My research lies at the interface between optimization, statistics and machine learning. I am currently interested in the long-run behavior of stochastic gradient methods and in distributionally robust learning. Before moving to Toulouse in 2023, I was an Assistant Professor (Maître de Conférences) at Université Grenoble Alpes.
I have a (wonderful) small child, which means that I do not travel much at the moment and may decline most invitations.
With a finite training budget, stochastic gradient descent (SGD) outputs a random model rather than a minimizer of the training loss. Using large deviations theory, we characterize the probability distribution of this model and the time needed to reach a global minimizer.
Wasserstein distributionally robust optimization trains models that perform well on all probability distributions close to the empirical one. We study its statistical guarantees and its regularization, and we develop the Python library skwdro.
Marie Boyer started her PhD, co-supervised with Xiaoyi Mai (Univ. Toulouse Jean Jaurès). Her project is centered around bias amplification in statistical learning.
Our project SO-LAR (StatOptim – Landscape Analysis of Risks), led by P. Escande, is funded by the ANR as a single-team project (PRME).
Congratulations to Waïss Azizian for defending his stellar PhD! The jury included Francis Bach, John Duchi, Gabriel Peyré, Niao He, Anatoli Juditsky, Julien Mairal and his co-advisors: J. Malick, P. Mertikopoulos, and I.
Iyad Zekhnini started his PhD. He is co-supervised with Jérôme Malick and Tung Le (CNRS & LJK, Grenoble) and works on stochastic optimization for non-convex machine learning objectives.
My Tremplin-ERC project DROPS is funded by the ANR (awarded because my ERC Consolidator Grant proposal was rated A at step 2 but not funded).
Our project MadLearning, led by F. Malgouyres, is funded by the PEPR IA program.
Older news
I am now in charge of the Statistics and Optimization team of the lab.
FocusTech ANITI on skwdro: here are the presentation and code (zip).
I am a co-PI of an ANITI (Toulouse's AI cluster) chair on Trust and Responsibility in Artificial Intelligence led by J.-M. Loubes and J. Bolte.
I am taking the responsibility of the second year of the Master SID in Data Science & Engineering.
Our library skwdro has a v1! We provide efficient code for machine learning with Wasserstein distributional robustness. The project is still under development, but this version is a first step towards popularizing this approach. Available on GitHub and pip.
Our project MAD (Maths of Automatic Differentiation), led by S. Vaiter (CNRS & Univ. Côte d'Azur), is funded by the ANR!
Yu-Guan Hsieh's PhD thesis received the Université Grenoble Alpes PhD award!
I am the happy father of Louise! This also means that I will be travelling less in the coming months (maybe years?).
Congratulations to Yu-Guan Hsieh for his fantastic PhD! The jury included Constantinos Daskalakis, Sylvain Sorin, Nicolò Cesa-Bianchi, Alexandre d'Aspremont, Anatoli Juditsky, Maryam Kamgarpour, and his co-advisors: J. Malick, P. Mertikopoulos, and I.
After seven great years in Grenoble, I have moved to Toulouse!
Congratulations to Gilles Bareilles for passing his PhD with flying colors! The jury included Nadia Brauner, Jalal Fadili, Claudia Sagastizábal, Jean-Charles Gilbert, Mathurin Massias, Claude Lemaréchal and his co-advisors: J. Malick and I.
Waïss Azizian started his PhD. He is co-supervised by Panayotis Mertikopoulos and Jérôme Malick (CNRS & LJK, Grenoble) and works on non-convex optimization for learning problems.
Two interns have joined the team: Waïss Azizian (student at ENS Paris, for his third internship with us!) and Julien Prando (student at UGA). Both work on regularization aspects of distributionally robust methods.
The second “Journée SMAI MAS-MODE” was a success: 7 talks and around 40 participants (in person!) at the intersection between optimization, probability, and statistics. Website
I defended my Habilitation à Diriger des Recherches (HDR) on December 15th! The jury was composed of Alexandre d'Aspremont (reviewer), Jérôme Bolte (reviewer), Adrian Lewis (reviewer), Jalal Fadili, Adeline Leclercq-Samson (president), and Julien Mairal. Manuscript, defense slides.
Waïss Azizian (student at ENS Paris) visited our team again for six weeks, as part of his (previously online) internship, to work on distributionally robust methods.
Congrats to Dmitry “Mitya” Grishchenko for successfully defending his PhD! The jury included Pascal Bianchi, Peter Richtárik, Julien Mairal, Samuel Vaiter, Alexander Gasnikov and his co-advisors: J. Malick, M.-R. Amini, and I.
Waïss Azizian (student at ENS Paris) visited our team for six weeks, as part of his (previously online) internship, to work on extragradient methods.
I was awarded an ANR JCJC grant (young researcher grant of the French Research Agency) for the project STROLL: Harnessing Structure in Optimization for Large-scale Learning.
I am part of the chair on Optimization and Learning, led by Jérôme Malick and Yurii Nesterov, funded by the Grenoble AI Institute (MIAI).
Mathias Chastan began his PhD between STMicroelectronics and our team. He is co-supervised by A. Lam, J. Malick and I, and works on machine learning for industrial defect detection.
Around distributional robustness and generalization in machine learning, Journée du département SI de l'IRIT, Toulouse (France)
A study of the long-run behavior of Stochastic Gradient Descent via large deviations, MAD-Stat seminar, Toulouse School of Economics, Toulouse (France)
What is the long-run distribution of SGD?, JS2O, Perpignan (France)
What is the long-run distribution of SGD?, Colloquium de Mathématiques, Avignon (France)
What is the long-run distribution of SGD?, Séminaire Mathématiques de l'IA, Bordeaux (France)
