normalizing flows

1. highlights

gets you an exact likelihood, rather than a lower bound on the likelihood, like from variational auto encoders