Learning an optimization solver for a class of inverse problemsPP

We consider methods for solving large scale optimization problems based on machine learning. In particular, we specify a class of optimization algorithms using only linear operations and applications of proximal operators, which is general enough to span first-order solvers like Chambolle-Pock. We then apply unsupervised learning to find the best solver in this class for solving the TV-problem in tomography, with constraint on the computation time. Finally, the trained solver is compared to state-of-the-art solvers.

This is poster number 54 in Poster Session

Axel Ringh (KTH - Royal Institute of Technology)
Jonas Adler (KTH Royal Institute of Technology)
Sebastian Banert (KTH - Royal Institute of Technology)
Ozan Öktem (KTH - Royal Institute of Technology)
Johan Karlsson (KTH - Royal Institute of Technology)
computed tomography, deep learning, image reconstruction, inverse problems, machine learning, nonlinear optimization