Abstract
Evidential Deep Learning (EDL) has emerged as a widely adopted framework for quantifying uncertainty in classification tasks by modeling a higher-order prior on classification probabilities using a Dirichlet distribution. Recent advances in EDL have notably progressed, primarily by calibrating the predicted Dirichlet distribution to enhance uncertainty estimation. However, these methods often overlook the role of feature representations, the fundamental components of deep learning models that directly impact both classification performance and uncertainty estimation. This paper adopts a feature-centric perspective to explore the relationship between predicted evidence and feature representations in EDL. We demonstrate that existing EDL approaches struggle to learn sufficiently discriminative features as the number of classes increases, which significantly undermines their capacity to produce reliable uncertainty estimates. To tackle this issue, we propose Tangential Gradient Descent (TGD), a novel feature optimization method that explicitly decomposes the loss gradient with respect to feature representations into tangential and radial components. By accelerating updates along the tangential component, TGD promotes feature alignment to a simplex Equiangular Tight Frame (ETF), thereby maximizing inter-class angular separability for enhanced classification. At the same time, updates along the radial component foster discriminative feature projections that effectively differentiate between certain and uncertain samples, leading to improved uncertainty estimation. Extensive experiments on benchmarks such as CIFAR-10, CIFAR-100, and ImageNet-200, covering both OOD detection and generalization across various class scales, reveal that our approach consistently surpasses conventional EDL methods in both uncertainty estimation and scalability.