Epistemic Uncertainty

Bias contamination in uncertainty estimates

Epistemic uncertainty estimation methods are fundamentally incomplete

We argue that current epistemic uncertainty methods are fundamentally incomplete: unaccounted bias contaminates aleatoric estimates, and existing methods capture only partial variance contributions.

Faithfulness violations in evidential deep learning

Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods?

We show that evidential deep learning methods face fundamental optimisation difficulties and that their epistemic uncertainty estimates do not satisfy basic faithfulness properties.