Generative Adversarial Networks: Overview

Deep neural networks (DNNs) have attained boundless accomplishments in dealing with high-dimensional data, particularly images. However, generating naturalistic images for tasks such as classification, detection, segmentation, reconstruction, and so on remains a difficult task. As a result, researchers are compelled to propose new algorithms and methods to overcome these challenges. These learning algorithms, on the other hand, frequently need a large volume of training data to be efficient. Also, its associated real-world challenges have limited data accessibility, and gathering sufficient amounts of data may be expensive. Additionally struggling with the issues of class imbalance, data shortage, and label propagation has turned the attention toward generative approaches, i.e., using modern procedures for producing new sample data with similar features as real sample data to enhance the efficiency of learning algorithms. Generative adversarial networks (GANs), a class of DNNs offer a novel way to model and generate data in an unsupervised manner. The deep learning community is increasingly pursuing an interest in GANs. The paper gives a basic introduction to GANs and discusses their variants in the study.