Visualization Analysis of Convolutional Neural Network Processes

Convolutional neural network visualization can help deep learning researchers better understand the concepts and principles of convolution. However, the existed analysis methods have some shortness, such as input and internal features of the model. Through visualization, researchers can have a clearer view of each step in convolution operations, including input, convolution kernel, convolution operation, and output, thereby better understanding the essence and function of convolution is needed. So, we propose a visualization analysis methods based on the random meaning perturbation (VARMP). The proposed method adds the random meaning perturbation to input image, then we can find the minimum deleting mask. From the experiment results, we can find that the visualization analysis results show the real train process of convolutional neural network.