A survey on generative adversarial networks
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Since the Generative Adversarial Network (GAN) was first introduced by Goodfellow \textit{et al.} in 2014, it has become one of the most popular topics in machine learning, especially in computer vision domain. Extensive studies have been conducted to analyze and improve GAN from different angles, and many new variants of GANs have also been proposed and shown to achieve remarkable performance. Despite this success of GANs, there are, unfortunately, still many unresolved issues associated with the training of GANs, such as mode collapse, gradient vanishing, and lack of benchmark for evaluation. The purpose of this thesis is to provide a thorough survey on GANs, in order for researchers in this area to be quickly exposed to the opportunity and challenges with GANs. In particular, this thesis is organized in the following three segments: First, we survey the original GAN, Autoencoder, and other modified GANs with motivation, mathematical representations, and structure presented. Secondly, several evaluation metrics will be surveyed and discussed. Thirdly, GAN training obstacles and techniques for GAN training and performance improvement will be demonstrated.
Description
Keywords
Evaluation, Generative adversarial networks, Generative models, Training