e<\/h2>\n<ul>\n<li>\u65b0\u5de5\u4f5c: AI \u8a13\u7df4\u5e2b\n<ul>\n<li>\u9700\u8981\u8a13\u7df4\u5e2b\u9078\u64c7 model, \u4e0d\u540c model \u9069\u5408\u8655\u7406\u4e0d\u540c\u554f\u984c<\/li>\n<li>\u9700\u8981\u8db3\u5920\u7684\u7d93\u9a57\uff0c\u624d\u80fd\u99d5\u99ad best model\uff0c\u7279\u5225\u662f\u50cf DNN \u9019\u9ebc powerful \u7684 model \u554a<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>2016\/9\/24 \u53bb\u4e2d\u7814\u9662\u53c3\u52a0\u674e\u5b8f\u6bc5\u8001\u5e2b\u7684\u4e00\u5929\u641e\u61c2\u6df1\u5ea6\u5b78\u7fd2\u8ab2\u7a0b\uff0c\u6536\u7a6b\u883b\u591a\u7684\uff0c\u6295\u5f71\u7247\u5728\u9019\u88e1\u3002\u5c0d\u6211\u4f86\u8aaa Part &hellip; <\/p>\n<p class=\"link-more\"><a href=\"https:\/\/ihower.tw\/blog\/8574-%e4%b8%80%e5%a4%a9%e6%90%9e%e6%87%82%e6%b7%b1%e5%ba%a6%e5%ad%b8%e7%bf%92-%e5%bf%83%e5%be%97%e7%ad%86%e8%a8%98\" class=\"more-link\">\u95b1\u8b80\u5168\u6587<span class=\"screen-reader-text\">\u3008\u53f0\u7063\u8cc7\u6599\u79d1\u5b78\u611b\u597d\u8005\u5e74\u6703: \u4e00\u5929\u641e\u61c2\u6df1\u5ea6\u5b78\u7fd2 \u5fc3\u5f97\u7b46\u8a18\u3009<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":false,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[74],"tags":[],"class_list":["post-8574","post","type-post","status-publish","format-standard","hentry","category-data-science","entry"],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","jetpack_shortlink":"https:\/\/wp.me\/p1q6tG-2ei","jetpack_sharing_enabled":true,"jetpack_likes_enabled":true,"_links":{"self":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/8574","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/comments?post=8574"}],"version-history":[{"count":19,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/8574\/revisions"}],"predecessor-version":[{"id":11995,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/8574\/revisions\/11995"}],"wp:attachment":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/media?parent=8574"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/categories?post=8574"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/tags?post=8574"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}