Generative Adversarial Networks: Models and Techniques - A Review

Deep learning has exhibited a tremendous amount of success in the domain of artificial intelligence, numerous deep learning models have been built. Generative Adversarial Network (GAN) is a deep learning approach which has recently emerged as a popular study area. Since 2014, GANs have been extensively researched, and especially in various fields like image classification, language and speech processing, computer vision, etc. It has shown a lot of success. GANs are models that can be utilized to generate fresh samples with data distributions that are comparable to the training data-set. The importance of model variation is to generate synthetic realistic data by using unsupervised learning to achieve the desired data distribution. They can deal with representations that can be employed in a variety of applications.

I. Introduction

With the advancement of artificial intelligence research, machine’s ability to mimic human behavior is increasing at an exponential rate. Generative Adversarial Networks (GANs) are an emerging approach introduced by Goodfellow [1] which have been well crafted for performing certain tasks. Being a hot topic for research, it is a relatively new approach for both semi-supervised and unsupervised learning. These networks achieve results by modelling high-dimensional data distributions. GANs can avoid some shortcomings of traditional approaches by modelling the high dimensional distribution of data. It can optimize some loss functions that are difficult to handle, via adversarial learning, which allows us to realize semi supervised and unsupervised learning technology. The core premise of GANs is based on a two-person zero-sum game in which total gains are zero for both the players and loss or gain of utility of each player is exactly balanced by the gain or loss of utility of another player [2].