A Piggyback-Style Algorithm for Learning Improved Shearlets and TGV Discretizations

Abstract

Summary. This work demonstrates how to use a piggyback-style algorithm to compute derivatives of loss functions that depend on solutions of convex-concave saddle-point problems. Two application scenarios are presented, where the piggyback primal-dual al-gorithm is used to learn an enhanced shearlet transform and an improved discretization of the second-order total generalized variation.