Research

Research topics

Learning for inverse problems. Unrolling iterative thresholding algorithms yields neural networks with identifiability guarantees and practical acceleration. I study unrolled dictionary learning and learning priors without ground truth.

Sparsity, optimization and Bayesian approaches. Structured sparsity (mixed norms, iterative shrinkage), ℓ0 problems and Bayesian sparse models such as Bernoulli-Gaussian priors.

Applications, from audio to neuroimaging. Audio: underdetermined source separation in reverberant mixtures, sparse and low-rank models, music source separation with interpretable networks. Neuroimaging: M/EEG source localization, and MEG analysis for drug-resistant epilepsy.

Publications: see the Publications page.

Projects

  • PEPR SantéNum Optiscale (2026–2030). Local coordinator. PI: C.-G. Bénar
  • ANR BMWs – Bayesian Meets Wavelets (2020–2025). Local coordinator. PI: B. Torrésani

Earlier: LIDL (Digicosme, 2020–2023, with T. Moreau), SW2PB (Digicosme, 2019–2020, with L. Liberti), SESAM (Digiteo, 2014–2017), IMPULSE (PICS CNRS, 2015–2018), MAORI (PGMO, 2013–2016).

PhD students

Current

  • Dylan Sechet (2025–). Music source separation with interpretable neural networks. Co-supervised with M. Évrard.
  • Jad Yehya (2025–). Rare event detection with local pattern modeling. Co-supervised with T. Moreau.
  • Ismail Labiad (2025–). New training strategies for large language models. Co-supervised with M. Schoenauer, J. Kempe and R. Munos.
  • Maxence Dubernay (2026–). Self-supervised deep learning for spatio-temporal brain signals: methods and applications to MEG in epilepsy.
  • Joachim Jobard (2026–). Learning HJB solutions for continuous-time reinforcement learning. Co-supervised with A. Shilova.

Former