Inverse Autoregressive Flows

A trick that allows fast learning and also fast (but approximate) sampling is to build a masked autoregressive flow to learn the distribution (the teacher) and then use this to train an inverse autoregressive flow from which we can sample eﬀiciently (the student)