Neural Architecture Search for Generative Adversarial Networks: A Review

Abstract

Achieving competent image generation results in Generative Adversarial Networks (GANs) has been observed to be weighed down by barriers such as mode collapse and being error-prone, due to the time consumption, effort and domain expertise required to arduously design novel Neural Network Architectures through trial-and-error. Neural Architecture Search (NAS) which is a subfield of Automated Machine Learning (AutoML), has gained recent attention for its ability to automate the process of Neural Network architecture generation. NAS has achieved promising results in many computer-vision fields such as image classification, image segmentation, object detection and has now found its way into image generation tasks of Generative Adversarial Networks. Recent works on NAS systems that generate novel GAN architectures have proven this direction of applying NAS is promising by producing results that rival current state-of-the-art human-made architectures. The existing works related to NAS for GANs along with their approaches, performance and image generation results has been explored in this paper to assist and inspire research in the domain of Image Synthesis via NAS for GANs.