A deep generative model for astronomical images of galaxies

Simple parametric models (e.g. [1, 2, 3, 4, 5]) suffice to describe many idealized galaxy shapes, but severely misfit actual galaxies: they are not sufficiently flexible [5]. The popular program GALFIT copes with the limitations of simple parametric models by allowing users to fit an arbitrary number of mixture components [6, 7]. These mixtures are not learned from actual galaxies, so the models cannot provide meaningful uncertainty estimates. To our knowledge, no existing galaxy models are learned from a training set, which would allow for such uncertainty estimates. Indeed, only [2] attempts a Bayesian treatment of galaxy shapes1, albeit one based on just a few manually specified parameters. Yet modeling galaxies is an important part of learning about the universe from large-scale astronomical sky surveys [4, 5, 8, 9], and billions of images of galaxies are available for training.