High-Dimensional Bayesian Inference for Astronomical Imaging and Strong Gravitational Lensing.

Abstract

Modern telescopes such as Hubble and JWST deliver exquisitely detailed images that encode rich information about galaxies, stars, and dark matter. Extracting this information requires solving high-dimensional inverse problems in which each pixel represents a parameter of the sky. In this talk, I will present a Bayesian framework that uses diffusion-based generative models to perform inference directly in these high-dimensional spaces. By learning expressive priors over galaxy morphologies and likelihoods over complex, non-Gaussian noise distributions, this approach enables the reconstruction of astronomical images at unprecedented fidelity. This talk will cover a wide range of applications where we have applied these techniques, from the recovery of detailed images of protoplanetary discs using ALMA data and high redshift galaxies using Hubble data and strong gravitational lensing to a discussion about the misscalibration of the generative model. The generative model often needs to be trained on low redshift galaxies or using simulations, all of which could introduce biases in our inference. To address this, I will present a framework which can update the parameters of the generative model using corrupted data only. Together, these results illustrate how generative modeling can be used in a principled statistical framework for astronomy.