Research map / 3D vision / 3D generation
Reliable Conflictive Multi-View 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 Reliable Conflictive Multi-View Learning (RCML) and an Evidential Conflictive Multi-view Learning (ECML) method that outputs decisions plus reliabilities for conflictive multi-view data via evidential DNNs and a conflictive opinion aggregation rule.
Problem
Multi-view classification where views may be noisy or unaligned (conflictive instances); existing methods eliminate such instances, whereas applications need decisions with calibrated reliabilities for them.
Approach
View-specific evidential DNNs produce Dirichlet opinions (belief, uncertainty); a conflictive opinion aggregation rule averages evidence and makes uncertainty grow when views conflict, with a conflict-minimizing consistency loss during training.
Key results
- On conflictive test sets, ECML reaches 84.00% accuracy on PIE versus 68.16% for the best baseline TMDL-OA (+23.24%).
- On conflictive Scene15, ECML achieves 56.97% vs 48.42% for TMDL-OA (+17.66%).
- On normal test sets, ECML is best on all six datasets, e.g. 90.84% on HMDB vs 88.20% for TMDL-OA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ECML instead of TMDL-OA | Multi-view classification · HandWritten normal test set | accuracy (%): 99.25±0.45 -> 99.40 ± 0.00 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · CUB normal test set | accuracy (%): 95.43±0.20 -> 98.50 ± 2.75 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · HMDB normal test set | accuracy (%): 88.20±0.58 -> 90.84 ± 1.86 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · Scene15 normal test set | accuracy (%): 75.57±0.02 -> 76.19 ± 0.12 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · Caltech101 normal test set | accuracy (%): 94.63±0.04 -> 95.36 ± 0.38 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · PIE normal test set | accuracy (%): 92.33±0.36 -> 94.71 ± 0.02 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · HandWritten conflictive test set | accuracy (%): 93.05±0.05 -> 94.40 ± 0.05 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · CUB conflictive test set | accuracy (%): 74.43±0.26 -> 76.50 ± 1.15 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · HMDB conflictive test set | accuracy (%): 67.62±0.28 -> 70.84 ± 1.19 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · Scene15 conflictive test set | accuracy (%): 48.42±1.02 -> 56.97 ± 0.52 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · Caltech101 conflictive test set | accuracy (%): 90.63±2.05 -> 92.36 ± 1.48 | Helped |
| Use ECML instead of TMDL-OA | Multi-view classification · PIE conflictive test set | accuracy (%): 68.16±0.34 -> 84.00 ± 0.14 | Helped |
Surprising findings
- Existing opinion aggregation methods reduce uncertainty after fusing another opinion; the authors argue this is wrong when views conflict.
Limitations
- Conflictive test sets are synthetically constructed (Gaussian noise and random view replacement), not real-world conflicts.
- Uncertainty ground truth is unavailable, so reliability is evaluated only indirectly via constructed conflicts.
Experimental setup
Datasets: HandWritten CUB HMDB Scene15 Caltech101 PIE
Benchmarks: HandWritten CUB HMDB Scene15 Caltech101 PIE
Metrics: accuracy uncertainty distribution
Baselines: DCCAE CPM-Nets DUA-Nets TMC TMDL-OA
Builds on: Evidential Deep Learning Trusted Multi-View Classification subjective logic Dempster-Shafer combination
Closest papers by meaning
- Navigating Conflicting Views: Harnessing Trust for Learning (ICML 2025 · similarity 0.89)
- Deep Fuzzy Multi-view Learning for Reliable Classification (ICML 2025 · similarity 0.86)
- Enhancing Multi-View Classification Reliability with Adaptive Rejection (AAAI 2025 · similarity 0.85)
- Biased Incomplete Multi-View Learning (AAAI 2025 · similarity 0.83)
- Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (AAAI 2025 · similarity 0.83)
- Robust Multi-view Clustering against Imperfect Information (NeurIPS 2026 · similarity 0.82)
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.