Deep learning classification network -- zfnet

Preface

ZFNet In this paper motivation There are two points ：

• Why does deep neural network perform well ？
• How to improve the network ？

The author passes Deconvnet To visualize the characteristic map of each layer of the deep neural network , And according to this, we can pertinently treat AlexNet Improvement , made 2013 year ImageNet The champion of the classification challenge .

1. Network structure

On the basis of visual feature map , The author of ALexNet Two changes have been made ：

• The size of the first convolution is determined by 11×11 Change it to 7×7
• The first convolution stride from 4 Change it to 2
  

Of course , The author's modification is not fabricated out of thin air , But there is evidence to rely on , This basis is the visualization result of the feature map .

2. The main points of

• Deconvnet Visualization methods
  

• Characteristics of each layer of the network
  
  

• Feature learning process of network layers
  

• Improve network structure according to feature visualization
  

• Feature invariance
  

• Occlusion sensitivity
  

Reference material

• ZFNet The paper ： Visualizing and Understanding Convolutional Networks
• “ Intuitive understanding ” Convolutional neural networks ( One )： deconvolution (Deconvnet)