Réunion d'hiver SMC 2025
Toronto, 5 - 8 decembre 2025
Mathématiques de l'apprentissage automatique
Org: Ben Adcock (Simon Fraser University), Ricardo Baptista (University of Toronto) et Giang Tran (University of Waterloo)
Org: Ben Adcock (Simon Fraser University), Ricardo Baptista (University of Toronto) et Giang Tran (University of Waterloo)
- ISAAC GIBBS, University of California, Berkeley
- AVI GUPTA, Simon Fraser University
- MOHAMED HIBAT-ALLAH, University of Waterloo
- SPENCER HILL, Queen’s University
- ANASTASIS KRATSIOS, McMaster University
- SOPHIE MORIN, Polytechnique Montreal
- RACHEL MORRIS, Concordia University
- CAMERON MUSCO, University of Massachusetts Amherst
- ESHA SAHA, University of Alberta
- MATTHEW THORPE, Warwick University
- ALEX TOWNSEND, Cornell University
- YUNAN YANG, Cornell University
Training Distribution Optimization in the Space of Probability Measures [PDF]
- AVI GUPTA, Simon Fraser University
-
A central question in data-driven modeling is: from which probability distribution should training samples be drawn to most effectively approximate a target function or operator? This work addresses this question in the setting where “effectiveness” is measured by out-of-distribution (OOD) generalization accuracy across a family of downstream tasks. We formulate the problem as minimizing the expected OOD generalization error, or an upper bound thereof, over the space of probability measures. The optimal sampling distribution depends jointly on the model class (e.g., kernel regressors, neural networks), the evaluation metric, and the target map itself. Building on this characterization, we propose two adaptive, target-dependent data selection algorithms based on bilevel and alternating optimization. The resulting surrogate models exhibit significantly improved robustness to distributional shifts and consistently outperform models trained with conventional, non-adaptive, or target-independent sampling across benchmark problems in function approximation, operator learning, and inverse modeling.