Improving Multi-Manifold GANs with a Learned Noise Prior

Matthew Amodio, Smita Krishnaswamy

Keywords: GAN, generative adversarial network, ensemble

TL;DR: A multi-generator GAN framework with an additional network to learn a prior over the input noise.

Abstract: Generative adversarial networks (GANs) learn to map samples from a noise distribution to a chosen data distribution. Recent work has demonstrated that GANs are consequently sensitive to, and limited by, the shape of the noise distribution. For example, a single generator struggles to map continuous noise (e.g. a uniform distribution) to discontinuous output (e.g. separate Gaussians) or complex output (e.g. intersecting parabolas). We address this problem by learning to generate from multiple models such that the generator's output is actually the combination of several distinct networks. We contribute a novel formulation of multi-generator models where we learn a prior over the generators conditioned on the noise, parameterized by a neural network. Thus, this network not only learns the optimal rate to sample from each generator but also optimally shapes the noise received by each generator. The resulting Noise Prior GAN (NPGAN) achieves expressivity and flexibility that surpasses both single generator models and previous multi-generator models.