Fuzzy Evidential Learning: Structured Evidence Generation for Uncertainty-Aware Vision Networks

Reliable uncertainty estimation is critical for deploying deep learning models in settings where inputs may be ambiguous or out-of-distribution (OOD). Although evidential deep learning (EDL) provides a principled framework for modelling Dirichlet distributions over class probabilities, existing formulations rely on unstructured evidence representations that struggle to capture ambiguity and often require delicate regularisation to avoid overconfidence. In this work, we introduce ‘Fuzzy Evidential Learning’, a novel framework that integrates fuzzy logic with the subjective reasoning underpinning EDL. We propose a fuzzy evidential block that replaces the conventional classifier with a structured prototype-based fuzzy inference mechanism, where learnable rules enable local feature-space regions to contribute non-negative, class-wise evidence that naturally parameterises a Dirichlet distribution. By instantiating this framework within two existing methods (Fisher Information-based EDL ($\mathcal{I}$-EDL) and Relaxed EDL (Re-EDL)) we demonstrate that this can be integrated into existing EDL variants without modifying their training objectives. Experiments on standard benchmarks (MNIST, CIFAR-10) and OOD settings (FMNIST, SVHN, CIFAR-100, Food-101) show that the proposed method improves in-distribution performance and enhances OOD detection, particularly under far distribution shifts. For $\mathcal{I}$-EDL, structured fuzzy aggregation yields consistent improvements in both AUPR and AUROC. In contrast, for Re-EDL, the fuzzy block degrades performance under near or semantically similar OOD conditions, suggesting that smoothing induced by the fuzzy block can lead to overconfidence when not strongly regularised. Although performance decreases, this behaviour reveals a trade-off between local smoothing and class separability in near-OOD regimes, where structured evidence aggregation can blur decision boundaries between closely related classes. We further show that deep ensembling amplifies these gains to state-of-the-art performance, with fuzzy models benefiting most from increased ensemble diversity. These results establish fuzzy evidential learning as a flexible and effective approach for robust uncertainty estimation in deep neural networks.