Abstract

We introduce a tunable GAN, called $\alpha$-GAN, parameterized by $\alpha \in(0,\infty]$, which interpolates between various $f$-GANs and IntegralProbability Metric based GANs (under constrained discriminator set). Weconstruct $\alpha$-GAN using a supervised loss function, namely, $\alpha$-loss,which is a tunable loss function capturing several canonical losses. We showthat $\alpha$-GAN is intimately related to the Arimoto divergence, which wasfirst proposed by \"{O}sterriecher (1996), and later studied by Liese and Vajda(2006). We posit that the holistic understanding that $\alpha$-GAN introduceswill have practical benefits of addressing both the issues of vanishinggradients and mode collapse.

 

Quick Read (beta)