Uncertainty-of-Information-Driven GAN (UoI GAN): Quantifying and Communicating Uncertainty to Decision-Makers

Abstract

Generative Adversarial Networks (GANs) are cutting-edge machine learning algorithms that can generate realistic data like images and time series. Their application has been explored in autonomous systems, fraud detection, and medical diagnostics. The “black box” nature and intrinsic instability of these models pose significant and frequently hidden risks. Decision-makers struggle to interpret GAN outputs because, unlike predictive models, GANs do not explicitly measure confidence. Using simply GAN average performance statistics like the Fréchet Inception Distance can obscure important failure modes. GAN systems must be designed to explicitly expose and communicate their uncertainty. The hybrid model we propose in this research combines explicit uncertainty representation with quantitative uncertainty of information measurement. Our model re-designs the GAN to intrinsically represent uncertainty by examining output variance. It then provides a simple way for translating these complex metrics into interpretable signals that expose model behavior and limitations.