English

$\alpha$-GAN: Convergence and Estimation Guarantees

Machine Learning 2022-05-16 v1 Information Theory math.IT Machine Learning

Abstract

We prove a two-way correspondence between the min-max optimization of general CPE loss function GANs and the minimization of associated ff-divergences. We then focus on α\alpha-GAN, defined via the α\alpha-loss, which interpolates several GANs (Hellinger, vanilla, Total Variation) and corresponds to the minimization of the Arimoto divergence. We show that the Arimoto divergences induced by α\alpha-GAN equivalently converge, for all αR>0{}\alpha\in \mathbb{R}_{>0}\cup\{\infty\}. However, under restricted learning models and finite samples, we provide estimation bounds which indicate diverse GAN behavior as a function of α\alpha. Finally, we present empirical results on a toy dataset that highlight the practical utility of tuning the α\alpha hyperparameter.

Cite

@article{arxiv.2205.06393,
  title  = {$\alpha$-GAN: Convergence and Estimation Guarantees},
  author = {Gowtham R. Kurri and Monica Welfert and Tyler Sypherd and Lalitha Sankar},
  journal= {arXiv preprint arXiv:2205.06393},
  year   = {2022}
}

Comments

Extended version of a paper accepted to ISIT 2022. 12 pages, 7 figures

R2 v1 2026-06-24T11:16:03.816Z