Disentangle aleatoric and epistemic uncertainty in F-EDL

Develop a methodology within flexible evidential deep learning (F-EDL) to achieve full disentanglement of aleatoric uncertainty and epistemic uncertainty, surpassing the current variance-based decomposition and yielding structurally separated, interpretable components.

Background

F-EDL introduces label-wise variance-based uncertainty measures and a decomposition into aleatoric and epistemic components. However, the authors note that this decomposition does not fully disentangle the two sources of uncertainty, which is a longstanding challenge in UQ.

A complete disentanglement would enhance interpretability and reliability, especially in complex or noisy scenarios, and could improve downstream tasks such as misclassification and distribution shift detection.