Abstract

We present Multi-chart flows, a flow-based model for concurrently learningtopologically non-trivial manifolds and statistical densities on them. Currentmethods focus on manifolds that are topologically Euclidean, enforce strongstructural priors on the learned models or use operations that do not scale tohigh dimensions. In contrast, our model learns the local manifold topologypiecewise by "gluing" it back together through a collection of learnedcoordinate charts. We demonstrate the efficiency of our approach on syntheticdata of known manifolds, as well as higher dimensional manifolds of unknowntopology, where we show better sample efficiency and competitive or superiorperformance against current state-of-the-art.

 

Quick Read (beta)