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?

为什么CNN 在imagenet上取得如何显著的结果？可以归结为下面三个当面：

1、larger training sets;  millions of labeled examples. 

2、GPU makes practical

3、Better model regularization strategies: Dropout

输入：2D images;  输出：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 through a rectified linear function (relu(x) = max(x,0));   Ensure the feature maps are always positive.

(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.