Conference Paper


Credal wrapper of model averaging for uncertainty estimation in classification

Abstract

This paper presents an innovative approach, called credal wrapper, to formulating a credal set representation of model averaging for Bayesian neural networks (BNNs) and deep ensembles (DEs), capable of improving uncertainty estimation in classification tasks. Given a finite collection of single predictive distributions derived from BNNs or DEs, the proposed credal wrapper approach extracts an upper and a lower probability bound per class, acknowledging the epistemic uncertainty due to the availability of a limited amount of distributions. Such probability intervals over classes can be mapped on a convex set of probabilities (a credal set) from which, in turn, a unique prediction can be obtained using a transformation called intersection probability transformation. In this article, we conduct extensive experiments on several out-of-distribution (OOD) detection benchmarks, encompassing various dataset pairs (CIFAR10/100 vs SVHN/Tiny-ImageNet, CIFAR10 vs CIFAR10-C, CIFAR100 vs CIFAR100-C and ImageNet vs ImageNet-O) and using different network architectures (such as VGG16, ResNet-18/50, EfficientNet B2, and ViT Base). Compared to the BNN and DE baselines, the proposed credal wrapper method exhibits superior performance in uncertainty estimation and achieves a lower expected calibration error on corrupted data.

Attached files

Authors

Wang Kaizheng
Cuzzolin, Fabio
Shariatmadar, Keivan
Moens, David
Hallez, Hans

Oxford Brookes departments

School of Engineering, Computing and Mathematics

Dates

Year of publication: 2025
Date of RADAR deposit: 2025-01-29


Creative Commons License This work is licensed under a Creative Commons Attribution 4.0 International License


Related resources

This RADAR resource is Identical to Credal Wrapper of Model Averaging for Uncertainty Estimation in Classification

Details

  • Owner: Joseph Ripp
  • Collection: Outputs
  • Version: 1 (show all)
  • Status: Live
  • Views (since Sept 2022): 236