Banner

About

I am a PhD student in machine learning and optimization in Grenoble. I have the honour of having the amazing trio Franck Iutzeler, Jérôme Malick and Panayotis Mertikopoulos as adviors. More precisely, I am at the LJK lab, which is part of UGA. I had the chance to study at ENS Paris and to graduate from the MVA master.

See arXiv, Google Scholar, DBLP, Github, LinkedIn for more information.

Contact

Research

My current interest are robust optimization, non-convex stochastic optimization and understanding LLM internal mechanisms.

Stochastic optimization in deep learning

In this line of work, we focus on a simple yet fundamental question: what is the long-run behaviour of stochastic gradient descent (SGD) on non-convex objectives? In the first part, we seek to describe the asymptotic distribution of SGD on general non-convex objectives. Leveraging large deviation theory, we obtain a description of the invariant measure of SGD (ICML 2024, poster). This work was presented at Thoth seminar (slides), at the Séminaire de Statistique of the LPSM lab in Paris (slides) and at the Séminaire de Probabilités et Statistiques of the Nice University (slides).

In the second part, we focus on estimating the time it takes for SGD to reach the global minimum of a non-convex function. This reveals an intricate interplay between the loss landscape, the noise structure and the behavior of SGD (ICML 2025, poster).

Internal mechanisms of Large Language Models

I had the chance of working with Michael Kirchhof, Eugene Ndiaye, Louis Bethune, Michal Klein, Pierre Ablin and Marco Cuturi while interning at Apple Machine Learning Reserach. We studied how uncertainty estimates for LLMs behave under distribution shifts (R2FM Wrokshop@ICML 2025).

Wasserstein Distributionnally Robust Optimization

Inspired by the success of entropic regularization in optimal transport, we study the regularization of WDRO (ESAIM COCV). We also show that these estimators enjoy attractive generalization guarantees (NeurIPS 23, slides).

I presented early versions of these works at a workshop in Erice in May 2022, (slides), and the second part at FOCM 2023 in Paris, (poster) as well at Neurips@Paris 2023 (slides).

Last-iterate convergence of mirror methods

We characterize the last iterate convergence rate of mirror methods in variational inequalities as a function of the local geometry of the Bregman divergence near the solution, both in the deterministic (to be published in SIOPT) and stochastic settings (COLT 21).

The latter was presented at COLT 21 (slides, poster) and at ICCOPT 22 (slides) while the former was presented at SMAI MODE 2024 (slides).

Graph Neural Networks

With Marc Lelarge, we precisely describe the approximatyion cabapilities of invariant and equivariant graph neural networks (ICLR 21). It was presented at a MIPT-UGA workshop and at the Thoth team seminar (slides).

Smooth game optimization for Machine Learning

With Gauthier Gidel, Ioannis Mitliagkas and Simon Lacoste-Julien, we propose a tight and unified analysis of gradient-based methods in games (AISTATS 20, slides) and leverage matrix iteration theory to study accelerated methods in games (AISTATS 20).

Teaching