arxivmath.OCcs.LG2026-06-29
A Distributionally Robust Framework for Learned Reconstructions in Inverse Problems
Floor van Maarschalkerwaart, Subhadip Mukherjee, Christoph Brune, Marcello Carioni
Learned reconstruction operators for inverse problems are typically trained under a fixed noise model, and generalize poorly when the distribution during testing differs from the one assumed during training. Distributionally robust optimization (DRO) addresses this by optimizing…