An Information Analysis Approach into Feature Understanding of Convolutional Deep Neural Networks

摘要

This paper is centered on feature analysis and visualization of pre-trained deep neural networks based on responses of neurons to input images. In order to address this problem, first the information content of learned encodings of neurons is investigated based on the calculation of the salient activation map of each neuron. The salient activation map is considered to be the activation map that has the highest aggregative value over all its cells. Second, neurons are visualized based on their maximum activation response at each location. The results put forward the uncertainty reduction over the stage of deeper layers as well as a decrement pattern in Variation of Information.