Research map / Label-efficient and robust learning / Recognition and long-tail
Navigating Conflicting Views: Harnessing Trust for Learning
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 instance-wise, probability-sensitive trust discounting (TF/ETF) on top of Evidential multi-view classification, using a learned referral evidence network and BCF fusion, to resolve view conflicts.
Problem
Multi-view classification suffers when views conflict (some views express distinct or misaligned information), causing confident but wrong fused predictions in evidential fusion frameworks.
Approach
Adds a separate referral evidential network that outputs a per-view, per-instance degree of trust, used to discount each view's functional opinion via subjective-logic trust discounting before Belief Constraint Fusion, trained with a stage-wise schedule.
Key results
- ETF achieves average Top-1 accuracy of 90.22% across six datasets, versus 88.74% for ETMC and 86.60% for TMC.
- ETF reaches Fleiss' Kappa 0.48 on Scene15 vs 0.37 for ETMC-style baselines, and 0.65 on HMDB where ETMC scores -0.15.
- On Food101 end-to-end training, ETF attains 93.09% Top-1 accuracy versus 92.49% for ETMC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trust discounting (TD) (ablation) | ETF · Scene15 | Top-1 accuracy: 81.06±0.16 -> 82.01±0.17 | Helped |
| add trust discounting (TD) (ablation) | ETF · Scene15 | Fleiss' Kappa: 0.46±0.01 -> 0.48±0.01 | Helped |
| add trust discounting (TD) (ablation) | TF · Scene15 | Top-1 accuracy: 76.82±0.33 -> 77.83±0.32 | Helped |
| add trust discounting (TD) (ablation) | TF · Scene15 | Fleiss' Kappa: 0.37±0.01 -> 0.39±0.00 | Helped |
| vary warm-up label smoothing factor from 0.9 to 1.0 (ablation) | ETF · Scene15 | Top-1 accuracy: 82.01±0.17 -> 82.07±0.12 | No clear effect |
| vary warm-up label smoothing factor from 0.9 to 0.8 (ablation) | ETF · Scene15 | Fleiss' Kappa: 0.48±0.01 -> 0.49±0.01 | No clear effect |
| add pseudo-view incorporation | TF · PIE | Top-1 accuracy: 93.31±0.40 -> 94.63±0.34 | Helped |
| add pseudo-view incorporation | TF · Scene15 | Top-1 accuracy: 77.83±0.32 -> 82.01±0.17 | Helped |
| add trust discounting to ETMC | ETMC · PIE | Top-1 accuracy: +2.94% | Helped |
| add trust discounting to ETMC | ETMC · Scene15 | Top-1 accuracy: +3.60% | Helped |
| add trust discounting to TMC | TMC · Scene15 | Top-1 accuracy: +3.84% | Helped |
| add trust discounting to TMC | TMC · PIE | Top-1 accuracy: +5.88% | Helped |
Surprising findings
- Applying trust discounting only to existing functional opinions at test time does not improve cross-view consistency; gains come from incorporating trust discounting during training.
- ETMC has negative Fleiss' Kappa (-0.15) on HMDB, indicating worse-than-chance view agreement.
Limitations
- Performance on CUB drops in AUROC (TF 53.52) likely due to small training set (480 instances) and unstable optimization.
- Requires a multi-stage training schedule with warm-up, adding training complexity.
- TMNR cannot be run in end-to-end training because it needs frozen pre-extracted features.
Experimental setup
Datasets: Handwritten Caltech101 PIE Scene15 HMDB CUB Food101
Benchmarks: Handwritten Caltech101 PIE Scene15 HMDB CUB Food101
Metrics: Top-1 accuracy Fleiss' Kappa MVAGT AUROC
Baselines: MGP ECML TMNR CCML TMC ETMC
Builds on: Trusted Multi-View Classification (TMC) Evidential Deep Learning subjective logic trust discounting Belief Constraint Fusion ETMC
Closest papers by meaning
- Reliable Conflictive Multi-View Learning (AAAI 2024 · similarity 0.89)
- Deep Fuzzy Multi-view Learning for Reliable Classification (ICML 2025 · similarity 0.87)
- Enhancing Multi-View Classification Reliability with Adaptive Rejection (AAAI 2025 · similarity 0.87)
- Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild (CVPR 2025 · similarity 0.86)
- Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (AAAI 2025 · similarity 0.85)
- Trusted Multi-View Classification with Expert Knowledge Constraints (ICML 2025 · 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.