Investigating Shifts in GAN Output-Distributions

TL;DR

This paper introduces a systematic method to investigate distribution shifts in GAN outputs, revealing significant differences from real data distributions and questioning existing convergence guarantees.

Contribution

It proposes a loop-training scheme and bounded measures for analyzing distribution shifts, providing new tools for understanding GAN limitations.

Findings

Large shifts observed between real and generated data distributions

Existing theoretical guarantees may not hold in practice

Method applicable across multiple datasets and GAN architectures

Abstract

A fundamental and still largely unsolved question in the context of Generative Adversarial Networks is whether they are truly able to capture the real data distribution and, consequently, to sample from it. In particular, the multidimensional nature of image distributions leads to a complex evaluation of the diversity of GAN distributions. Existing approaches provide only a partial understanding of this issue, leaving the question unanswered. In this work, we introduce a loop-training scheme for the systematic investigation of observable shifts between the distributions of real training data and GAN generated data. Additionally, we introduce several bounded measures for distribution shifts, which are both easy to compute and to interpret. Overall, the combination of these methods allows an explorative investigation of innate limitations of current GAN algorithms. Our experiments on different data-sets and multiple state-of-the-art GAN architectures show large shifts between input and output distributions, showing that existing theoretical guarantees towards the convergence of output distributions appear not to be holding in practice.