Codes and Expansions (CodEx) Seminar

Ben Adcock (Simon Fraser University)
Deep learning for inverse problems: confident hallucinations and new theoretical guarantees for Bayesian recovery

Deep learning is currently transforming how inverse problems arising in imaging reconstruction are solved. However, it is increasingly well-known that such deep learning-based methods are susceptible to hallucinations.  In this talk, I will present several theoretical explanations for why hallucinations occur, in both deterministic and statistical estimators. I will conclude by observing that hallucinations can only be avoided by careful design of the forwards operator in tandem with the recovery algorithm, and then provide a theoretical framework for how this can be achieved in a Bayesian setting in the case of posterior sampling using generative models.