Epistemic Wrapping for Uncertainty Quantification
Maryam Sultana, Neil Yorke-Smith, Kaizheng Wang, Shireen Kudukkil Manchingal, Muhammad Mubashar, Fabio Cuzzolin
TL;DR
This paper introduces Epistemic Wrapping, a novel method that transforms Bayesian Neural Network outputs into belief function posteriors to improve uncertainty estimation in classification tasks, demonstrating enhanced generalization and reliability.
Contribution
The paper proposes Epistemic Wrapping, a new approach that improves uncertainty quantification in neural networks by converting BNN outputs into belief functions, applicable across various datasets.
Findings
Enhanced uncertainty estimation in classification tasks.
Improved model generalization on benchmark datasets.
Effective transformation of BNN outputs into belief functions.
Abstract
Uncertainty estimation is pivotal in machine learning, especially for classification tasks, as it improves the robustness and reliability of models. We introduce a novel `Epistemic Wrapping' methodology aimed at improving uncertainty estimation in classification. Our approach uses Bayesian Neural Networks (BNNs) as a baseline and transforms their outputs into belief function posteriors, effectively capturing epistemic uncertainty and offering an efficient and general methodology for uncertainty quantification. Comprehensive experiments employing a Bayesian Neural Network (BNN) baseline and an Interval Neural Network for inference on the MNIST, Fashion-MNIST, CIFAR-10 and CIFAR-100 datasets demonstrate that our Epistemic Wrapper significantly enhances generalisation and uncertainty quantification.
Peer Reviews
Decision·Submitted to ICLR 2026
1. The idea of wrapping Bayesian posteriors in belief functions to capture higher-order epistemic uncertainty in the parameter space is quite novel (at least I haven't seen much). 2. The methodology is general well connected and nicely fit with each other. 3. Across datasets and architectures, this methods yields consistent accuracy and OOD gains.
1. While the method combines existing ingredients in an interesting way, several components (belief functions, Dirichlet evidential layers, random-set neural networks, interval networks) have all appeared in prior works. The authors also point out the novel lies in combing these techniques but I think then this is considered as rather incremental work maybe more suitable for a journal paper given nowadays AI conference requires more novel concepts. 2. With so much components is it hard to know
* Epistemic Uncertainty in parameter space * Purely visible improvements in results on public datasets
* There is no information on the speed of the algorithm, especially in comparison to other baselines * No formal derivation on why $\alpha_{k}$ can be expressed through L-moments like it was used in Eq. (6) * No analysis on why INNs to be better Before Fine-Tuning in Table 1 for Random-Selection, and why INNs are good especially for the deep NNs (n = 8) After Fine-Tuning? * Some minor remarks: * Figure 1 is very small * line 130: No clarification of what is "SPDE" * Table 3, for Cifar-10
1. The paper is easy to follow. 2. Potentially, the approach might be interesting.
1. I do not understand the motivation behind the approach in the paper. The authors say that the "current efforts model epistemic uncertainty in the model's target space, rather than its parameter space", which is somewhat true. But this target space is induced by the parameter space. Existing approaches take the parameter space, and I do not clearly understand in what regard the paper under review differs from existing papers. 2. Conceptually, why is it a good idea to make this "discretization
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAI-based Problem Solving and Planning · Risk and Safety Analysis · Anomaly Detection Techniques and Applications