Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage?

NeurIPS 2024 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Epistemic versus aleatoric decomposition · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows theoretically that reverse-KL evidential deep learning objectives (RPriorNet, Belief Matching, PostNet/NatPN, EDL-MSE) converge to a fixed sample-size-independent target meta distribution, argues EDL acts as an energy-based OOD detector, and proposes bootstrap distillation for better epistemic uncertainty.

Problem

Whether evidential deep learning (EDL) methods learn meaningful epistemic and aleatoric uncertainties, given they reportedly perform well on OOD detection and selective classification despite theoretical pitfalls such as non-vanishing distributional uncertainty.

Approach

A unifying objective shows that many EDL losses (PriorNet, MSE, VI, UCE) reduce to divergence matching against a fixed target meta distribution, so the learned uncertainty is sample-size-independent; EDL is then reinterpreted as an energy-based OOD detector, and a bootstrap-based distillation of model uncertainty is proposed to recover faithful epistemic uncertainty.

Key results

• Theory shows reverse-KL EDL objectives converge to a fixed target Dirichlet Dir(alpha_0 + nu*eta(x)), so epistemic uncertainty does not vanish even with infinite data.
• Bootstrap Distillation performs at or near best on CIFAR10/CIFAR100 OOD detection and selective classification, with EnD2 also strong, especially on CIFAR100.
• Classical EDL methods perform comparably regardless of auxiliary techniques (density parameterization or OOD data).

Measured effects reported in the paper

Surprising findings

• Classical EDL methods achieve comparable OOD detection regardless of density parameterization or OOD-data regularization, and the likelihood choice barely matters.
• Smaller regularization lambda consistently improves OOD detection, even though it makes the learned uncertainty less meaningful.

Limitations

• Bootstrap Distillation requires training multiple models, increasing computational cost.
• Theoretical vanishing of epistemic uncertainty for bootstrap distillation is only argued informally and left as future work.

Experimental setup

Datasets: CIFAR10 CIFAR100 SVHN FMNIST TinyImageNet

Benchmarks: CIFAR10 OOD detection CIFAR100 OOD detection CIFAR10 selective classification CIFAR100 selective classification

Metrics: AUROC AUPR

Baselines: RPriorNet Belief Matching PostNet NatPN EDL (MSE) Fisher-EDL EnD2 S2D

Builds on: Prior Networks Posterior Networks Natural Posterior Network Deep Ensembles Energy-based OOD detection

Closest papers by meaning

• Uncertainty Estimation by Flexible Evidential Deep Learning (NeurIPS 2025 · similarity 0.94)
• Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods? (ICML 2024 · similarity 0.92)
• Uncertainty Estimation by Density Aware Evidential Deep Learning (ICML 2024 · similarity 0.91)
• Calibrating LLMs with Information-Theoretic Evidential Deep Learning (ICLR 2025 · similarity 0.90)
• Uncertainty Regularized Evidential Regression (AAAI 2024 · similarity 0.89)
• Epistemic Uncertainty Quantification For Pre-Trained Neural Networks (CVPR 2024 · similarity 0.89)