UNLIP - Unsupervised Learning for Non-Linear Inverse Problems

This website contains information about the ANR JCJC project UNLIP.

ANR project website: anr.fr

Dates: November 2023 - November 2027

Abstract Deep neural networks have revolutionized the field of imaging inverse problems, obtaining state-of-the-art performance in a wide range of applications such as medical and astronomical imaging. Deep learning-based image reconstruction techniques are transforming many areas of science, industry, and medicine. The predominant approach for tackling imaging problems with deep learning consists of training neural networks in a supervised way, i.e., using a dataset of pairs of measurements and associated images. Widespread deployment of supervised learning solutions in various scientific and medical imaging applications has been so far limited as: it is often very expensive or even impossible to obtain large datasets of ground-truth signals, and supervised networks can fail to reconstruct structures and patterns which do not appear in the ground-truth training examples. Recent advances have highlighted the possibility to learn from noisy and incomplete measurements alone by minimizing an unsupervised learning loss. These methods can obtain a performance on par with supervised learning and even surpass it, as measurement data is vastly more available than ground truth data. However, most existing unsupervised approaches are limited to linear inverse problems, hindering their use in various practical non-linear imaging settings. This project aims to push the frontiers of unsupervised learning beyond linear imaging problems, paving the way for learning-based solutions trained from quantized, phaseless or uncalibrated measurement data alone. The project will study necessary and sufficient conditions for learning from measurement data alone, and propose new algorithmic solutions based on deep neural networks which will be demonstrated in practical imaging problems, including auto-calibrating magnetic resonance imaging, astronomical polarimetric imaging and coherent diffractive imaging.

Team