- Cet évènement est passé
Séminaire IMAGE : Multi-domain translation method with little supervision, applied to medical scans and face generation (Tsiry Mayet)
11 avril / 14:00 - 15:30
Nous aurons le plaisir d’accueillir Tsiry Mayet, de l’INSA Rouen.
Il donnera un séminaire IMAGE, le jeudi 11 avril 2024, à 14h00, en salle de séminaire F-200.
Titre: Multi-domain translation method with little supervision, applied to medical scans and face generation.
Résumé:
Multi-domain translation is the task of learning a translation between any partition of domains within a set (such as (D1, D2) → D3, D2 → (D1, D3), D3 → D1, etc. for 3 domains) without the need to train separate models for each configuration.
The more domains considered, the harder it is to achieve full supervision, especially when human intervention is required to label data.
Semi-supervised multi-domain translation involves a setting where arbitrary domains may be missing during training.
For this purpose, we introduce Multi-Domain Diffusion (MDD). The purpose of this method is twofold: i) to be able to reconstruct any missing views for any new data object and ii) to learn in a semi-supervised context where any configuration of supervision is possible.
This is done in Multi-Domain Diffusion (MDD) by exploiting the noise formulation in diffusion models and modeling one noise level per domain. Similar to existing approaches, MDD learns the translation between any partition of domains. Unlike existing approaches, our formulation handles semi-supervised context without modification by modeling missing views as noise in the diffusion process.
The more domains considered, the harder it is to achieve full supervision, especially when human intervention is required to label data.
Semi-supervised multi-domain translation involves a setting where arbitrary domains may be missing during training.
For this purpose, we introduce Multi-Domain Diffusion (MDD). The purpose of this method is twofold: i) to be able to reconstruct any missing views for any new data object and ii) to learn in a semi-supervised context where any configuration of supervision is possible.
This is done in Multi-Domain Diffusion (MDD) by exploiting the noise formulation in diffusion models and modeling one noise level per domain. Similar to existing approaches, MDD learns the translation between any partition of domains. Unlike existing approaches, our formulation handles semi-supervised context without modification by modeling missing views as noise in the diffusion process.
