Le mercredi 17 juin à 10h30 en salle B206 se tiendra un séminaire SOP où nous aurons le plaisir d’écouter un exposé de Yohan Petetin !

From Kalman to diffusion models.

In a first part, I revisit linear state-space models on which the popular Kalman filter relies. I show that it is possible to derive a more general class of linear state – space models which have some invariant properties. In these models, the computation of the posterior distributions of interest remain feasible.

In a second part, I show that this generalization can benefit to diffusion models. I generalize the framework of the Denoising Diffusion Implicit model (DDIM) and propose a non Markovian diffusion model based on latent random variables.