Two-Sample Testing with Sliced Wasserstein

Cet évènement est passé.
Two-Sample Testing with Sliced Wasserstein

Quand

5 novembre 2026    
14 h 00 min - 16 h 00 min

Où

Room 02 124 Building 5 St Priest campus
860 Rue de St - Priest, Montpellier, Hérault, 34090 , Occitanie

Machine Learning in Montpellier, Theory & Practice – Nicolas Schreuder (CNRS (DI ENS))

Given two datasets, how can we tell whether they come from the same probability distribution? This classical problem, known as two-sample testing, arises in many applications, from scientific data analysis to the evaluation of generative models. In this talk, I will introduce the two-sample testing problem and discuss how distances between probability distributions can be used to construct tests, with a particular focus on permutation-based calibration. I will then discuss the choice of the distance, from kernel-based approaches to optimal transport.
Finally, I will present a two-sample test based on the Sliced Wasserstein distance. The test has finite-sample Type I error control and achieves the minimax separation rate over broad classes of distributions, while remaining computationally efficient. I will conclude with numerical experiments and open questions on the comparison between Sliced Wasserstein and kernel-based tests.

Carte non disponible