Research map / Label-efficient and robust learning / Noisy labels
Self-cognitive Denoising in the Presence of Multiple Noisy Label Sources
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 SDM, which uses neural networks' self-cognition (small-loss memorization) to estimate instance-wise noise and annotator-wise source quality from multiple noisy label sources, mutually denoises a multi-tower MLP, and distills a lightweight single-tower model.
Problem
Learning binary classifiers from multiple noisy label sources without ground truth, where prior methods overlook neural networks' inherent ability to autonomously distinguish noisy samples and sources during training.
Approach
Theoretically shows that networks fit higher-quality noisy sources with smaller loss and predict closer to the truth there, then uses instance-wise small-loss filtering and cosine-similarity-based source quality estimates to weight cross-source mutual denoising, plus a selective distillation into a single-tower model.
Key results
- On class-dependent noise SDM reaches 85.43 AUC on Yelp versus 82.48 for the best baseline CVL (and 80.30 for HE A).
- Distilled single-tower model gθ matches the multi-tower GΘ performance with fewer parameters, consistently beating Major, HE A, SLF, and ADMoE on average over six datasets.
- Ablation on Yelp shows adding instance-wise filter M and annotator-wise quality q raises ensemble AUC from 70.30 to 77.07 under instance-dependent noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · Yelp, class-dependent noise | AUC: 82.48 -> 85.43 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · IMDb, class-dependent noise | AUC: 69.61 -> 75.93 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · AgNews, class-dependent noise | AUC: 75.12 -> 81.39 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · SVHN, class-dependent noise | AUC: 71.19 -> 69.01 | Hurt |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · MNIST, class-dependent noise | AUC: 88.65 -> 87.50 | Hurt |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · Bank, class-dependent noise | AUC: 90.61 -> 91.91 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · Yelp, instance-dependent noise | AUC: 75.85 -> 77.07 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · IMDb, instance-dependent noise | AUC: 66.07 -> 66.50 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · AgNews, instance-dependent noise | AUC: 88.28 -> 89.45 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · SVHN, instance-dependent noise | AUC: 82.67 -> 83.75 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · MNIST, instance-dependent noise | AUC: 98.00 -> 98.17 | Helped |
| use SDM instead of the strongest baseline | learning from multiple noisy label sources · Bank, instance-dependent noise | AUC: 93.41 -> 93.44 | Helped |
Surprising findings
- Several baselines (e.g., Major voting, CrowdLayer) degrade as the number of noisy sources increases, whereas SDM improves with more sources.
Limitations
- Theory and method restricted to binary classification; multi-class extension requires new assumptions.
- Theoretical analysis assumes class-dependent noise, while experiments also use instance-dependent noise where assumptions are broken.
Experimental setup
Datasets: Yelp IMDb AgNews SVHN MNIST Bank
Benchmarks: Yelp IMDb AgNews SVHN MNIST Bank
Metrics: AUC
Baselines: Single Majority voting EBCC DoctorNet CrowdLayer CVL WeaSEL HyperEnsemble SLF ADMoE
Builds on: co-teaching mean teacher (EMA) HyperEnsemble Crowd Layer ADMoE
Closest papers by meaning
- Noise-Robust Learning from Multiple Unsupervised Sources of Inferred Labels (AAAI 2022 · similarity 0.89)
- Understanding Instance-Level Label Noise: Disparate Impacts and Treatments (ICML 2021 · similarity 0.89)
- Understanding Self-Distillation in the Presence of Label Noise (ICML 2023 · similarity 0.89)
- Revisiting Learning with Noisy Labels: Active Forgetting and Noise Suppression (CVPR 2026 · similarity 0.88)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.87)
- Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.87)
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.