Learned Adversarial Regularisers
Monday, November 9, 2020 - 10:40am - 11:25am
In this talk we review some recent advances in the derivation of data-driven regularisers for inverse problems. We are in particular interested in those which are represented by a neural network and which are trained using an unsupervised, adversarial loss. The latter has links to optimal transport as the loss is derived as an approximation to the Wasserstein distance - which also allows us to prove some things. We discuss non-convex and convex versions of these learned regularisers and show applications to image denoising, deconvolution, low-dose and limited angle tomography. This talk is based on joint works with Sebastian Lunz, Subhadip Mukherjee, Sören Dittmer, Zak Shumaylov and Ozan Öktem.