Evolving GAN Formulations for Higher Quality Image Synthesis

TL;DR

This paper introduces TaylorGAN, a novel method that evolves customized Taylor expansion-based loss functions for GANs, leading to improved image quality and stability in image translation tasks.

Contribution

It proposes a new technique to optimize GAN loss functions via multiobjective evolution, enhancing training stability and output quality.

Findings

Qualitative improvement in generated images

Quantitative improvement in GAN performance metrics

Enhanced stability in training process

Abstract

Generative Adversarial Networks (GANs) have extended deep learning to complex generation and translation tasks across different data modalities. However, GANs are notoriously difficult to train: Mode collapse and other instabilities in the training process often degrade the quality of the generated results, such as images. This paper presents a new technique called TaylorGAN for improving GANs by discovering customized loss functions for each of its two networks. The loss functions are parameterized as Taylor expansions and optimized through multiobjective evolution. On an image-to-image translation benchmark task, this approach qualitatively improves generated image quality and quantitatively improves two independent GAN performance metrics. It therefore forms a promising approach for applying GANs to more challenging tasks in the future.