Bilevel optimization for machine learning

par Mathieu Dagréou

mardi 31 mars 2026, 11:15 → 12:15 Europe/Paris

Salle K. Johnson (1R3, 1er étage)

Bilevel problems are optimization problems characterized by a hierarchical structure. In these problems, one seeks to minimize an outer function subject to the constraint that certain variables solve an inner optimization problem. These problems are gaining popularity in the machine learning community due to their wide range of applications, including hyperparameter optimization and data reweighting. In this talk, we introduce bilevel optimization and show how various machine learning problems can be formulated within this framework. We then focus on the algorithmic aspects of solving bilevel problems. Specifically, we present a general algorithmic framework that enables the adaptation of first-order stochastic solvers—originally designed for single-level problems—to the bilevel setting. We provide theoretical guarantees for specific instances of this framework. We conclude by presenting recent work on optimal transport under group fairness constraints, where bilevel optimization arises as one possible formulation.