Iterative Updates of Score-Based Priors for Inverse Imaging in Strong Gravitational Lenses.

Gabriel Missael Barco ( Université de Montréal )

Score-based/diffusion models (SBMs) can be used as priors for a wide range of inverse problems that allow for a fully Bayesian approach in high-parametric inference problems, such as when working with images. This approach has been applied and studied in different scientific setups, such as strong gravitational lensing and radio interferometry. In the former, it could be used for inverse imaging of the source of the strong lensing system. However, having a misspecified prior could bias the posterior sampling procedure and negatively affect downstream tasks. Since SBMs are amortized priors, based completely on data, this is a possibility. We explore the effect of having a misspecified prior under this setup, and how to address it by performing iterative updates to the SBM prior.