A Seminar by Guillaume Laura (Université Côte d'Azur, France)
Abstract: Multilevel methods have harnessed dimension reduction to accelerate the solving of image restoration problems. Their framework, however, offers mild theoretical guarantees. On the other hand, block coordinate descent (BCD) algorithms have been thoroughly studied but have shown little success in accelerating image restoration problems.
By introducing a general BCD framework for non-smooth & non-convex problems, we are able to bridge a gap between both classes of algorithms, hence opening theoretical perspectives for multilevel algorithms and practical perspectives for BCD algorithms. We illustrate these findings with experiments.
Biography: Guillaume Laura is a Postdoctoral Researcher at Université Côte d’Azur.