Bonjour — I'm Antoine
I'm broadly interested in mathematical and algorithmic questions in deep learning, especially in the structure and behavior of trained models.
-
Postdoc @ EPFL (with Nicolas Boumal) Dec 2024 – present
-
PhD @ ENS Lyon (with Rémi Gribonval, Elisa Riccietti, Nicolas Brisebarre) 2021 – 2024
What I work on
A first line of my work studies structures that training selects, beyond just fitting the data. Recent examples include diffusion language models learning support before frequencies, and match-and-copy mechanisms learned by Transformers.
Another line of my work studies structure in the parameterization and implementation of neural networks. This includes complexity measures, perturbation bounds, and pruning rules that respect rescaling symmetries, as well as GPU implementations of structured sparse layers.