Approximate Inference for Generative Models of Astronomical Images
Jeffrey Regier, University of California, Berkeley
A central problem in astronomy is to infer locations, colors, and other properties of stars and galaxies appearing in astronomical images. In this talk, I present a hierarchical probabilistic model for the problem: the number of photons arriving at each pixel during an exposure is Poisson distributed, with a rate parameter that is a deterministic function of the latent properties of the imaged stars and galaxies. I then propose a procedure based on variational inference to approximate the posterior distribution of the latent properties. The procedure attains state-of-the-art results. It demonstrates the scaling characteristics necessary to construct an astronomical catalog from the full 20-terabyte Sloan Digital Sky Survey dataset, with a computational budget of 1 million CPU hours. Finally, I present an enhanced model for galaxies that uses a neural network, also trained with variational inference, to encode the conditional distribution of the data given the latent random variables. The proposed model assigns higher probability to held-out data than the current standard practice and provides a rare example of successful application of a neural network to unsupervised learning.