Generative Priors with Tunable Complexity for Inverse Problems

Data Science Seminar

Paul Hand
Northeastern University

Abstract

Generative models have emerged as powerful priors for solving inverse problems, such as compressed sensing. These models typically represent a class of natural images using a single fixed complexity or dimensionality.  In this talk, we will study the benefits of training a family of generative models with varying levels of complexity.  We will see experimental evidence that appropriately trained families of generative models lead to decreased signal reconstruction error on a variety of imaging inverse problems.  We will also see theoretical results illustrating this effect in the case of denoising and compressed sensing in the setting of a linear generative prior.