Évènements

Modern statistical learning problems often involve heterogeneous, high-dimensional, or distributed data, raising two complementary challenges: controlling model complexity in high dimensions, and enabling frugal inference under distributional constraints, while preserving both structure and statistical consistency.

In this talk, I present contributions centered on mixture-of-experts (MoE) models, a flexible latent variable framework with well-established approximation capabilities and learning guarantees for conditional densities.

I first consider MoE models with high-dimensional predictors, including functional data such as curves and time series, and their training via Lasso-type regularization in unsupervised settings, enabling sparse and interpretable representations together with non-asymptotic model selection guarantees.

I then address learning from data distributed across multiple sites due to storage, computational, or governance constraints, and present a frugal aggregation strategy for MoE models based on optimal transport, which constructs a reduced global estimator from locally trained models in a single communication round while preserving both model structure and statistical consistency, making it particularly well suited to large-scale settings where communication is a major bottleneck.

Together, these contributions position MoE as a unified framework for principled, interpretable, and scalable statistical learning, with perspectives on frugal learning of generative models at the interface of modern machine learning and AI, and applications in constrained industrial environments.

[exposé en français.]

I will present a construction of the Euclidean sine–Gordon quantum field theory using probabilistic methods. The approach formulates the model through a stochastic control problem and a weak forward–backward stochastic differential equation (FBSDE). The starting point is a detailed study of the perturbation theory associated with the sine–Gordon potential across the full subcritical regime.  For a portion of the subcritical regime extending above 6π, the construction can be made rigorous beyond perturbation theory using a stochastic control problem. I will explain the different regimes of the model and why we currently cannot construct the model in the full subcritical regime.

En 1999, Wen Chao Lu a donné une démonstration par l’analyse réelle du théorème des nombres premiers avec terme d’erreur, “à epsilon près” celui obtenu un siècle auparavant par La Vallée Poussin au moyen de l’analyse complexe. En 2024, Gozé a, dans sa thèse, donné une version quantitative de ce résultat.

Dans un travail en cours avec Gozé et Bruno Martin, nous reprenons les démonstrations de Lu et Gozé, et tentons d’en dégager les idées essentielles. L’exposé présentera sous une forme simple certaines d’entre elles.