Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Fuzzy Multi-view Learning for Reliable Classification
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 FUML, a fuzzy-set-theory multi-view classifier that models network outputs as memberships, learns category credibility (possibility plus necessity) with a tailored loss, and fuses views via a dual-reliable (uncertainty and conflict) weighting for robust, reliable classification under conflicting views.
Problem
Evidential deep learning based trusted multi-view classification is sensitive to conflicting views, and its uncertainty estimates underestimate uncertainty for conflicting multi-view instances because they ignore conflicts between belief masses.
Approach
Replace evidence/Dempster-Shafer modeling with fuzzy memberships whose category credibility averages possibility and necessity; estimate entropy-based uncertainty from credibility; fuse views in one step with weights from view uncertainty and inter-view conflict (Dual-reliable Multi-view Fusion).
Key results
- On normal test sets FUML outperforms all 12 baselines on all eight datasets, e.g. +1.62% on Scene, +1.59% on Leaves, +1.47% on PIE over second-best.
- On conflicting test sets FUML leads with gains of 4.83% (Scene), 7.31% (LandUse), 14.60% (Leaves) over second-best.
- On Fashion conflicting-set OOD identification, FPR95 drops to 0.510 vs 0.930 (ETMC) and 0.967 (ECML).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FUML instead of the strongest reported baseline | multi-view classification · Normal test sets across eight datasets | accuracy: ∆0.20; ∆0.95; ∆1.45; ∆0.01; ∆1.62; ∆0.97; ∆1.59; ∆1.47 | Helped |
| Use FUML instead of the strongest reported baseline | multi-view classification · Conflicting test sets across eight datasets | accuracy: ∆0.08; ∆2.38; ∆3.51; ∆3.96; ∆4.83; ∆7.31; ∆14.60; ∆4.01 | Helped |
| Add necessity measure to category credibility (ablation) | FUML · Fashion normal test set | accuracy: 97.70±0.41 -> 98.96±0.25 | Helped |
| Add necessity measure to category credibility (ablation) | FUML · Fashion conflicting test set | accuracy: 94.91±0.47 -> 96.68±0.32 | Helped |
| Add necessity measure to category credibility (ablation) | FUML · LandUse normal test set | accuracy: 44.64±3.65 -> 76.71±0.46 | Helped |
| Add necessity measure to category credibility (ablation) | FUML · LandUse conflicting test set | accuracy: 39.00±3.29 -> 69.14±2.43 | Helped |
| Use DRF instead of arithmetic mean fusion (ablation) | FUML · Fashion conflicting test set | accuracy: 96.15±0.22 -> 96.68±0.32 | Helped |
| Use DRF instead of arithmetic mean fusion (ablation) | FUML · LandUse conflicting test set | accuracy: 67.71±2.30 -> 69.14±2.28 | Helped |
| Include view uncertainty in DRF weights (ablation) | FUML · Fashion conflicting test set | accuracy: 96.27±0.22 -> 96.68±0.32 | Helped |
| Include view conflict in DRF weights (ablation) | FUML · LandUse conflicting test set | accuracy: 68.50±2.02 -> 69.14±2.28 | Helped |
| Use DRF instead of concatenating view features (ablation) | FUML · Fashion conflicting test set | accuracy: 88.45 -> 96.68 | Helped |
| Use DRF instead of concatenating view features (ablation) | FUML · LandUse conflicting test set | accuracy: 59.69 -> 69.14 | Helped |
Surprising findings
- ECML's sequential evidence-averaging fusion is order-dependent, making its decisions sensitive to the order of conflicting views.
- Existing EDL-based methods' uncertainty barely increases on conflicting test sets and even decreases on LandUse, indicating underestimated uncertainty.
Limitations
- Conflicting test sets are synthetically constructed via Gaussian noise and random view label alteration rather than natural conflicts.
- Evaluation is restricted to relatively small multi-view benchmarks using shallow fully connected backbones.
- Hyperparameter p (norm order) must be tuned; performance peaks for p within 2-5.
Experimental setup
Datasets: Handwritten (HW) MSRC-V1 NUS-WIDE-OBJ Fashion-MV Scene15 LandUse Leaves100 PIE
Benchmarks: HW MSRC-V1 NUS-WIDE-OBJ Fashion-MV Scene15 LandUse Leaves100 PIE
Metrics: accuracy precision F-score FPR95 uncertainty density
Baselines: DFTMC DCP QMF PDF DUA-Nets TMC ETMC TMDL-OA UIMC ECML TMNR CCML
Builds on: Evidential Deep Learning Trusted Multi-view Classification (TMC) Enhanced TMC (ETMC) Evidential Conflictive Multiview Learning (ECML) Dempster-Shafer theory
Closest papers by meaning
- Enhancing Multi-View Classification Reliability with Adaptive Rejection (AAAI 2025 · similarity 0.87)
- Navigating Conflicting Views: Harnessing Trust for Learning (ICML 2025 · similarity 0.87)
- Reliable Conflictive Multi-View Learning (AAAI 2024 · similarity 0.86)
- Hyper Evidential Deep Learning to Quantify Composite Classification Uncertainty (ICLR 2024 · similarity 0.86)
- Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (AAAI 2025 · similarity 0.85)
- Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild (CVPR 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.