Foundation models for fMRI and brain–behaviour prediction in the MIND team at Inria Saclay, and foundation and normative models of brain anatomy.
Amortized brain–behaviour prediction with Prior-Data Fitted Networks. A transformer meta-trained on synthetic regression tasks built from real FC matrices predicts behavioural scores from a labelled context in one forward pass.
An open-access fMRI timeseries collection for self-supervised training of brain foundation models. It is built entirely from public OpenNeuro studies, uniformly preprocessed with a single pipeline, and released as parcellated ROI timeseries and FC matrices for several atlases, with dataset-level provenance.
Joint work with G. Marraffini and D. Wassermann.
Pre-training on large anatomical MRI datasets to learn brain representations that transfer to many downstream tasks, and modelling the healthy population to detect and follow deviations from it.
Contrastive representation learning, debiasing, and medical imaging applications.