Machine learning · optimization · scientific inference
Building reliable machine learning from data and structure
At Inria THOTH, I develop principled learning methods that combine data with prior knowledge—such as structure, simulators, and implicit constraints—for reliable scientific discovery and decision-making.
Research directions
Learning with implicit constraints
Methods for learning when part of the problem is specified by structure, optimality conditions, physical knowledge, or dynamical systems.
Bilevel optimization and scientific inference
Reliable optimization tools that combine deep learning with simulations and expert knowledge for scientific applications.
Predictive representation learning
Theory and algorithms for self-supervised learning that recover useful representations with explicit objectives and guarantees.
Short biography
I am currently a Research Scientist (Chargé de recherche) at Inria Grenoble – Rhône-Alpes, within the THOTH team. My research focuses on representation learning, optimization, and learning with structure and simulations. I received an ERC Starting Grant in 2026 for the DELPHI project.
Previously, I was a Starting Research Fellow in the same team, working with Julien Mairal. I completed my PhD in 2021 at the Gatsby Computational Neuroscience Unit, University College London, under the supervision of Arthur Gretton.
Prospective students and postdocs
I welcome outstanding candidates who enjoy both mathematical depth and empirical work.
See current opportunities and application guidance to learn about research fit, formal openings, and how to make a focused application.
News
| October 2026 | New work on PEIRA, a predictive self-supervised learning method, accepted at NeurIPS 2026. |
|---|---|
| October 2026 | Research update: BlockFormer and ProxiMAP join our NeurIPS 2026 work; flow-matching calibration for simulation-based inference appeared at ICML 2026. |
| September 2026 |  I received an ERC Starting Grant for the DELPHI project, which is scheduled to start in October 2027.
|
| August 2026 | Publication roundup: EquiTabPFN, learning theory for kernel bilevel optimization, and MAP estimation with denoisers appeared at NeurIPS 2025; LUDVIG appeared at ICCV 2025. |
| February 2025 | New work on foundation models for tabular data. |
| February 2025 | New work on learning theory for kernel bilevel optimization. |
| December 2024 | Our work on functional bilevel optimization received a spotlight at NeurIPS 2024. |
| March 2024 |  Official start of the BONSAI project funded by ANR JCJC.
|