Research And Application Of Differential Visual Analysis For Convolutional Neural Network

As an important branch of neural network model,convolutional neural network is widely used in image recognition,automatic driving,cancer detection and semantic analysis because of its good performance.However,because there is no perfect explanation for the internal operation mechanism of the network at present,users generally regard the model as a "black box" when using the neural network for data analysis,which will lead to the difficulty of users to modify the model to achieve satisfactory results when the actual demand changes.One of the commonly used methods to improve the performance of the model is to modify the relevant parameters of the model.This method is simple and feasible.However,there is no perfect explanation about the influence of parameters on the network at present,so the modification of parameters is mostly in a "trial-and-error" state.In order to solve above problems,this paper designs and implements a visual analysis system of convolutional neural network DiffCNNVis based on difference analysis.The purpose of this is to help users understand how various parameters of convolutional network affect the network.The core idea of this paper is to compare the internal information of two convolutional neural networks with different parameters through a single variable experiment,and visualize the corresponding information differences to reflect the way the parameters act on the network.This paper designs a benchmark network for comparison,and modifies parameters such as activation function,convolution layer structure,learning rate and discarding rate on the basis of the network,so as to obtain the experimental control network,and then extracts the internal features of the network such as convolution kernel tensor and convolution layer output characteristics in the experiment.Then,according to the user's needs,we design and implement the whole system and the difference visualization components by using B/S architecture.According to the principle of overview before exploration,the visual components are divided into two parts: the difference overview component and the difference exploration component.Different overview component integrates the different information of different networks,including convolution layer difference solar map,convolution core dimension reduction scatter map and convolution layer output data difference pixel map.Users can first observe the difference information of different networks through the overview component to find suspicious information,and then explore the convolution layer output device in the component by using difference.The suspicious points are further analyzed by volume feature map display component and weight detail display component.Finally,a specific experimental case is given to prove the effectiveness of the system designed in this paper.