AGAN: Towards Automated Design of Generative Adversarial Networks

TL;DR

AGAN introduces an automated neural architecture search method tailored for GANs, outperforming human-designed models on image generation tasks and demonstrating transferability across datasets.

Contribution

This paper presents the first neural architecture search algorithm specifically designed for GANs, reducing reliance on human expertise and trial-and-error.

Findings

Outperforms state-of-the-art GAN architectures on CIFAR-10.

Achieves competitive results on supervised tasks at 32x32 resolution.

Modules learned are transferable to other datasets like STL-10.

Abstract

Recent progress in Generative Adversarial Networks (GANs) has shown promising signs of improving GAN training via architectural change. Despite some early success, at present the design of GAN architectures requires human expertise, laborious trial-and-error testings, and often draws inspiration from its image classification counterpart. In the current paper, we present the first neural architecture search algorithm, automated neural architecture search for deep generative models, or AGAN for abbreviation, that is specifically suited for GAN training. For unsupervised image generation tasks on CIFAR-10, our algorithm finds architecture that outperforms state-of-the-art models under same regularization techniques. For supervised tasks, the automatically searched architectures also achieve highly competitive performance, outperforming best human-invented architectures at resolution $32\times32$. Moreover, we empirically demonstrate that the modules learned by AGAN are transferable to other image generation tasks such as STL-10.