Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LITIS, INSA Rouen Normandie, Université de Rouen N
Durée : 6 mois
Contact : aleite@insa-rouen.fr
Date limite de publication : 2027-01-16

Contexte :
Machine-learning models are commonly trained assuming that deployment data follow the training distribution. In practice, changes in populations, sensors, environments, or acquisition conditions produce a mismatch between a source domain PS and a target domain PT . Domain adaptation seeks to transfer predictive knowledge across such shifts, while uncertainty quantification (UQ) estimates the reliability of a model’s predictions. In particular, a source-trained model may be more uncertain about target inputs that are poorly represented in its training data. Yet domain shift and uncertainty are not equivalent: a shift need not increase prediction difficulty, and uncertainty can be high even without a shift. This internship will study when predictive uncertainty can meaningfully reveal, measure, or be corrected for a change of domains.

Sujet :
This internship aims to establish a conceptual and empirical framework connecting predictive uncertainty, distribution shift, and domain adaptation.

Profil du candidat :
Master’s 2 or final-year engineering level; strong foundations in machine learning and probability/statistics; ability to implement reproducible experiments and critically analyze results. Familiarity with uncertainty quantification, transfer/domain adaptation, or probabilistic learning is appreciated but not required. Interest in both theoretical reasoning and rigorous empirical research is important.

Formation et compétences requises :

Adresse d’emploi :
Rouen Normandie

Document attaché : 202610201148_LITIS_CRAN_internship_proposal_From Predictive Uncertainty to Domain Adaptation_onepage.pdf