Abstract

In this work, we propose a new family of generative flows on an augmenteddata space, with an aim to improve expressivity without drastically increasingthe computational cost of sampling and evaluation of a lower bound on thelikelihood. Theoretically, we prove the proposed flow can approximate aHamiltonian ODE as a universal transport map. Empirically, we demonstratestate-of-the-art performance on standard benchmarks of flow-based generativemodeling.

 

Quick Read (beta)