Intelligent Integration of Enterprise Convolutional Neural Network Outline

智慧化企業整合 Intelligent Integration of Enterprise Convolutional Neural Network 助教: 陳可馨

Outline • Why CNN for image • Convolutional Neural Network üConvolution Layer üPooling Layer-Max Pooling üFlatten • Demo • Class Assignment & Homework

Why CNN for image? • Some patterns are much smaller than the whole image • The same patterns appear in different regions. • Subsampling the pixels will not change the object

Convolutional Neural Network

The Whole CNN Can repeat many times

Why CNN for image? • Some patterns are much smaller than the whole image Convolution • The same patterns appear in different regions. • Subsampling the pixels will not change the object Max Pooling

Convolution Layer network parameters to be learned. What we see What computer see

Convolution Layer Kernels/filters: each filter detects a small pattern (for example: 3 x 3).

Convolution v. s. Fully connected 1 0 0 1 1 -1 -1 -1 0 1 0 -1 1 -1 0 0 1 1 0 0 -1 -1 1 0 0 0 1 0 1 0 convolution image 0 0 0 1 0 0 0 1 1 0 0 0 1 0 0 0 1 0 … Fullyconnected 1

1 -1 -1 -1 1 1: 1 Filter 1 2: 0 3: 0 4: 0 0 0 1 0 0 1 1 0 0 6 x 6 image Less parameters! 7: 0 8: 1 9: 0 10: 0 … 0 1 0 0 … 1 0 0 1 3 13: 0 14: 0 15: 1 16: 1 Only connect to 9 input, not fully connected … Shared weights

Pooling Layer-Max Pooling 1 0 0 1 0 0 0 1 0 1 1 0 0 0 1 1 0 0 6 x 6 image New image but smaller Conv Max Pooling 3 -1 0 3 1 0 1 3 2 x 2 image Each filter is a channel

Flatten 3 3 0 30 1 -1 0 1 3 1 Flatten 3 -1 1 0 3 Fully Connected Feedforward network

Demo

References • http: //speech. ee. ntu. edu. tw/~tlkagk/courses/ML_2016/ Lecture/CNN%20(v 2). pdf • https: //medium. com/jameslearningnote/%E 8%B 3%87 %E 6%96%99%E 5%88%86%E 6%9 E%90%E 6%A 9%9 F%E 5%99%A 8%E 5%AD%B 8%E 7%BF%9 2 -%E 7%AC%AC 5 -1%E 8%AC%9 B%E 5%8 D%B 7%E 7%A 9%8 D%E 7%A 5%9 E%E 7%B 6%9 3%E 7%B 6%B 2%E 7%B 5%A 1%E 4%BB%8 B%E 7%B 4% B 9 -convolutional-neural-network-4 f 8249 d 65 d 4 f

Class assignment • Please train a CNN with 3 Conv 2 D layers and 2 Maxpooling 2 D layers( try different activations, e. g. sigmoid, relu, etc. ). to predict the class of input images in Fashion Mnist dataset, and the testing accuracy should be at least 95%. • Turn in your work with the format of. ipynb , and please write some brief comments in your ipynb to illustrate your results. • File name: class 6_Your Chinese Name 15

Homework • Please use the Cifar-10 dataset and what we taught in TA class to train a CNN model ( you may design your own CNN model), and the testing accuracy should be at least 60%. • You are encouraged to implement different methods to train your model. (EX: dropout or different optimizers) • Turn in your work with the format of. ipynb , and please write some brief comments in your ipynb to illustrate your results. • File name: hw 6_Your Chinese Name 16