Visualizing and Understanding convolutional networks
Large convolutional networks model on ImageNet
(Krizhevsky, A., Sutskever, I. and Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks , Advances in Neural Information Processing 25, 2012)
Why they perform so well?
How they migh be improved?
為什麼CNN 在imagenet上取得如何顯著的結果?可以歸結為下面三個當面:
1、larger training sets; millions of labeled examples.
2、GPU makes practical
3、Better model regularization strategies: Dropout
輸入:2D images; 輸出:a probability vector over the C different classes.
Convolutional layers 包含:
(1) Convolution of the previous layer output with a set of learned filters
(2) Passing the responses
(3) Max-pooling over local neighborhoods (optionally)
(4) Normalize the responses across feature maps.(optionally)
Top few layers of NN: Fully-connected networks.
Final layer: a softmax classifier.
Visualization with a Deconvnet by mapping these activities back to the input pixel space.