Research map / Label-efficient and robust learning / Noisy labels
Label Noise Correction via Fuzzy Learning Machine
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 FLM-LNC, a two-stage Fuzzy Learning Machine framework that scores fuzzy membership of samples to their labels via exemplar sets, filters high-confidence samples, retrains, and corrects noisy labels by exemplar-based prediction.
Problem
Label noise filtering and correction in classification: existing filters rely on raw features or deep networks overfit noisy labels, leading to high noise-identification error and poor generalization on large noisy datasets.
Approach
Uses a Fuzzy Learning Machine (cosine-similarity fuzzy relation network trained with a fuzziness permissible loss) to compute exemplar-based fuzzy membership of each sample to its own label, selects high-confidence samples, retrains on them, and corrects labels by nearest exemplar-set prediction.
Key results
- FLM-LNF reaches F1 of 99.6 (MNIST), 98.9 (CIFAR-10), 99.3 (SVHN) at 5% noise, outperforming RD, vRD, MNN, MVF, CRF and Adap mCRF.
- Label noise correction accuracy is 97.2% and 98.0% on MNIST at 15% and 20% noise, but only 60.5%-63.8% on CIFAR-10.
- Classifier trained on FLM-filtered CIFAR-10 at 5% noise reaches 66.1% vs 61.4% without filtering.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FLM-LNF instead of the compared filtering methods | MNIST label-noise filtering · MNIST, 5%-20% noise | Acc | Helped |
| Train the classifier on data filtered by FLM-LNF rather than the compared filtering methods | MNIST classifier · MNIST, 5%-20% noise | PreAcc | Helped |
| Train the classifier on data filtered by FLM-LNF rather than the compared filtering methods | CIFAR-10 classifier · CIFAR-10, 5%-20% noise | PreAcc | Helped |
| Train the classifier on data filtered by FLM-LNF rather than the compared filtering methods | SVHN classifier · SVHN, 5%-20% noise | PreAcc | Helped |
Surprising findings
- Correction accuracy on MNIST is only 65.4% at 5% noise, lower than at higher noise ratios.
- Several baseline filters reduce classifier accuracy below the unfiltered baseline on CIFAR-10 and SVHN, which the authors attribute to over-filtering.
Limitations
- Correction accuracy on CIFAR-10 is modest (about 60-64%).
- Hyperparameters such as the number of exemplars and threshold gamma are manually specified.
- Evaluated only on synthetic uniform random noise on three image datasets.
Experimental setup
Datasets: MNIST CIFAR-10 SVHN
Benchmarks: MNIST CIFAR-10 SVHN
Metrics: Accuracy NfAcc Precision Recall F1 classification accuracy (PreAcc) label correction accuracy
Baselines: RD vRD MNN MVF CRF Adap mCRF NOF
Builds on: Fuzzy Learning Machine Relative Density filter Complete Random Forest filter
Closest papers by meaning
- FINE Samples for Learning with Noisy Labels (NeurIPS 2021 · similarity 0.84)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.83)
- Which Is More Effective in Label Noise Cleaning, Correction or Filtering? (AAAI 2024 · similarity 0.83)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.82)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.82)
- Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning (CVPR 2024 · similarity 0.82)
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.