A Visual Method for Interpreting Convolutional Neural Networks

Abstract

The invention relates to a visualization method for explaining convolutional neural networks, comprising: preparing a convolutional neural network model M and its training set S; extracting all the judgment conditions used by the model M in the decision-making process; using neurons and human corpus The matching degree of the semantics in the middle determines the semantics of the neurons, and generates understandable semantics for all judgment conditions; forms a decision tree T, uses its decision-making process as the decision-making process of the model M; converts the decision tree T into a tree flow graph; makes neurons Semantic view; making neuron relationship diagram; making decision data flow diagram; building an interactive visualization system.

Description

technical field [0001] The present invention relates to machine learning and visualization techniques, especially visualization methods for explaining deep convolutional neural networks. Background technique [0002] Machine learning has become one of the most efficient data analysis tools. It has received extensive attention in the industrial and academic fields. Despite the high efficiency of machine learning models, their opacity and inexplicability are the most criticized places. If you look at the machine learning model according to its interpretability and learning ability, you will find that linear regression has the highest interpretability and the lowest learning ability, while the neural network model, on the contrary, has the lowest interpretability and the highest learning ability. At the same time, in industry, users who use neural networks to make predictions need to understand how neural networks make decisions. Academically, researchers also hope to have a...

AI Technical Summary

Problems solved by technology

However, this method is only applicable to shallow neural networks (only one hidden layer). Once a deep neural network is encountered, the extracted rules are too complex for humans to analyze and understand, which leads to the failure of this method. When it comes to convolutional neural networks, the complexity of the rules extracted by this method will further increase

The third is the visualization of deep learning features proposed by Fu Kun et al. (Patent Publication No. CN106909945A), but it is only a qualitative analysis of convolution features, which is equivalent to verifying that the features learned by the deep learning model are from low-level to high-level, and Failure to understand the model's decision-making process

Therefore, the scope of these methods is limited and not universal

Images