Research map / Label-efficient and robust learning / Noisy labels
Which Is More Effective in Label Noise Cleaning, Correction or Filtering?
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.
Compares label correction and filtering via a rebuilt error bound under label noise, then proposes FCF, a two-stage fusion that corrects high-noise-probability labels and filters unconfident ones, reaching SOTA on noisy CIFAR-10/100.
Problem
Choosing between label correction and label filtering for cleaning noisy labels is unresolved; their effectiveness, applicability, and hyper-parameter sensitivity have not been systematically compared from an error-bound perspective.
Approach
Rebuilds a generalization error bound for label noise that depends on noise rate, then uses it to compare correction and filtering at dataset and sample levels; combines them in a fusion framework (FCF) where correction runs first on high-noise-probability labels and filtering removes remaining unconfident samples, with noise probabilities estimated by a beta mixture model over label confidences.
Key results
- FCF achieves 93.7% on CIFAR-10 with 20% symmetric noise and 94.4% with 20% asymmetric noise, outperforming all listed baselines.
- On CIFAR-100 with 40% symmetric noise, FCF reaches 69.8% versus 69.7% for SFT and 68.7% for MSLC-style correction-only ablation at 68.7%.
- Ablation on CIFAR-100 shows correction alone (68.7%) beats filtering alone (65.1%) at 40% symmetric noise, and FCF reaches 69.8%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine label correction followed by filtering rather than use no cleaning (ablation) | FCF on CIFAR-100 · symmetric noise 20%, 40%; asymmetric noise 20%, 40% | accuracy: 47.6 -> 74.3; 45.3 -> 69.8; 51.9 -> 76.3; 37.7 -> 73.4 | Helped |
| add correction and filtering rather than correction alone (ablation) | FCF on CIFAR-100 · symmetric noise 20%, 40%; asymmetric noise 20%, 40% | accuracy: 72.1 -> 74.3; 68.7 -> 69.8; 72.7 -> 76.3; 69.3 -> 73.4 | Helped |
| add correction and filtering rather than filtering alone (ablation) | FCF on CIFAR-100 · symmetric noise 20%, 40%; asymmetric noise 20%, 40% | accuracy: 69.4 -> 74.3; 65.1 -> 69.8; 70.1 -> 76.3; 65.5 -> 73.4 | Helped |
| use correction instead of filtering at the same cleaning size | label noise cleaning · datasets with close cleaned-dataset noise rates | error bound | Helped |
| correct labels with large noise probabilities instead of filtering them | label noise cleaning · samples with pi > pT | error bound | Helped |
| filter samples with unconfident median noise probabilities instead of correcting them | label noise cleaning · samples with 0.5 < pi < pT | error bound | Helped |
| use filtering rather than correction when cleaning size exceeds the number of noises | label noise cleaning · over-cleaning | error bound | Helped |
Surprising findings
- Theory implies that filtering is more tolerant of imperfect cleaning-size hyper-parameters, while correction is more effective when noise rates are close.
- Filtering removes correctly labeled samples (over-cleaning), and wrong corrections can raise the noise rate in later stages.
Limitations
- Theoretical analysis assumes known noise probability and noise rate, while real cleaning uses estimated noise probabilities.
- Experiments are limited to CIFAR-10/100 with synthetic noise and a single backbone (ResNet-34).
- Theory considers binary classification with random label noise, while experiments are multi-class.
Experimental setup
Datasets: CIFAR-10 CIFAR-100
Benchmarks: CIFAR-10 CIFAR-100
Metrics: test accuracy label purity
Baselines: Standard Bootstrap Joint-Opt Co-teaching MW-Net SELFIE AdaCorr TopoFilter MSLC FINE SFT CTO
Builds on: Co-teaching MW-Net TopoFilter FINE MSLC
Closest papers by meaning
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.85)
- Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.85)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.84)
- Enhancing Sample Selection Against Label Noise by Cutting Mislabeled Easy Examples (NeurIPS 2025 · similarity 0.84)
- How Much is a Noisy Image Worth? Data Scaling Laws for Ambient Diffusion. (ICLR 2025 · similarity 0.84)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.84)
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.