Research map / Label-efficient and robust learning / Recognition and long-tail
ULAREF: A Unified Label Refinement Framework for Learning with Inaccurate Supervision
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 ULAREF, a unified label refinement framework that globally detects unreliable predictions via GMM-based entropy and locally enhances them with a consistency-regularized fine-tuned model, evaluated on noisy-label and partial-label learning.
Problem
Learning with inaccurate supervision (noisy labels and partial labels), where models overfit to mistaken or ambiguous annotations and specialized algorithms are needed for each annotation form.
Approach
A single label refinement pipeline that splits training data into reliable/unreliable sets using entropy scores modeled by a two-component GMM, then fine-tunes an enhanced model on reliable data and its neighborhood with a KL-based consistency loss (mixup-style manifold regularization), and mixes its outputs into the refined labels; theory shows the enhanced model can be more accurate on the unrel
Key results
- On CIFAR-100 with 0.4 asymmetric noise, ULAREF exceeds the second-best method by 3.76% accuracy.
- ULAREF reaches 93.72% on CIFAR-10 uniform partial labels with gamma=0.3, the best among compared PLL methods.
- ULAREF achieves 74.78% on Clothing1M and 77.80% on WebVision, the best among compared noisy-label methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ULAREF instead of the strongest listed baseline | Noisy-label learning · CIFAR-100 asymmetric noise 0.4 | accuracy: 72.13 -> 76.11 | Helped |
| Use ULAREF instead of the strongest listed baseline | Noisy-label learning · CIFAR-10 symmetric noise 0.2 | accuracy: 95.69 -> 95.28 | Hurt |
| Use ULAREF instead of the strongest listed baseline | Noisy-label learning · CIFAR-10 symmetric noise 0.5 | accuracy: 94.37 -> 94.31 | Hurt |
| Use ULAREF instead of the strongest listed baseline | Noisy-label learning · CIFAR-10 symmetric noise 0.8 | accuracy: 90.82 -> 91.47 | Helped |
| Use ULAREF instead of the strongest listed baseline | Noisy-label learning · CIFAR-10 asymmetric noise 0.4 | accuracy: 92.23 -> 92.56 | Helped |
| Use ULAREF instead of the strongest listed baseline | Noisy-label learning · Clothing1M | accuracy: 74.42 ± 0.27% -> 74.78 ± 0.14 | Helped |
| Use ULAREF instead of the strongest listed baseline | Noisy-label learning · Webvision | accuracy: 77.37 ± 0.27% -> 77.80 ± 0.24 | Helped |
| Use ULAREF instead of the strongest listed baseline | Partial-label learning · CIFAR-10 uniform partial labels 0.3 | accuracy: 93.13 0.15 -> 93.72 0.06 | Helped |
| Use ULAREF instead of the strongest listed baseline | Partial-label learning · CIFAR-10 instance-dependent partial labels | accuracy: 86.43 0.23 -> 87.16 0.10 | Helped |
| Use ULAREF instead of the strongest listed baseline | Partial-label learning · CIFAR-100 uniform partial labels 0.1 | accuracy: 65.27 0.18 -> 67.93 0.24 | Helped |
| Use ULAREF instead of the strongest listed baseline | Partial-label learning · BirdSong | accuracy: 74.23 ± 0.17% -> 74.72 ± 0.24% | Helped |
| Use ULAREF instead of the strongest listed baseline | Partial-label learning · Yahoo!News | accuracy: 66.62 ± 0.19% -> 67.12 ± 0.14% | Helped |
Limitations
- Not best on CIFAR-10 with symmetric noise at 0.2 and 0.5 (second place to DivideMix or ELR+).
- Theoretical guarantee relies on mild but strong cluster and consistency assumptions.
- Hyperparameters (tau_s, T_s, T_e, beta, lambda) require tuning, though sensitivity is reported as stable.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Clothing1M WebVision Lost BirdSong MSRCv2 Soccer Player Yahoo!News
Benchmarks: CIFAR-10 CIFAR-100 Clothing1M WebVision Lost BirdSong MSRCv2 Soccer Player Yahoo!News
Metrics: classification accuracy pairwise t-test significance
Baselines: CE MIXUP CO-TEACHING DIVIDEMIX ELR ELR+ ADACORR PROMIX SOP IDGP PLCR PICO
Builds on: DivideMix mixup manifold regularization PRODEN
Closest papers by meaning
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.89)
- Collaborative Refining for Learning from Inaccurate Labels (NeurIPS 2024 · similarity 0.87)
- Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label Configurations (NeurIPS 2024 · similarity 0.87)
- Controller-Guided Partial Label Consistency Regularization with Unlabeled Data (AAAI 2024 · similarity 0.86)
- Self Iterative Label Refinement via Robust Unlabeled Learning (NeurIPS 2025 · similarity 0.86)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 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.