The Unreasonable Effectiveness of Deep Evidential Regression

There is a significant need for principled uncertainty reasoning in machine learning systems as they are increasingly deployed safety-critical domains. A new approach with uncertainty-aware regression-based neural networks (NNs), based on evidential distributions aleatoric and epistemic uncertainties, shows promise over traditional deterministic methods typical Bayesian NNs, notably the capabilities to disentangle uncertainties. Despite some empirical success of Deep Evidential Regression (DER), there important gaps mathematical foundation that raise question why proposed technique seemingly works. We detail theoretical shortcomings analyze performance synthetic real-world data sets, showing heuristic rather than an exact quantification. go discuss corrections redefinitions how uncertainties should be extracted from NNs.