Research map / Label-efficient and robust learning / Recognition and long-tail
Beyond Magnitude: Scale-Invariant Evidential Fusion for Multi-View Classification
Wei Liu, Yufei Chen, Jie Shi, Xiaodong Yue
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add instance-wise standardization of logits (ablation) | SAEF · AVE (semantic hijacking) | conflict detection AUROC: 14.06 -> 93.55 | Improved |
| use Stouffer fusion instead of Sum or Mean (ablation) | SAEF · AVE (sensor noise) | noise rejection AUROC: 27.21 (Sum) / 28.19 (Mean) -> 87.03 | Improved |
| use Stouffer fusion instead of Max (ablation) | SAEF · AVE | clean accuracy: 80.10 -> 82.84 | Improved |
| instance standardization instead of batch norm (ablation) | SAEF · AVE | conflict detection AUROC: 88.19 -> 93.55 | Improved |
| Stouffer fusion instead of Sum/Mean on clean accuracy (ablation) | SAEF · AVE | clean accuracy: 83.57 (Sum), 83.66 (Mean) vs 82.84 | No clear effect |
| vary canonical scale beta from 1 to 20 (ablation) | SAEF · AVE | clean accuracy: 83.66 -> 82.41 | No clear effect |
| vary noise floor epsilon up to 0.1 (ablation) | SAEF · AVE | clean accuracy: 82.84 for all epsilon up to 0.1 | No clear effect |
| replace SAEF with canonical fusion under logit scale perturbation of one view | DS/Sum/Mean/Weighted fusion · SUNRGBD | accuracy: 48% -> 24% | Hurt |
| use SAEF instead of canonical fusion | multi-view fusion · SUNRGBD | accuracy: 46.03 -> 48.17 | Improved |
| use SAEF instead of Mean fusion | multi-view fusion · inference runtime | ms per sample: latency difference bounded within about 0.1ms | No clear effect |
Closest papers by meaning
- Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (AAAI 2025 · similarity 0.87)
- Navigating Conflicting Views: Harnessing Trust for Learning (ICML 2025 · similarity 0.86)
- Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild (CVPR 2025 · similarity 0.85)
- Deep Fuzzy Multi-view Learning for Reliable Classification (ICML 2025 · similarity 0.84)
- Hyper Evidential Deep Learning to Quantify Composite Classification Uncertainty (ICLR 2024 · similarity 0.84)
- Uncertainty Estimation by Flexible Evidential Deep Learning (NeurIPS 2025 · similarity 0.84)
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.
