Latest work exploring the use of the ProxSkip algorithm
We are excited to share our latest work exploring the use of the ProxSkip algorithm as an efficient solution for accelerating iterative methods in imaging inverse problems. This project was led by our Senior Research Scientist Evangelos Papoutsellis, in collaboration with Kostas Papafitsoros (Queen Mary University) and Zeljko Kereta (University College London). By randomly skipping regularisation steps, ProxSkip significantly reduces computational time without compromising convergence. We also introduce a novel variant, PDHGSkip, which further enhances performance. Extensive numerical experiments demonstrate that these methods deliver faster computations while maintaining high-quality reconstructions.
We acknowledge funding from from the Analysis for Innovators (A4i) Denoising of chemical imaging and tomography data project, in collaboration with National Physical Laboratory which supported early development. As part of this effort, we also extended the stochastic optimisation framework in the Core Imaging Library (CIL) to incorporate these new algorithms. We are pleased to announce that this work has been accepted for presentation at the 10th International Conference on Scale Space and Variational Methods in Computer Vision (SSVM2025). For more information, we refer to the preprint version https://arxiv.org/abs/2411.