Triple-GAN with Fixed Memory Step Gradient Descent Method and Xwish Activation Function

Abstract

Generative adversarial network is an important generation model in the fields of big data, machine learning and deep learning. For semi-supervised learning, the improved generative adversarial network plays an important role in image generation and classification. Triple-GAN is the best one among them, which consists of classifier generator and discriminator, and it guarantees that the data distribution of generator and classifier can converge to the real distribution. However, the activation function and the gradient descent method of the original generative adversarial network have not been changed, and still have traditional limitations, which are prone to gradient disappearance and gradient explosion. Therefore, this paper adopts the new activation function Xwish and the fractional gradient step descent method with fixed memory step gradient descent method instead of the original activation function and the gradient descent method. It makes up for the disadvantages of inactivation caused by neuron flowing into large gradient while preventing gradient explosion and gradient disappearance with certain sparsity and introduces adaptive parameters to achieve better activation efficiency, which greatly improves the efficiency of deep learning.