Visualization of Convolutional Neural Networks Based on Gaussian Models

Convolutional neural networks (CNN) have made breakthrough progress in tasks such as image classification, target detection and scene recognition. The trained model has excellent automatic feature extraction and prediction performance, and can provide users with end-to-end “input-output” form Solution. However, due to distributed feature encoding and increasingly complex model structures, people have never been able to accurately understand the internal knowledge representation of the convolutional neural network model and the underlying reasons that prompted it to make specific decisions. We can find that the output and input of CNN with an appropriate prior over the weights and biases is a Gaussian process (GP). For the CNN, it is a signal Gaussian model. So, we can use single Gaussian model to explain the computational process of CNN. We show the GP distribution and computational process over the weights and biases. The experiment results show that this analysis method can give a good visualization effective.