Title: Improved triple generative adversarial nets

Authors: Yaqiu Liu; Qinghua Zhao; Yunlei Lv; Kun Wang

Addresses: Department of Information and Computer Engineering, Northeast Forestry University, Harbin City, 150040, Heilongjiang Province, China ' Department of Information and Computer Engineering, Northeast Forestry University, Harbin City, 150040, Heilongjiang Province, China ' Department of Information and Computer Engineering, Northeast Forestry University, Harbin City, 150040, Heilongjiang Province, China ' College of Information and Electronic Technology, Jiamusi University, Jiamusi City, 154007, Heilongjiang Province, China

Abstract: Generative Adversarial Nets (GANs) have shown excellent performance in image generation and Semi-Supervised Learning (SSL). However, existing GANs have three problems: (1) the generator G and discriminator D tends to be optimal out of sync, and are not good at processing labelled data. (2) the generator G easy to generate chaotic semantics; and (3) the GANs cannot learn the inverse mapping (projecting data back into the latent space which is benefit to feature representation of related semantics). The problems caused by the limitation of two-player mode, where D can only share incompatible roles of identifying fake samples and predicting labels and distinguishes the data without label. To solve the problems, we propose an Improved Triple Generative Adversarial Nets (ITGANs) which consists of four parts-a generator G, a classifier C, an encoder EN and a discriminator D. The G and the C characterise the conditional distributions between images and labels, the discriminator distinguishes whether a pair of data (x, y) comes from the true distribution, and the EN maps data x to latent representations z. The experimental results show that the ITGANs achieve the state-of-the-art classification results among deep generative models, which demonstrate that the additional encoder can enhance the classification accuracy effectively.

Keywords: GANs; triple-GAN; encoder; deep convolutional networks.

DOI: 10.1504/IJCAT.2019.098029

International Journal of Computer Applications in Technology, 2019 Vol.59 No.2, pp.114 - 122

Received: 23 Apr 2018
Accepted: 23 Apr 2018
Published online: 27 Feb 2019 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article