Vous êtes ici : FR > Nos domaines d'action > Doctorat

ED 512 - InfoMaths (Informatique et Mathématiques)
Publié le 4 novembre 2026 | Mis à jour le 4 novembre 2026

Bayesian Transfer Learning on Manifolds: Adversarial Domain Adaptation

Bayesian models of brain ageing must not only predict how anatomical structures change with age, but also quantify how confident those predictions are, an especially delicate requirement when the model is trained on one cohort and deployed on another with a different age range, scanner, or acquisition protocol, since a poorly calibrated transfer can silently understate uncertainty exactly where a clinician or researcher most needs to trust it.
When the model carried from a source to a target ageing cohort is no longer a linear map but a deep network or a stochastic process, transporting its parameters along the manifold geometry is not enough to guarantee that the transfer helps rather than harms the target task