Convolution neural network visual analysis method based on difference comparison

Abstract

The invention discloses a convolution neural network visual analysis method based on difference comparison, which comprises the following steps of designing a basic network model by using Tensorflow,and modifying parameters to obtain a control model; training the two models and extracting the parameters of the model after training; inputting the obtained model parameters into the variance analysis system to display the variance; by observing the variance overview component of the variance analysis system, identifying the possible key variance points quickly; through the interactive exploration component provided by the system, analyzing the possible key differences in further detail, and drawing the conclusion. The method of the invention effectively realizes the difference visual analysis, and the user can find the key problem more efficiently in the process of actually modifying the neural network model by understanding the difference.

Description

technical field [0001] The invention relates to a visual analysis method of a convolutional neural network. Background technique [0002] Deep convolutional neural network has a wide range of applications in pattern recognition fields such as automatic driving, cancer detection, sentence analysis, etc. In view of its good performance, more and more companies and researchers want to explore and apply it based on various needs. Characteristic research. However, the explanation of the operation mechanism of the convolutional neural network model is not perfect at this stage. Most of the time, we still treat it as a "black box", which will make it difficult for us to locate the problem and optimize the performance of the model when the output result of the model does not meet expectations. [0003] There are two main difficulties in understanding the convolutional neural network model. One is the large number of parameters inside the model. The convolution layer may contain hun...

AI Technical Summary

Problems solved by technology

[0005] Although many works have analyzed the problems of feature extraction and pattern clustering in neural networks, and analyzed various neural networks such as convolutional neural networks and recurrent neural networks, but for the sensitivity of affecting the performance of neural networks Differential visual analysis after changes in factors such as network model structure, input data characteristics and their correlations, activation function settings, and network parameters, there is still no very effective solution

Images