
Deep Generative Models for Bayesian Inference in Astrophysics
Biwei Dai of the Institute for Advanced Study presents a seminar on using deep generative models to solve inference problems in astrophysics. The first half covers field-level cosmological inference, showing how generative models can build likelihood functions that extract more information from data than traditional two-point statistics, with an application to weak gravitational lensing analysis using the Hyper Suprime-Cam survey. The second half turns to Bayesian inverse problems, demonstrating how generative models serve as physically informed priors for reconstructing images of AGN accretion disks from intensity interferometry data, where phase information is lost and only Fourier amplitudes are measured. Dai shows how sampling from the resulting posterior distributions yields high-fidelity reconstructions with uncertainty quantification that outperform traditional iterative methods across different noise levels and data coverage. The talk runs about an hour and assumes familiarity with Bayesian statistics and astrophysical data analysis.