MDS20 Minitutorial: Data-Driven Methods for Inverse Problems by Ozan Öktem

Abstract: The mini-tutorial aims to provide a survey of different data-driven approaches to solve inverse problems with emphasis on deep learning based approaches that have gained widespread interest the last 2-3 years. The presentation will start out from the Bayesian approach to regularisation where a reconstruction method is represented by an estimator (decision rule). The various deep learning based approaches for solving inverse problems can now be seen as different ways to approximate estimators related to the posterior (deep direct estimation). Alternatively, one may also use trained deep neural networks to sample from the posterior (deep posterior sampling). The mini-tutorial will show examples of these methods in the context of tomographic imaging.