Extend F-EDL to regression

Extend the flexible evidential deep learning (F-EDL) framework, which models uncertainty in classification by predicting a flexible Dirichlet distribution over class probabilities, to regression tasks by constructing an evidential regression variant that provides calibrated uncertainty estimates for continuous targets (e.g., by adapting evidential regression formulations).

Background

The paper introduces F-EDL, a single-forward-pass uncertainty quantification framework for classification that predicts parameters of a flexible Dirichlet distribution. While demonstrating strong results across multiple classification scenarios, the approach is currently scoped to discrete-label settings.

The authors explicitly identify the extension to regression as an open challenge, pointing to evidential regression models as a potential foundation. Addressing this would broaden the applicability of F-EDL to continuous outputs while preserving its efficiency and principled uncertainty modeling.