Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild
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 a testing-time evidence filtering mechanism that uses conformal-style p-values and Benjamini-Hochberg thresholds to down-weight unreliable view evidence in trusted multi-view classification, improving robustness to noisy views.
Problem
Trusted multi-view classification (TMVC) methods lose accuracy at test time when views are noisy or corrupted, because the uncertainty estimates for unreliable views are not high enough to suppress their evidence during fusion.
Approach
Frame identification of unreliable view evidence as a multiple hypothesis testing problem: compute p-values of test-view uncertainty against a validation set, select reliable evidence with a BH-style threshold to control false discovery rate, and down-weight flagged evidence by a factor alpha before fusion. The mechanism is plug-in compatible with BCF, A-CBF and ABF fusion rules and existing TMVC
Key results
- On corrupted views with A-CBF fusion on CUB, accuracy rises from 56.67% to 79.17%, roughly a 23-point gain.
- Applying the filter to six TMVC methods improves accuracy under corrupted views on all six datasets, per radar-chart comparison.
- On clean views the method yields mixed results, with small drops on SCENE and CAL (e.g., SCENE ABF 70.94 to 70.69).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add evidence filtering to A-CBF | A-CBF · CUB with corrupted views | ACC (%): 56.67±6.67 -> 79.17±1.67 | Helped |
| Add evidence filtering to BCF | BCF · CUB with corrupted views | ACC (%): 65.33±3.33 -> 83.75±1.25 | Helped |
| Add evidence filtering to ABF | ABF · CUB with corrupted views | ACC (%): 64.17±5.00 -> 80.00±0.83 | Helped |
| Add evidence filtering to ABF | ABF · NUS with corrupted views | ACC (%): 27.70±1.60 -> 37.70±0.85 | Helped |
| Add evidence filtering to A-CBF | A-CBF · SCENE with corrupted views | ACC (%): 50.75±2.00 -> 51.38±1.94 | Helped |
| Add evidence filtering to ABF | ABF · SCENE clean views | ACC (%): 70.94±0.06 -> 70.69±0.31 | Hurt |
| Add evidence filtering to A-CBF | A-CBF · CAL clean views | ACC (%): 97.63±0.38 -> 97.25±0.25 | Hurt |
| Add evidence filtering to BCF | BCF · CUB clean views | ACC (%): 97.08±0.42 -> 97.92±0.42 | Helped |
| Add evidence filtering to ABF | ABF · CAL with corrupted views | ACC (%): 56.75±1.88 -> 76.75±5.00 | Helped |
| Add evidence filtering to A-CBF | A-CBF · HAND clean views | ACC (%): 99.00±0.05 -> 99.00±0.75 | No clear effect |
| Vary the filtering parameter α from 0.005 to 0.100 (ablation) | Evidence filtering with ABF · CAL with corrupted views | ACC (%): 53.75 ± 1.75 -> 53.75 ± 1.75; 55.13 ± 0.38 -> 78.63 ± 1.38 | Mixed |
| Vary the filtering parameter α from 0.005 to 0.100 (ablation) | Evidence filtering with A-CBF · CAL with corrupted views | ACC (%): 69.60 ± 2.70 -> 69.60 ± 2.70; 71.80 ± 1.10 -> 71.90 ± 1.10 | Mixed |
Surprising findings
- Minor accuracy decreases on clean SCENE and CAL data, attributed to down-weighting of higher-uncertainty but informative evidence.
Limitations
- Robustness gains depend on the choice of alpha; clean-view performance can slightly decrease on some datasets.
- Evaluation uses synthetic noise injection (random noise in half of views) rather than naturally corrupted real sensor data.
Experimental setup
Datasets: CUB SCENE CAL HAND PIE NUS
Benchmarks: CUB SCENE CAL HAND PIE NUS
Metrics: accuracy (ACC)
Baselines: TMC TMDOA ETMC SMDC ECML CCML BCF A-CBF ABF
Builds on: Trusted Multi-view Classification (TMC) Benjamini-Hochberg procedure Subjective Logic ETMC ECML
Closest papers by meaning
- Trusted Multi-View Classification with Expert Knowledge Constraints (ICML 2025 · similarity 0.89)
- Enhancing Multi-View Classification Reliability with Adaptive Rejection (AAAI 2025 · similarity 0.87)
- Navigating Conflicting Views: Harnessing Trust for Learning (ICML 2025 · similarity 0.86)
- Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (AAAI 2025 · similarity 0.86)
- Deep Fuzzy Multi-view Learning for Reliable Classification (ICML 2025 · similarity 0.85)
- Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios (ICML 2025 · similarity 0.83)
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.