Research map / Label-efficient and robust learning / Noisy labels
Noisy Label Calibration for Multi-View 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 NLC, a noisy-label calibration method for multi-view classification that detects noisy labels via neighbor-based JS-divergence confidence and relabels them, combined with cross-view max-margin ranking and MixUp.
Problem
Multi-view classification degrades when training labels are noisy (instance-dependent noise), since existing multi-view methods memorize noisy labels and cannot separate clean from noisy instances.
Approach
Use the label distributions of reliable neighbors (in a fused multi-view representation) to detect noisy instances via a JS-divergence confidence score and to relabel them, combined with cross-view max-margin ranking loss and MixUp.
Key results
- On UCI at 50% noise, NLC reaches 92.15% accuracy versus 88.90% for the next best method (TMNR).
- On Scene15 at 50% noise, NLC reaches 68.61% versus 65.15% for TMC.
- Ablation at 0.5 noise shows removing label calibration drops UCI accuracy from 92.15% to 78.75%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NLC instead of the strongest listed baseline | Multi-view classification · UCI (50% noise) | accuracy (%): 88.90±0.49 -> 92.15±3.67 | Helped |
| Use NLC instead of the strongest listed baseline | Multi-view classification · Caltech101 (50% noise) | accuracy (%): 72.89±1.97 -> 75.85±2.66 | Helped |
| Use NLC instead of the strongest listed baseline | Multi-view classification · Leaves100 (50% noise) | accuracy (%): 73.81±3.61 -> 75.19±3.23 | Helped |
| Use NLC instead of the strongest listed baseline | Multi-view classification · Scene15 (50% noise) | accuracy (%): 65.15±1.72 -> 68.61±1.74 | Helped |
| Use NLC instead of the strongest listed baseline | Multi-view classification · LandUse21 (50% noise) | accuracy (%): 36.57±2.34 -> 48.29±2.41 | Helped |
| Use NLC instead of the strongest listed baseline | Multi-view classification · ALOI-100 (50% noise) | accuracy (%): 88.64±0.96 -> 94.76±0.71 | Helped |
| Add penalty term Lp (ablation) | NLC · UCI (50% noise) | accuracy (%): 79.80 -> 92.15 | Helped |
| Add cross-view ranking learning (ablation) | NLC · UCI (50% noise) | accuracy (%): 90.70 -> 92.15 | Helped |
| Add label calibration (ablation) | NLC · UCI (50% noise) | accuracy (%): 78.75 -> 92.15 | Helped |
| Add MixUp (ablation) | NLC · UCI (50% noise) | accuracy (%): 91.95 -> 92.15 | Helped |
| Add penalty term Lp (ablation) | NLC · Caltech101 (50% noise) | accuracy (%): 50.23 -> 75.85 | Helped |
| Add cross-view ranking learning (ablation) | NLC · Caltech101 (50% noise) | accuracy (%): 75.60 -> 75.85 | Helped |
Surprising findings
- NLC performance drops sharply at noise rates above 0.7 and remains poor at 0.9, attributed to noisy neighbors.
Limitations
- Performance collapses at very high noise rates (0.9) because neighbors contain too much noise.
- Requires tuning of thresholds theta and theta' (adjusted for noise above 0.5) and loss weights alpha, beta.
Experimental setup
Datasets: UCI Caltech101 Leaves100 Scene15 LandUse21 ALOI-100
Benchmarks: UCI Caltech101 Leaves100 Scene15 LandUse21 ALOI-100
Metrics: classification accuracy
Baselines: DCCAE TMC TMDLO ETMC DCP UIMC ECML TMNR
Builds on: MixUp TMC TMNR DCP
Closest papers by meaning
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.88)
- ROLL: Robust Noisy Pseudo-label Learning for Multi-View Clustering with Noisy Correspondence (CVPR 2025 · similarity 0.88)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 2025 · similarity 0.88)
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.87)
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.86)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (ICML 2024 · similarity 0.86)
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.