Séminaire IMAGE: « Self-Supervised Learning, Relational Learning and Multi-Domain Diffusion for Image Generation Under Constraints » (Romain Hérault, GREYC).

Pour ce séminaire de rentrée, nous aurons le plaisir d’écouter Romain Hérault, professeur UFR Sciences UNICAEN qui fait partie de l’équipe IMAGE du GREYC, à Caen / France.

Il donnera un séminaire IMAGE, le jeudi 11 janvier 2024, à 14h00, en salle de séminaire F-200.

In a first part, we will cover self-supervised learning and relational learning. Self-supervised representation learning, facilitated by soft contrastive learning, enables pretraining neural networks without labels, enhancing downstream task performance with minimal annotations.

We also introduces a novel approach to sample negative examples using OCSVM.

In a second part, we wil focus on image generation under constraints. Conditioning Generative Adversarial Networks through auxiliary tasks allows explicit control over the content of generated images. The talk delves into domain-transfer tasks, specifically translating color images to the polarimetric domain with hard physics-based constraints. The cyclic-consistency approach is employed, extending generative model training with handcrafted tasks to enforce constraints.

The discussion further explores Multi-Domain Diffusion (MDD), a conditional diffusion framework for semi-supervised multi-domain translation. MDD facilitates learning in various supervision configurations, utilizing noise formulation to shift from a basic reconstruction task to a domain translation task. Results on a challenging multi-domain synthetic image translation dataset with semantic domain inversion will be presented.