Gdr IASIS: Advances in learning-based image restoration

I recently attended a day organized by the GdR IASIS, focused on advances in learning-based image restoration. This event felt like a perfect match for my research interests, sitting at the intersection of applied mathematics and machine learning. As one speaker pointed out, what were once separate communities have now converged, with everyone working on Bayesian statistics, optimization, Markov chains, and deep denoisers.While reading papers for my thesis, I was surprised to discover that many of the leading contributors in this field are French. Most of the major authors in my bibliography were present, and the talks were of a very-high quality.

Among the talks I really enjoyed was the one by Jean-Christophe Pesquet, about the Lipschitz properties of neural networks, which is of particular interest to design convergent Plug-and-Play algorithms. The talk of Pierre Weiss (which was supposed to be given by Hai Nguyen) was also very intereseting, about the differences between Plug-and-Play and unfolded methods, and between MMSE and MAP estimators. I think it was a summary of Hai’s latests work. I loved the talk of Pierre Chainais about HQS like updates for a sort of Langevin algorithm, I think I’ve already read the paper somehow. Then Andres Almansa gave us an overview of Posterior sampling algorithms, from PnP-Langevin algorithms reminiscent of the work of Rémi Laumont, to modern diffusion models, to his recent work with Diffusion EM for blind inverse problems.