Visual Analysis Of Convolutional Neural Network Filters Based On Image Datasets

In recent years,convolutional neural networks have achieved remarkable success in many fields,especially in the field of image recognition,the accuracy of which has been very close to or even exceeds the level of related experts.However,the working mechanism of convolutional neural networks has not yet had a reasonably rigorous explanation,and it is still a "black box" that is incomprehensible to experts.Training a good model is often a trial and error process,which requires a lot of time and computer resources.In order to speed up the training process and reduce the number of trials,experts need to understand what happened during the training process.However,although the current mainstream machine learning frameworks provide visual tools to help users analyze,they only provide a small amount of information and cannot be further explored.Far from meeting the analysis needs of some experts.Convolutional neural networks are usually composed of dozens or even hundreds of hidden layers,and each hidden layer is composed of dozens or even hundreds of filters.Even though the convolutional neural network has the characteristic of sharing weights,its internal independent weight parameters are still hundreds of thousands or even millions,and these weight parameters will be updated at any time as the iteration progresses.Simply observing the value is impossible.Visual analysis is an effective means to assist users in understanding such complex numerical problems.Compared with traditional numerical analysis methods,it is not only more intuitive and easier to understand,but also has lower professional requirements for users.By establishing a multi-level exploration and interactive visual analysis framework,it helps users explore the evolution of filters in convolutional neural networks,understand the role of filters in each hidden layer in convolutional neural networks,and assist users in finding convolutional neural networks Abnormal phenomena that occurred during the training process and analyze their causes,so as to further optimize the network structure parameters and adjust the training method.This paper proposes a method to efficiently explore the hidden layers of convolutional neural networks by evaluating the training quality of each level of convolutional neural networks,and introduces visualization techniques to perform in-depth analysis of convolutional neural network training log data at multiple levels.The main research contents and contributions of this article are as follows:We propose a visual exploration process from the network layer to the hidden layer,from the hidden layer to the filter,design multiple intuitive and easy-to-understand views with rich visual coding,to help the analyst to intuitively and efficiently discover the abnormal conditions generated during the training process and Can further explore the causes of its anomalies.An evaluation method of the hidden layers in the convolutional neural network is studied.Consider that the number of features that the filter can independently identify in the convolutional neural network is an important indicator to measure the ability of the hidden layer to identify the feature pattern.The filter's confusion is calculated statistically Frequency distribution to evaluate the training quality of the hidden layer.We suggest a multi-dimensional visual perception method of filters,display the filters that users are interested in from multiple perspectives,analyze and compare the distance between different filters,and design multiple visual views to help users quickly locate abnormal filters.An integrated visual analysis system is designed and developed.Through multilevel visual views and rich multi-picture linkage technology,it can provide guidance to users in terms of optimizing network structure,assisting parameter settings,and exploring abnormal causes...