Addressing GAN Training Instabilities via Tunable Classification Losses
IEEE Journal on Selected Areas in Information Theory · 2024 · 5 · 534-553
DOI: 10.1109/jsait.2024.3415670 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Journal on Selected Areas in Information Theory, 2024. Verfasst von Monica Welfert und 3 weiteren Autor:innen. Die Arbeit wurde bisher 24-mal zitiert (Crossref) und nennt 70 Literaturangaben. Gefördert von: National Science Foundation.
Abstract (Originalsprache)
Generative adversarial networks (GANs), modeled as a zero-sum game between a generator (G) and a discriminator (D), allow generating synthetic data with formal guarantees. Noting that D is a classifier, we begin by reformulating the GAN value function using class probability estimation (CPE) losses. We prove a two-way correspondence between CPE loss GANs and f-GANs which minimize f-divergences. We also show that all symmetric f-divergences are equivalent in convergence. In the finite sample and model capacity setting, we define and obtain bounds on estimation and generalization errors. We specialize these results to $\alpha $ -GANs, defined using $\alpha $ -loss, a tunable CPE loss family parametrized by $\alpha \in (0,\infty $ ]. We next introduce a class of dual-objective GANs to address training instabilities of GANs by modeling each player’s objective using $\alpha $ -loss to obtain $(\alpha _{D},\alpha _{G})$ -GANs. We show that the resulting non-zero sum game simplifies to minimizing an f-divergence under appropriate conditions on $(\alpha _{D},\alpha _{G})$ . Generalizing this dual-objective formulation using CPE losses, we define and obtain upper bounds on an appropriately defined estimation error. Finally, we highlight the value of tuning $(\alpha _{D},\alpha _{G})$ in alleviating training instabilities for the synthetic 2D Gaussian mixture ring as well as the large publicly available Celeb-A and LSUN Classroom image datasets.
Förderung
- National Science Foundation
Zitieren
APA
Welfert, M., Kurri, G. R., Otstot, K., & Sankar, L. (2024). Addressing GAN Training Instabilities via Tunable Classification Losses. IEEE Journal on Selected Areas in Information Theory, 5, 534-553. https://doi.org/10.1109/jsait.2024.3415670
Vancouver
Welfert M, Kurri GR, Otstot K, Sankar L. Addressing GAN Training Instabilities via Tunable Classification Losses. IEEE Journal on Selected Areas in Information Theory. 2024;5:534-553. doi:10.1109/jsait.2024.3415670
BibTeX
@article{Welfert2024,
author = {Welfert, Monica and Kurri, Gowtham R. and Otstot, Kyle and Sankar, Lalitha},
title = {Addressing GAN Training Instabilities via Tunable Classification Losses},
journal = {IEEE Journal on Selected Areas in Information Theory},
year = {2024},
volume = {5},
pages = {534--553},
doi = {10.1109/jsait.2024.3415670}
}
EUVIMED