Universum GANs: Improving GANs through contradictions

TL;DR

This paper introduces Universum GANs, a novel approach that enhances discriminator accuracy and generates high-quality data in limited labeled-data scenarios, supported by theoretical guarantees and empirical evidence.

Contribution

The paper proposes Universum GANs with an evolving discriminator loss, improving discriminator performance and data quality under limited data conditions.

Findings

Improved discriminator accuracy in limited data settings.

Generation of high-quality realistic data.

Theoretical guarantees supporting the approach.

Abstract

Limited availability of labeled-data makes any supervised learning problem challenging. Alternative learning settings like semi-supervised and universum learning alleviate the dependency on labeled data, but still require a large amount of unlabeled data, which may be unavailable or expensive to acquire. GAN-based data generation methods have recently shown promise by generating synthetic samples to improve learning. However, most existing GAN based approaches either provide poor discriminator performance under limited labeled data settings; or results in low quality generated data. In this paper, we propose a Universum GAN game which provides improved discriminator accuracy under limited data settings, while generating high quality realistic data. We further propose an evolving discriminator loss which improves its convergence and generalization performance. We derive the theoretical guarantees and provide empirical results in support of our approach.