Research map / Label-efficient and robust learning / OOD and anomaly detection
Hyper-opinion Evidential Deep Learning for Out-of-Distribution Detection
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.
Proposes Hyper-opinion Evidential Deep Learning (HEDL), which models sharp and vague evidence over a hyper-domain of composite sets and projects hyper-opinions to multinomial opinions to improve OOD detection and avoid the EDL vanishing-gradient problem.
Problem
Evidential deep learning (EDL) only collects evidence for single categories, ignoring evidence for composite sets of in-distribution classes, which degrades OOD detection and, with many classes, causes vanishing gradients in the fully-connected layer that hurt classification accuracy.
Approach
Extend EDL's multinomial opinion to a hyper-opinion over the hyper-domain (powerset of classes), splitting belief into sharp (singleton) and vague (composite-set) mass, then project it to a multinomial opinion via a Heaviside-masked linear map whose detached weights make the gradient path well-conditioned.
Key results
- On CIFAR-10 with SVHN/Textures/Places365 OOD, HEDL reaches average FPR95 15.55% and AUROC 96.27% versus EDL's 16.96% and 95.39%, with 95.66% accuracy.
- On CIFAR-100, HEDL achieves average FPR95 55.14% and AUROC 89.59%, against EDL's 95.94% FPR95 and 76.51% AUROC.
- On CUB-200-2011, HEDL improves average FPR95 to 3.82% and AUROC to 98.91% compared with EDL's 98.03% FPR95 and 75.27% AUROC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HEDL instead of LogitNorm | CIFAR-10 OOD detection · CIFAR-10, average OOD | AUROC: 95.97 -> 96.27 | Helped |
| Use HEDL instead of RED | CIFAR-100 OOD detection · CIFAR-100, average OOD | AUROC: 82.05 -> 89.59 | Helped |
| Use HEDL instead of GEN | Flower-102 OOD detection · Flower-102, average OOD | AUROC: 98.85 -> 99.07 | Helped |
| Use HEDL instead of VIM | CUB-200-2011 OOD detection · CUB-200-2011, average OOD | AUROC: 98.26 -> 98.91 | Helped |
| Add hyper-opinion to multinomial-opinion without opinion projection (ablation) | HEDL · Flower-102 | AUROC: 67.23 -> 97.61 | Helped |
| Add hyper-opinion to multinomial-opinion without opinion projection (ablation) | HEDL · CUB-200-2011 | AUROC: 75.27 -> 97.82 | Helped |
| Add hyper-opinion without opinion projection (ablation) | HEDL · CUB-200-2011 | accuracy: 59.87 -> 52.30 | Hurt |
| Add opinion projection to hyper-opinion (ablation) | HEDL · Flower-102 | AUROC: 97.61 -> 99.07 | Helped |
| Add opinion projection to hyper-opinion (ablation) | HEDL · Flower-102 | accuracy: 81.40 -> 84.13 | Helped |
| Add opinion projection to hyper-opinion (ablation) | HEDL · CUB-200-2011 | accuracy: 52.30 -> 74.62 | Helped |
| Add opinion projection to hyper-opinion (ablation) | HEDL · CUB-200-2011 | AUROC: 97.82 -> 98.91 | Helped |
Surprising findings
- Standard EDL shows zero gradient norms for several fully-connected parameters during training, correlating with very low per-class accuracy on CIFAR-100.
Limitations
- Best performance depends on the pre-trained model being adapted via transfer learning.
- Requires reducing dependence on pre-trained models as future work.
- Evaluated only on image classification datasets with a limited set of OOD benchmarks.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Flower-102 CUB-200-2011
Benchmarks: SVHN Textures Places365
Metrics: AUROC FPR95 AUPR accuracy
Baselines: MSP ODIN openGAN GradNorm VIM KNN DICE RankFeat ASH SHE GEN MCDropout
Builds on: Evidential Deep Learning Subjective Logic OpenOOD Regularized Evidential Distribution (RED)
Closest papers by meaning
- Uncertainty Estimation by Flexible Evidential Deep Learning (NeurIPS 2025 · similarity 0.87)
- ANEDL: Adaptive Negative Evidential Deep Learning for Open-Set Semi-supervised Learning (AAAI 2024 · similarity 0.87)
- Hyper Evidential Deep Learning to Quantify Composite Classification Uncertainty (ICLR 2024 · similarity 0.87)
- Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage? (NeurIPS 2024 · similarity 0.87)
- Uncertainty Estimation by Density Aware Evidential Deep Learning (ICML 2024 · similarity 0.85)
- Towards Evidential and Class Separable Open Set Object Detection (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.