Description
Date depot: 28 septembre 2023
Titre: Similarity Measure Learning for Analogical Transfer
Directrice de thèse:
Marie-Jeanne LESOT (LIP6)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Analogical transfer is the process by which some information about a new situation is inferred directly from a mapping established between the new situation and some analog situations retrieved from memory. To enable a widespread application of transfer methods to real scenarii related to machine learning tasks, the main objective of this Phd thesis is to provide a methodology to learn a similarity measure that is optimized for a given transfer task, developing novel metric learning methods, dedicated to this case. The propositions will be applied to two types of real data, to study their robustness to different applications: the first one deals with the cooking domain and learning recipe transfers, which is a traditional application domain of transfer methods. The second one , which will be developed in collaboration with the APHP (Assistance Publique – Hôpitaux de Paris), deals with the domain of breast cancer management. Indeed, it has been shown that in medicine, analogical transfer is used by practitioners to evaluate the plausibility of decisions when they are confronted to complex cases (e.g., when the usual treatment is not applicable).
Doctorant.e: Fan Chunyang