Robustness Analysis And Application Research Of Deep Neural Network

The research of deep neural networks has been rapidly expanding,with its robustness becoming a focus.Thesis investigates and designs a powerful,stability-preserving generation antagonism network,introducing the structure of its generation and discrimination networks,as well as the stability structure incorporated into the discrimination network.In view of the instability of the training of the generated countermeasure network and its application in the prediction of sea surface temperature,the loss function of the model is improved,and the controllability of Wasserstein distance is analyzed.Robustness research methods are divided into accurate methods and approximate methods.Different analysis methods can be selected for different deep learning tasks.The research center of this paper is to study the robustness of the deep neural network based on the Lipscihtz constant method,give the linearization method of the nonlinear deep neural network model,and study the robustness conditions when the data or network parameters are disturbed.The specific research content is divided into two parts.The first part studies the robust conditions of deep neural networks.Firstly,the method of linearly mapping a nonlinear deep neural network model is studied.Secondly,the conditions for the robustness of the deep neural network model when data or network parameters are disturbed are studied,namely,the value range of the weight matrix norm of each network layer of the deep neural network model and the value range of the eigenvalues of the activation function transformation matrix.At the same time,the selection of activation functions for each network layer is studied when network parameters are disturbed.The second part designs and builds a robust strong stability preserving generative adversarial network model.Firstly,a generating network and a discriminant network with stable structures are designed,and the stability of the discriminant network is proved.Secondly,the controllability of Wasserstein distance and the loss function of the model are analyzed.To solve the problem of unstable training of generating confrontation networks,a Lipschitz condition is applied to the gradient of the discriminating network,that is,the Lipschitz condition is applied to the real data distribution area of sea surface temperature,the generated data distribution area,and the area between them.Limiting its gradient to around 1 can achieve a good effect.The selection of the loss function of the generated network can predict the mean square deviation formula of sea surface temperature.Finally,the strong stability preserving generative adversarial network is trained and tested,and the accuracy and robustness of the model in predicting sea surface temperature are verified through experiments.