《智能控制智能控制智能控制 (37).pdf》由会员分享,可在线阅读,更多相关《智能控制智能控制智能控制 (37).pdf(16页珍藏版)》请在得力文库 - 分享文档赚钱的网站上搜索。
1、Convolutional Neural NetworksIntroduction Family of CNN:Atable of developmentIntroduction Family of CNN:Accuracy plexityInput Xkernel KOutput Y22 sizesNetwork Layers Convolution:Kernel K is learnable.00+11+32+43=1933 sizes1-1-1-11-1-1-11?Network Layers Convolution:Kernel Stride1-1-1-11-1-1-11same ke
2、rnel Network Layers Convolution:Kernel Stride Padding adds rows/columns around inputa)p=1b)p=2 Size of the output:Network Layers Convolution:Kernel Stride Padding Channels Color image may have three RGB channels Assign different kernel for each channelNetwork Layers Pooling layer Types Max pooling A
3、verage pooling Affect Compact feature representation Reducing processing costNetwork Layers Fully Connected layer The same as a traditional MLP Used for classification1x2x36x100001010010001100100010010010001010Loss Function Evaluating model performance during training Gradual improving due to optimi
4、zer Multiple loss functions for one model possible(one for each output variable)Category of Loss Function Binary Classification Binary Cross-Entropy loss SVM hinge loss Squared hinge loss?=?+1?1?Category of Loss Function Multi-class Classification Cross-entropy loss(Softmax loss)Expectation loss?(?,
5、?)=?log(?),?1,?An example:LeNetY.LeCun,L.Bottou,Y.Bengio,P.Haffner,1998Gradient-basedlearning applied todocument recognition Handwritten Digit Recognition MNIST Dataset 50,000 training data 10,000 test data 28 x 28 images 10 classesThe Architecture of LeNet Convolutional layer:learn space information Fully connected layer:convert to the category spaceExpensive if we have many outputs.Results Basic Layers:Convolution/Pooling/Fully Connected layer Loss Function for different tasks Asimple CNN architectureSummary