Turning pass-through network architectures into iterative ones, which use their own output as input, is a well-known approach for boosting performance. In this paper, we argue that such architectures offer an additional benefit: The convergence rate of their successive outputs is highly correlated with the accuracy of the value to which they converge. Thus, we can use the convergence rate as a useful proxy for uncertainty. This results in an approach to uncertainty estimation that provides state-of-the-art estimates at a much lower computational cost than techniques like Ensembles, and without requiring any modifications to the original iterative model. We demonstrate its practical value by embedding it in two application domains: road detection in aerial images and the estimation of aerodynamic properties of 2D and 3D shapes.
@article{arxiv.2403.16732,
title = {Enabling Uncertainty Estimation in Iterative Neural Networks},
author = {Nikita Durasov and Doruk Oner and Jonathan Donier and Hieu Le and Pascal Fua},
journal= {arXiv preprint arXiv:2403.16732},
year = {2025}
}