Institut de Mathématiques de Toulouse

Accueil > Événements Scientifiques > Séminaires & Groupes de Travail > Groupes de Travail > GdT Mathématiques de l’apprentissage

Mathématiques de l’apprentissage

par Aurélien Garivier, Sébastien Gerchinovitz - publié le , mis à jour le

Ce groupe de travail hebdomadaire est dédié à l’étude mathématique des problèmes et des algorithmes de machine learning. Bien que rattaché à l’équipe-projet AOC qui réunit des membres de l’IMT et de l’IRIT, il est ouvert à toutes les personnes se sentant concernées par cette thématique sur le site toulousain, afin d’échanger des idées et de susciter des collaborations. Nous alternons entre exposés formels, séances de lecture, debriefings post-conférences, ou simples discussions.

* lieu : 1R3 - salle MIP si possible, sinon à voir selon disponibilités.
* fréquence : toutes les semaines le jeudi 12h30-13h30.




  • Lundi 1er avril 12:30-13:30 - Tommaso Cesari - University of Milan

    Cooperative Online Learning : Keeping your Neighbors Updated

    Résumé : We study an asynchronous online learning setting with a network of agents. At each time step, some of the agents are activated, requested to make a prediction, and pay the corresponding loss. The loss function is then revealed to these agents and also to their neighbors in the network. When activations are stochastic, we show that the regret achieved by a network of agents running the standard online Mirror Descent algorithm scales with the independence number of the network rather than with the number of agents. We also show a matching lower bound that holds for any given network. When the pattern of agent activations is arbitrary, the problem changes significantly : we prove a linear lower bound on the regret that holds for any online algorithm oblivious to the feedback source.

    Lieu : Bâtiment 1R3, salle de conférence du premier étage (MIP)


  • Lundi 6 mai 12:30-13:30 - Bezirgen Veliyev - Aarhus University

    Functional sequential treatment allocation

    Résumé : In this paper we study a treatment allocation problem with multiple treatments, in which the individuals to be treated arrive sequentially. The goal of the policy maker is to treat every individual as well as possible. Which treatment is “best’’ is allowed to depend on various characteristics (functionals) of the individual-specific outcome distribution of each treatment. For example measures of welfare, inequality, or poverty. We propose the Functional Sequential Allocation policy, and provide upper bounds on the regret it incurs compared to the oracle policy that knows the best treatment for each individual. These upper bounds increase sublinearly in the number of treatment assignments and we show that this regret guarantee is minimax optimal. In addition, the assumptions under which this regret guarantee is established are as weak as possible—even a minimal weakening of them will imply non-existence of policies with regret increasing sub-linearly in the number of assignments.
    Link for the preprint : https://arxiv.org/pdf/1812.09408.pdf

    Lieu : Salle 106, bâtiment 1R1


iCal