Jisuanji kexue yu tansuo (Jan 2020)

Generative Adversarial Networks GAN Overview

  • LIANG Junjie, WEI Jianjing, JIANG Zhengfeng

DOI
https://doi.org/10.3778/j.issn.1673-9418.1910026
Journal volume & issue
Vol. 14, no. 1
pp. 1 – 17

Abstract

Read online

As a new unsupervised learning algorithm framework, generative adversarial networks (GAN) has been favored by more and more researchers, and it has become a research hotspot. GAN is inspired by the two-person zero-sum game theory in game theory. Its unique confrontation training idea can generate high-quality samples and has more powerful feature learning and feature expression ability than traditional machine learning algorithms. At present, GAN has achieved remarkable success in the field of computer vision, especially in the field of sample generation. Each year, there are a large number of GAN-related research papers. For the hotspot model of GAN, firstly this paper introduces the research status of GAN; then introduces the theory and framework of GAN, and analyzes the reasons why GAN has gradient disappearance and mode collapse during training; then discusses some typical models of GAN. This paper summarizes the improvement, advantages, limitations, application scenarios and implementation costs of the theory. At the same time, this paper compares GAN with VAE (variational autoencoder) and RBM (restricted Boltzmann machine) models, and summarizes the advantages and disadvantages of GAN. Finally, the application results of GAN in data generation, image super-resolution, image style conversion, etc. are presented, and the challenges and future research directions of GAN are discussed.

Keywords