Keywords: Demixing, Generative Models, GAN, Unsupervised Learning, Structured Recovery
TL;DR: An unsupervised learning approach for separating two structured signals from their superposition
Abstract: Recently, Generative Adversarial Networks (GANs) have emerged as a popular alternative for modeling complex high dimensional distributions. Most of the existing works implicitly assume that the clean samples from the target distribution are easily available. However, in many applications, this assumption is violated. In this paper, we consider the observation setting in which the samples from a target distribution are given by the superposition of two structured components, and leverage GANs for learning of the structure of the components. We propose a novel framework, demixing-GAN, which learns the distribution of two components at the same time. Through extensive numerical experiments, we demonstrate that the proposed framework can generate clean samples from unknown distributions, which further can be used in demixing of the unseen test images.