|
چکیده
|
Evidential Deep Learning (EDL) enables single-pass uncertainty estimation by predicting Dirichlet evidence, which is attractive for practical deployment where multi-pass inference is costly. However, many EDL approaches can remain overconfident under distribution shift, exhibit imperfect calibration, and provide limited expressiveness for multi-modal epistemic uncertainty especially near decision boundaries or when inputs deviate subtly from the training support. This thesis proposes Gated Evidential Mixtures (GEM), a family of single-pass uncertainty-aware classifiers that explicitly couples predictive confidence to representation-space support. The central idea is to learn an internal energy-like support signal end-to-end and use it to gate evidential outputs, encouraging strong evidence for well-supported ID inputs while suppressing unjustified evidence for weak-support and OOD-like inputs. The framework is developed incrementally to enable controlled ablation and clarify the contribution of each component. First, GEM-CORE learns a feature-level energy signal and maps it to a bounded integration gate that smoothly modulates evidential strength as support decreases. Second, to represent epistemic multi-modality without multi-pass ensembling, GEM-MIX introduces a lightweight mixture of evidential heads with learned routing weights, preserving single-pass inference while improving uncertainty expressiveness. Third, GEM-FI stabilizes mixture allocations using a Fisher-information–informed regularization/modulation mechanism, mitigating expert (head) collapse and improving uncertainty behavior in sensitive regions such as near decision boundaries. The proposed approach is evaluated across image classification and out-of-distribution (OOD) detection benchmarks under far-OOD and near-OOD shifts, as well as corruption-based distribution shifts. The results indicate that GEM maintains competitive in-distribution (ID) accuracy while substantially improving confidence reliability and ID/OOD separability. In particular, support-aware gating improves calibration by reducing overconfident errors under low support, mixture modeling enhances epistemic separation, and Fisher-informed stabilization yields more robust routing and more consistent uncertainty quality. Overall, GEM provides a practical single-pass framework for deployment-oriented uncertainty estimation, achieving consistent gains over strong EDL baselines on metrics including accuracy, Brier score, AUROC, and AUPR.
|