Research map / Label-efficient and robust learning / Noisy labels
Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning
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 Part-Level Multi-labeling (PLM), which crops images into five parts, labels each part with a noisy-trained classifier, and jointly trains a single-to-multiple transition matrix network with the noisy class posterior network to estimate noisy class posteriors and improve noisy label learning.
Problem
Noisy class posteriors estimated by classifiers trained on noisy labels overemphasize label-correlated feature parts that do not reflect instance characteristics, causing large errors in noisy class posterior estimation, which propagates to transition matrix and clean posterior estimation in noisy l
Approach
Augment noisy supervision with part-level multi-labels obtained by cropping each instance into five parts and labeling each part with a noisy-trained network, and link them to the single noisy label via a learned single-to-multiple transition matrix within a classifier-consistent joint training framework.
Key results
- On CIFAR-10 with Sym-50% noise, PLM reaches 85.08% accuracy versus 84.09% for VolMinNet and 74.33% for Forward.
- On Clothing1M, PLM achieves 73.30% accuracy, versus 72.42% for PTD and 71.67% for VolMinNet.
- On CIFAR-100 Pair-45%, PLM reaches 61.90% accuracy, versus 61.21% for VolMinNet; PLM-V reaches 64.05% on Pair-45%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use part-level multi-label supervision with cropping and a single-to-multiple transition matrix instead of CE posterior estimation | PLM noisy class posterior estimator · Synthetic datasets across noise rates [0.1, 0.2, 0.3, 0.4, 0.5] | noisy class posterior estimation error (l2-distance) | Helped |
| Use PLM instead of the strongest listed baseline | Noisy-label classification · MNIST Sym-20% | accuracy: 99.11 ± 0.06 -> 99.32 ± 0.02 | Helped |
| Use PLM instead of the strongest listed baseline | Noisy-label classification · MNIST Sym-50% | accuracy: 98.41 ± 0.08 -> 98.96 ± 0.07 | Helped |
| Use PLM instead of the strongest listed baseline | Noisy-label classification · CIFAR-10 Sym-20% | accuracy: 90.19 ± 0.17 -> 91.10 ± 0.24 | Helped |
| Use PLM instead of the strongest listed baseline | Noisy-label classification · CIFAR-10 Pair-45% | accuracy: 86.12 ± 1.26 -> 91.80 ± 0.63 | Helped |
| Use PLM instead of the strongest listed baseline | Noisy-label classification · CIFAR-100 Sym-20% | accuracy: 67.70 ± 0.99 -> 69.54 ± 0.23 | Helped |
| Use PLM instead of the strongest listed baseline | Noisy-label classification · CIFAR-100 Pair-45% | accuracy: 61.21 ± 2.98 -> 61.90 ± 1.82 | Helped |
| Use PLM instead of the strongest listed baseline | Noisy-label classification · Clothing1M test | accuracy: 72.42 -> 73.30 | Helped |
| Use PLM instead of the strongest listed baseline | Noisy-label classification · CIFAR-10 IDN-50% | accuracy: 72.17 ± 1.51 -> 76.87 ± 1.59 | Helped |
| Combine Forward transition-matrix estimation with PLM | CIFAR-10 and CIFAR-100 noisy-label classification · Symmetric and pair-flip noise settings in Table 4 | accuracy | Helped |
Surprising findings
- Plain Forward and T-Revision perform worse when run with the highest-confidence anchor selection, motivating a fair re-implementation; PLM consistently improves them.
Limitations
- Uniform five-crop cropping criterion for part-level labeling may be too simplistic.
- Does not combine with state-of-the-art methods such as DivideMix or NCR in the main comparison.
Experimental setup
Datasets: MNIST CIFAR-10 CIFAR-100 Clothing1M Animal-10N
Benchmarks: MNIST CIFAR-10 CIFAR-100 Clothing1M Animal-10N
Metrics: classification accuracy l2 estimation error of noisy class posterior
Baselines: Decoupling Co-teaching T-Revision Dual-T DMI VolMinNet Class2Simi Forward PTD TMDNN MEIDTM
Builds on: Forward correction T-Revision Dual-T VolMinNet Class2Simi
Closest papers by meaning
- Estimating Noise Transition Matrix with Label Correlations for Noisy Multi-Label Learning (NeurIPS 2022 · similarity 0.90)
- Dual T: Reducing Estimation Error for Transition Matrix in Label-noise Learning (NeurIPS 2020 · similarity 0.90)
- Part-dependent Label Noise: Towards Instance-dependent Label Noise (NeurIPS 2020 · similarity 0.90)
- Learning the Latent Causal Structure for Modeling Label Noise (NeurIPS 2024 · similarity 0.88)
- Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification (NeurIPS 2026 · similarity 0.87)
- Learning with Open-world Noisy Data via Class-independent Margin in Dual Representation Space (AAAI 2025 · 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.