Research Of Generative Adversarial Models In Deep Learning

With the wide application of artificial intelligence and deep learning in life,in order to meet the requirements of complex tasks,many scholars begin to research different neural network models,among which generative adversarial network is one of the excellent models.Generative adversarial network has become popular in deep learning since proposed,and the training for the network is actually a process of minimizing the maximum,therefore the generative adversarial models can be considered as a mathematical minimax optimization problem.We study the minimax problem and propose a new algorithm based on the existing optimization algorithms,which improve the efficiency of the problem and the generative adversarial network in turn.Therefore,it is significant to research the mathematical models for generative adversarial network.Firstly,we introduce the generative adversarial network and the corresponding minimax problem,and explain that the training and optimization of generative adversarial network are actually a process of minimizing the maximum.Optimizing the loss function of generating adversarial network can be seen as a non-convex-non-concave minimax problem.Secondly,we analyze the existing optimization algorithms for the minimax problem,and propose a new updating format.We combine the new updating format with the RMSprop algorithm and the Adam algorithm respectively to obtain new algorithms.The simulating data distribution experiment results on the generative adversarial network show that the new algorithms can make the loss function value lower.Finally,we analyze the factors affecting the optimization algorithms,including gradient direction and step-size,and propose a new type of learning step-size algorithm from the perspective of step-size.The new algorithm can update the step-size according to the current loss function during training dynamically and automatically.Comparing the new algorithm with the existing optimization algorithms for the minimax problem,the results show that the new algorithm can effectively improve training.