Lipschitz properties of neural networks and applications to unrolled proximal algorithms
Jean-Christophe Pesquet, CentraleSupélec, Université Paris-Saclay
26 November 2025
Lipschitz properties of neural networks play a crucial role in quantifying their robustness and establishing convergence guarantees for iterative algorithms that incorporate them. This talk will consist of two parts. The first part will review various analytical and numerical techniques for evaluating the Lipschitz constant of a neural network. The second part will focus on unrolling techniques, a recently emerging paradigm for designing mathematically interpretable neural architectures. Explicit expressions for the Lipschitz constant of an unrolled proximal algorithm used to solve inverse problems will also be presented.