Learned Regularizers for Inverse Problems

In the past years, there has been a surge of interest in methods to solve inverse problems that are based on neural networks and deep learning. A variety of approaches have been proposed, showing improvements in reconstruction quality over existing methods. Among those, a class of algorithms builds on the well-established variational framework, training a neural network as a regularization functional. Those approaches come with the advantage of a theoretical understanding and a stability theory that is built on existing results for variational regularization. We discuss various approaches for learning a regularization functional, aiming at giving an overview at the multiple directions investigated by the research community.