How AI Reshapes Inverse Problems in Imaging

Pierre WEISS
PhD, CNRS Researcher, CBI – CNRS
March 14th, 2024

Abstract
Inverse problems consist in reconstructing signals from incomplete information. They are widespread in applications, ranging from biomedical imaging (e.g. super-resolution microscopy, MRI) to image generation (e.g. Dall-E). In this talk, I plan to review some core ideas that are currently being explored to learn and use efficient image priors. In particular, I will show how standard Bayesian estimators (MAP, MMSE, posterior sampling), which previously yielded disappointing results, are now shining when coupled with carefully trained neural networks.

Supplementary Materials
Slides: https://www.math.univ-toulouse.fr/~weiss/presentations/learning_inverse_problems.pdf
Code: https://github.com/deepinv/deepinv