Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: Realizing GANs via a Tunable Loss Function
(Submitted on 9 Jun 2021 (v1), last revised 18 Oct 2021 (this version, v2))
Abstract: We introduce a tunable GAN, called $\alpha$-GAN, parameterized by $\alpha \in (0,\infty]$, which interpolates between various $f$-GANs and Integral Probability Metric based GANs (under constrained discriminator set). We construct $\alpha$-GAN using a supervised loss function, namely, $\alpha$-loss, which is a tunable loss function capturing several canonical losses. We show that $\alpha$-GAN is intimately related to the Arimoto divergence, which was first proposed by \"{O}sterriecher (1996), and later studied by Liese and Vajda (2006). We also study the convergence properties of $\alpha$-GAN. We posit that the holistic understanding that $\alpha$-GAN introduces will have practical benefits of addressing both the issues of vanishing gradients and mode collapse.
Submission history
From: Gowtham Raghunath Kurri [view email][v1] Wed, 9 Jun 2021 17:18:21 GMT (53kb)
[v2] Mon, 18 Oct 2021 17:45:25 GMT (58kb)
Link back to: arXiv, form interface, contact.