Normalizing Flows for Bayesian Experimental Design in Imaging Applications

Neural density estimators, such as invertible normalizing flows, are capable of estimating the Bayesian posterior distribution in a variety of imaging problems, including medical MRI and seismic imaging/monitoring. So far, few works explore the possibility to make explicit use of probabilistic information contained within the full Bayesian solution of these inverse problems. During our talk, we investigate how a simple modification to the training objective of conditional normalizing flows allows for Bayesian experimental design without modifying the normalizing flow's neural architecture itself. By establishing a key relationship between the expected information gain (EIG) and the maximum-likelihood, attained during the training of normalizing flows, we show that optimal experimental design can be achieved. During our talk, we first verify, on a stylized problem, that our method indeed maximizes the expected information gain, followed by demonstrating the advocacy of our methodology on large-scale medical and seismic problems.