Unsupervised Imaging Inverse Problems with Diffusion Distribution Matching

Title

Unsupervised Imaging Inverse Problems with Diffusion Distribution Matching

Speaker

Giacomo Meanti - INRIA Grenoble

Speaker: Giacomo Meanti (INRIA Grenoble)

Abstract:

I will talk about image restoration tasks addressed through the lens of inverse problems, having access to only unpaired datasets. The method proposed reduces the assumptions and priors needed with respect to knowledge of the forward model and data. By having access to just unpaired degraded and clean images we will see how to learn unknown forward models effectively. The method leverages conditional flow matching to model the distribution of degraded observations, while simultaneously learning the forward model via a distribution-matching loss that arises naturally from the framework.

Bio:

Giacomo Meanti is a postdoc at Inria Grenoble, working with Julien Mairal’s Thoth team. Right now, he’s all about solving tricky imaging problems and playing with generative models. Before that, he figured out how to make kernel methods cheaper. Prior to joining Inria he earned his PhD in computer science at the University of Genoa, lead by Lorenzo Rosasco.

When: November 3rd, h 15:00

Where: Room 322, DIBRIS Via Dodecaneso 35