Deep generative models offer powerful tools for solving astrophysical inference problems by enabling flexible representations of prior knowledge and likelihood functions. In this talk, I will present some of my recent works on applying deep probabilistic models to Bayesian inference in astrophysics.
In the first part of the talk, I will demonstrate how generative models can be leveraged to construct physically informed priors for Bayesian inverse problems. As an application, I will show how this approach enables robust image reconstruction from intensity interferometry data, where only the amplitudes of Fourier modes are measured while phase information is lost. By sampling from the resulting posterior distribution, we achieve high-fidelity reconstructions with uncertainty quantification, outperforming traditional iterative methods across varying noise levels and UV-plane coverage.
In the second part, I will discuss how generative models can be employed to construct likelihood functions for cosmological inference at the field level, enabling more effective extraction of information compared to traditional summary statistics like two-point statistics. This simulation-based inference framework facilitates anomaly detection of model misspecification, and enhances interpretability through sample generation. I will present applications to weak gravitational lensing analysis, particularly our ongoing work on applying this approach to the field-level analysis of the Hyper Suprime-Cam (HSC) survey.