Fiche de KIRCHMEYER Matthieu

Titre : Out-of-Distribution Generalization in Classification and Dynamics Forecasting
Date : 16 mai 2023
Heure: 14:00
Résumé : Deep learning has emerged as a powerful approach for modelling static data like images and more recently for modelling dynamical systems like those underlying times series, videos or physical phenomena. Yet, neural networks were observed to not generalize well outside the training distribution, in other words out-of-distribution. This lack of generalization limits the deployment of deep learning in autonomous systems or online production pipelines, which are faced with constantly evolving data. In this thesis, we design new strategies for out-of-distribution generalization on the tasks of classification of static data and dynamics forecasting of spatiotemporal systems. In the first part of this work, we investigate how we can efficiently leverage some observed data from a target domain for adaptation. In a second part, we explore how to generalize to unobserved domains without access to such data. Both parts handle the specific challenges posed by our two main application tasks.
Lieu : Sorbonne Université - 4 place Jussieu 75005 Paris. Salle 304- couloir 65-66, 3e étage.