I am a researcher in the MIND team @ INRIA Saclay, since autumn 2019, working on machine learning for physiological signals and on method evaluation in AI.
My work sits at the intersection of Machine Learning, Signal Processing, and Optimization, with a focus on deep learning methods for inverse problems, unsupervised learning and foundation models for time series, and bilevel optimization.
Before joining MIND, I did a postdoc in the Parietal team working with Alexandre Gramfort, and obtained my PhD at ENS Paris Saclay under the supervision of Nicolas Vayatis and Laurent Oudre.
I maintain several large-scale open-source projects for scientific computing: joblib, loky, and cloudpickle for parallel and distributed computing, and benchopt for reproducible benchmarks in machine learning. I have also contributed to a wide range of projects, from general-purpose ones like cpython and scikit-learn, to more research-oriented ones like braindecode/moabb, deepinv, and sbi.
Finally, I also co-organize scientific events at the intersection of AI and science, including the Séminaire Palaisien — a monthly seminar I co-founded, now in its 7th year — and a thematic semester at the AISSAI institute bringing together workshops on machine learning, signal processing, and HPC
Latest publication and projects
Geometry-Aware Discretization Error of Diffusion Models 2026
Samuel Hurault, Thomas Moreau, Gabriel Peyrépreprint ArXiv
Practical diffusion sampling is a numerical approximation problem: under a fixed inference budget, one must simulate a reverse-time ODE or SDE using only a limited number of denoising steps, so discretization error is often the dominant source of error. Existing non-asymptotic analyses provide convergence guarantees, but are typically too loose and too insensitive to diffusion parameters to guide practical design: broad families of schedules receive the same rates, which depend on ...
Practical diffusion sampling is a numerical approximation problem: under a fixed inference budget, one must simulate a reverse-time ODE or SDE using only a limited number of denoising steps, so discretization error is often the dominant source of error. Existing non-asymptotic analyses provide convergence guarantees, but are typically too loose and too insensitive to diffusion parameters to guide practical design: broad families of schedules receive the same rates, which depend on coarse worst-case quantities such as the dimension or the drift Lipschitz constant. We take a less ambitious but more informative route. In the exact-score setting, we derive first-order asymptotic expansions of the Euler-Maruyama weak and Fréchet discretization errors. These formulas hold for general smooth reverse diffusions and become fully explicit under Gaussian data. They show how discretization error adapts to the geometry of the data through the covariance spectrum, and how this geometry interacts with key diffusion parameters, including the diffusion schedules and the diffusion-term coefficient. This yields tractable objectives for geometry-aware parameter optimization. Finally, we show that the qualitative predictions of the Gaussian formulas remain robust across diffusion sampling problems with different geometries, including image generation on different datasets and image posterior sampling.
AI benchmarking infrastructures: lessons from Benchopt 06 May 2026,
At Imaging inverse problems and generating models workshop, ICMS, Edinburgh
Research and development in modern AI are primarily driven by empirical work, benchmarking new methods to evaluate relative performance. However, the statistical variability inherent in evaluation processes and long term maintainance objectives aer often poorly accounted for, leading to a validation crisis in which genuine advances are difficult to discern. This talk presents Benchopt, a framework designed to facilitate reproducible, reusable and extendable benchmarking of optimization and ...
Research and development in modern AI are primarily driven by empirical work, benchmarking new methods to evaluate relative performance.
However, the statistical variability inherent in evaluation processes and long term maintainance objectives aer often poorly accounted for, leading to a validation crisis in which genuine advances are difficult to discern.
This talk presents Benchopt, a framework designed to facilitate reproducible, reusable and extendable benchmarking of optimization and AI algorithms, which is a key component in addressing these crisis while also accounting for the need for frugality in modern AI research.
Loky Apr 2023
The aim of this project is to provide a robust, cross-platform and cross-version implementation of the ProcessPoolExecutor class of concurrent.futures.
The aim of this project is to provide a robust, cross-platform and cross-version implementation of the ProcessPoolExecutor class of concurrent.futures. It features:
Deadlock free implementation: one of the major concern in standard multiprocessing and concurrent.futures libraries is the ability of the Pool/Executor to handle crashes of worker processes. This library intends to fix those possible deadlocks and send back meaningful errors.
Consistent spawn behavior: All processes are started using fork/exec on POSIX systems. This ensures safer interactions with third party libraries.
Reusable executor: strategy to avoid respawning a complete executor every time. A singleton pool can be reused (and dynamically resized if necessary) across consecutive calls to limit spawning and shutdown overhead. The worker processes can be shutdown automatically after a configurable idling timeout to free system resources.
06 May 2026,
