B1413
Title: Langevin sampling for median of means based estimation
Authors: Stephane Chretien - NPL (United Kingdom)
Wenjuan Sun Stephane Chretien - National Physical Laboratory - University of Lyon 2 (United Kingdom) [presenting]
Abstract: Median of Means (MoM) estimators have attracted a lot of interest lately due to their ability to circumvent being corrupted by outliers and heavy tailed training data. However, designing fast algorithms for computing MoM estimators is still an interesting challenge both in practice and theory. We will show how the Langevin sampler can be put to work for MoM estimation. We show that the Langevin sampler reaches neighborhoods of the set of stationary points of the expected risk in polynomial time. We will also present experimental results for XCT reconstruction using deep learning based ``learned gradient'' methods.
