Research map / Label-efficient and robust learning / Noisy labels
Revisiting Interpolation for Noisy Label Correction
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.
Revisits interpolation-based label correction via gradient-scaling analysis and proposes DULC, which uses normalized JSD between predictions and noisy labels as a sample-wise dynamic, class-uniform interpolation weight, plus augmentation and prediction sharpening.
Problem
Label correction methods for learning with noisy labels suffer from false label correction because their interpolation weights are not sample-wise dynamic or class-uniform, leaving them behind sample-selection SOTA methods.
Approach
Theoretical gradient-scaling analysis shows an ideal interpolation weight should be sample-wise dynamic and class-wise uniform; normalized JSD (per-class min-max plus linear decay) is used as the weight, combined with weak/strong augmentation and a temperature sharpening function, and MixUp plus contrastive loss.
Key results
- On CIFAR-10 with symmetric 90% noise, DULC reaches 93.5% test accuracy versus 90.8% for UNICON.
- On WebVision, DULC achieves 79.9% top-1 and 93.7% top-5, above the prior best (RCAL+ at 79.6% top-1).
- On Clothing1M, DULC achieves 75.09% accuracy, versus 74.98% for UNICON.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DULC instead of UNICON | label correction on CIFAR-10 · CIFAR-10 symmetric 90% noise | test accuracy: 93.5 vs 90.8 | Helped |
| Use DULC instead of UNICON | label correction on CIFAR-100 · CIFAR-100 symmetric 90% noise | test accuracy: 52.8 vs 44.8 | Helped |
| Use DULC instead of UNICON | noisy-label learning on TinyImageNet · TinyImageNet symmetric 20% noise | average test accuracy: 58.5 vs 58.4 | Helped |
| Use DULC instead of UNICON | noisy-label learning on TinyImageNet · TinyImageNet symmetric 50% noise | average test accuracy: 52.1 vs 52.4 | Hurt |
| Add weak/strong augmentation (ablation) | DULC on CIFAR-10 · CIFAR-10 symmetric 90% noise | test accuracy: 82.3±2.64 -> 93.5±0.24 | Helped |
| Add prediction sharpening (ablation) | DULC on CIFAR-10 · CIFAR-10 symmetric 90% noise | test accuracy: 74.7±2.12 -> 93.5±0.24 | Helped |
| Use the dynamic JSD metric for interpolation weights (ablation) | DULC on CIFAR-10 · CIFAR-10 symmetric 90% noise | test accuracy: 60.7±3.69 -> 93.5±0.24 | Helped |
| Add class-wise JSD normalization (ablation) | DULC on CIFAR-10 · CIFAR-10 symmetric 90% noise | test accuracy: 24.6±0.09 -> 93.5±0.24 | Helped |
| Add linear decay to the interpolation weight (ablation) | DULC on CIFAR-10 · CIFAR-10 asymmetric 40% noise | test accuracy: 92.2±0.84 -> 95.2±0.67 | Helped |
| Replace the dynamic sample-wise weight with a fixed ensemble weight (ablation) | DULC on CIFAR-10 · CIFAR-10 symmetric 90% noise | test accuracy: 32.8% degradation | Hurt |
Surprising findings
- Removing the JSD normalization drops CIFAR-10 symmetric-90% accuracy by up to 68.9 points, showing class-wise uniformity matters greatly at high noise.
- On Tiny-ImageNet DULC is slightly below UNICON on average accuracy (52.1 vs 52.4 at 50% noise), which the authors attribute to more classes and fewer samples per class.
Limitations
- Performance on Tiny-ImageNet is slightly below UNICON, attributed to prediction confusion with many classes and few samples per class.
- Interpolation weight for hard classes at high noise is unreliable early in training, requiring a hand-set linear decay schedule with hyperparameters a, b, and gamma.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet WebVision Clothing1M ILSVRC12
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet WebVision ILSVRC12 Clothing1M
Metrics: test accuracy top-1 accuracy top-5 accuracy
Baselines: Standard CE MixUp ELR DivideMix JPL MOIT Sel-CL UNICON MILD OT-Filter HMW K-SPR
Builds on: Bootstrapping M-correction SELC UNICON MixUp
Closest papers by meaning
- Meta Label Correction for Noisy Label Learning (AAAI 2021 · similarity 0.90)
- Learning to Purify Noisy Labels via Meta Soft Label Corrector (AAAI 2021 · similarity 0.88)
- Does label smoothing mitigate label noise? (ICML 2020 · similarity 0.88)
- Robust Classification via Regression for Learning with Noisy Labels (ICLR 2024 · similarity 0.88)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.87)
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 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.