FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models

A promising class of generative models maps points from a simple distribution a complex distribution through an invertible neural network. -based training of these models requires restricting their to allow cheap computation of Jacobian determinants. , the Jacobian trace can be used if the transformation is by an ordinary differential equation. In this paper, we use's trace estimator to give a scalable unbiased estimate of the-density. The result is a continuous-time invertible generative model with density estimation and one-pass sampling, while allowing unrestricted network architectures. We demonstrate our approach on high-dimensional estimation, image generation, and variational inference, achieving the-of-the-art among exact likelihood methods with efficient sampling.