Bayesian Imaging with Deep Generative Priors

Explore a seminar talk on Bayesian imaging methodology that introduces a novel approach combining Bayesian analysis and computation for high-dimensional inference problems. Learn about constructing prior distributions supported on low-dimensional manifolds encoded by deep neural networks, where prior knowledge comes from training example datasets. Discover how the manifold hypothesis applies to real-world physical quantities and understand the process of learning manifolds and distributions through modern machine learning techniques for generative modeling. Examine theoretical and empirical studies of Bayesian models through challenging imaging inverse problems, with focus on uncertainty quantification, hypothesis testing, and model selection scenarios where ground truth is unavailable. Presented by Dr Marcelo Pereyra from Heriot-Watt University at the INI Seminar as part of the mathematical and statistical foundation of future data-driven engineering series.