Design an intrinsically stable training objective for epistemic control

Design an intrinsically stable training objective for flexible evidential deep learning (F-EDL) that controls epistemic uncertainty without relying on external regularization mechanisms, thereby improving theoretical soundness and practical robustness.

Background

The paper discusses critiques of EDL objectives and notes that, although F-EDL empirically mitigates some issues, it still requires external regularization to control epistemic uncertainty. This indicates a gap in the intrinsic stability of the training objective.

An intrinsically stable objective would reduce dependence on ad hoc regularization, address theoretical concerns raised in prior work, and potentially yield more reliable epistemic behavior across diverse settings.