Image Classification Method Based On Interpretable CNN

Convolutional Neural Networks(CNNs)^[1]has made outstanding performance since it was proposed.However,the end-to-end learning strategy of CNNs makes the interpretability of the network unreliable.The whole learning process of end-to-end learning is not the artificial sub problem division,but the mapping from the original data to the desired output.Its advantage is to give the model more space which can be automatically adjusted according to the data,and increase the fitness of the model,meanwhile,the interpretability of the model becomes unclear.The interpretability of the model is not only a challenge in the research and development of neural networks,but also an important characteristic and weakness of CNN.It has high figures in theory and practice.In this paper,we proposes an image classification method based on mutual information and a visualization method combining pattern mining and up-convolution network to explore the interpretability of Convolution Neural Networks.The specific work is as follows:(1)A method of feature extraction under a certain prior constraint is designed.A set of masking templates with Euclidean distance assignment is designed,which combines the local characteristics of convolution operation.By masking the feature map of the filter output,the active region is converged to enhance the feature representation of the filter.(2)A loss function based on mutual information is proposed to standardize the representation of the filter,so as to improve the interpretability of the filter and the whole network.Based on this,an interpretable CNN image classification model based on mutual information loss is built.By training ordinary CNN to compare with our interpretable CNN,we calculate the part interpretability of patterns and location instability of inference positions about two networks for single classification task and multi classification task.(3)Combined with the idea of pattern mining and up-convolution network,this paper proposes a method to visualize the convolution layer of CNN,which can be used in our interpretable CNN to verify the credibility and interpretability of our model.