Research On The Decision Attribution Interpretability Of Deep Learning Models For Image Classification

Deep learning models have been widely studied and applied in image classification tasks,but they are usually regarded as black-box models driven by data,and it is difficult to understand the decision-making basis of the model,so the output results of the model are in an uninterpretable state,making the credibility of the model is questionable.Therefore,designing an interpretation method for a deep learning model is an important way to improve model credibility and optimize model performance.In view of the lack of exactness and consistency of existing interpretation methods,the trade-off between visualization and exactness of interpretation,and the lack of theory to guide the optimization of models,this dissertation focuses on the deep neural networks(DNN),convolutional neural networks(CNN)and deep residual networks(Res Net),which are three typical deep learning models for image classification,conducting research on the interpretability of decision-making attribution,and further researches how to use the resulting interpretation methods to guide the optimization of model parameters for Res Net.The main work of this dissertation specifically includes the following aspects.(1)For the most basic DNN,aiming at the problem that most existing interpretation methods cannot measure the exactness and consistency of the interpretation,a global interpretation method of DNN that can prove the convergence and consistency is proposed.The main idea of this interpretation method is to first convert the DNN into a set of piecewise linear neural networks(PLNN)with global interpretability that can be proven to converge with DNN,and then convert the PLNN into a set of equivalent linear classifiers,and finally the decision features of each image with respect to the target class are interpreted by a linear classifier corresponding to each input image.Experimental results show that PLNN with global interpretability can sufficiently approximate with DNN under certain conditions,thereby ensuring the credibility of the interpretation,and also verifying the consistency of the interpretation method.In addition,the experiments show that the decision features obtained by the interpretation method have semantic meaning through the Fashion-MNIST dataset,and help to find the reason for making wrong decisions on some samples in the prediction.(2)For CNN with convolution operation,aiming at the problem that existing interpretation methods lack pixel-level visual interpretation to generate accurate classification decision features,a fine-grained local interpretation method of CNN is proposed to generate classification decision interpretation map that are consistent with the real behavior of CNN and have good visualization effects.First,through theoretical derivation,it is found that the classification decision of CNN can be decomposed into the sum of the gradient term of the input image and the gradient term of the bias,and then the gradient of the target class relative to the input is calculated to explain the contribution of each pixel of the input image for the decision result.To weigh the exactness of the interpretation and the visualization effect,an appropriate threshold is chose to selectively filter the gradient,so as to obtain a classification decision interpretation map that can show fine-grained decision features of the input image of CNN with respect to the target class.Experimental results show that this method has a finer-grained visual interpretation effect and higher interpretation exactness than other existing CAM-based methods,and has better practicability in troubleshooting and guiding optimization models.(3)For Res Net with a connection operation,firstly,aiming at the problem of the gradient vanishing and saturation issue of the traditional gradient-based interpretation methods,resulting in a lack of exactness in the interpretation,a provably exact and consistent local interpretation for Res Net using a local interpretation substitute model,a neural ordinary differential equation network(Neural ODE),is proposed.Experiments show that the interpretation method has high exactness and low computational cost while ensuring the effect of visual interpretation,and this interpretation can be effectively used to diagnose and optimize the model.Then,on the basis of the proposed interpretation method of Res Net,further aiming at the problem of low efficiency of artificial parameter adjustment in the random training of Res Net,the obtained interpretation method is used to establish the connection between the Res Net random training model and the stochastic optimal control theory,the expression of the optimal dropout probability in Res Net random training is deduced by the dynamic programming method,thus providing a theoretical guarantee for guiding and optimizing the model.And experiments show that choosing the optimal dropout probability can improve the generalization ability of Res Net compared with the randomly selected dropout probability.Finally,the main research content of this paper is summarized,and the future work is briefly introduced.