Overcoming inference challenges using score-based generative models

Ronan Legin ( Université de Montréal )

In this talk, I demonstrate the effectiveness of score-based generative models as a powerful tool for addressing challenging inference problems in the fields of astrophysics and cosmology. Specifically, I will discuss likelihood analysis, where we propose a framework called Score-based LIkelihood Characterization (SLIC) to build a data-driven model of the likelihood from data obtained by the Hubble Space Telescope (HST) and the James Webb Space Telescope (JWST). Our results show that SLIC can perform unbiased and precise inference even in the presence of highly non-Gaussian correlated and spatially varying noise. Overall, score-based generative models can serve as a versatile tool for addressing complex inference problems, which could potentially lead to important breakthroughs in the field of astrophysics and cosmology.