Research map / Label-efficient and robust learning / Noisy labels
IOFM: Using the Interpolation Technique on the Over-Fitted Models to Identify Clean-Annotated Samples
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 IOFM, a clean-sample identification score that measures how similar a sample's feature-space neighbors are along interpolation paths in input space using an over-fitted (MixUp-trained) DNN, and combines it with DivideMix for noisy-label learning.
Problem
Memorization-effect-based methods for noisy-label learning fail under imbalanced class distributions and heavy label contamination, and they require careful selection of the training epoch.
Approach
Over-fitted DNNs interpolate smoothly between clean same-label samples but not between a noisy sample and its feature-space neighbors; the area under the predicted probability along interpolation paths from a sample to its feature-space nearest same-label neighbors serves as a clean score, averaged over epochs and trained with MixUp.
Key results
- On imbalanced CIFAR100 (step, symmetric 0.3), IOFM reaches AUC 0.973 vs 0.966 for AUM and 0.958 for Loss-based baseline in best-epoch terms, with smaller gap between best and last epoch.
- On heavily noisy CIFAR10 with 90% symmetric noise, IOFM+DivideMix reaches 81.20% test accuracy vs 71.34% for DivideMix (re-implemented) and 78.7% for ELR+ (cited).
- On CIFAR10-N Worst noise, IOFM+DivideMix achieves 93.04% vs SOP+ 93.24%, a near tie.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ensemble IOFM scores across epochs instead of a single final-epoch IOFM score (ablation) | IOFM · Imbalanced CIFAR100 and heavily noisy CIFAR10/CIFAR100 | clean/noisy classification AUC | Helped |
| Use MixUp rather than standard cross-entropy training (ablation) | IOFM · CIFAR10 noisy-label detection | AUC | Helped |
| Increase the number of feature-space neighbors L (ablation) | IOFM · Noisy-label detection | clean/noisy classification performance | Helped |
| Increase interpolation trapezoids beyond H = 3 (ablation) | IOFM · Noisy-label detection | IOFM score approximation | No clear effect |
| Restrict neighbor search to 100 randomly sampled same-label candidates (ablation) | IOFM · Imbalanced CIFAR10 | clean/noisy classification performance | No clear effect |
| Replace DivideMix per-sample losses with IOFM scores | DivideMix · CIFAR10, imbalanced Step, symmetric noise 0.3 | test accuracy (%): 85.76 -> 88.19 | Helped |
| Replace DivideMix per-sample losses with IOFM scores | DivideMix · CIFAR10, imbalanced Step, symmetric noise 0.5 | test accuracy (%): 85.16 -> 88.07 | Helped |
| Replace DivideMix per-sample losses with IOFM scores | DivideMix · CIFAR10, imbalanced Step, asymmetric noise 0.2 | test accuracy (%): 87.15 -> 87.35 | Helped |
| Replace DivideMix per-sample losses with IOFM scores | DivideMix · CIFAR10, imbalanced Step, asymmetric noise 0.4 | test accuracy (%): 78.35 -> 78.73 | Helped |
| Replace DivideMix per-sample losses with IOFM scores | DivideMix · CIFAR10, symmetric noise 0.8 | test accuracy (%): 92.90 -> 93.48 | Helped |
| Replace DivideMix per-sample losses with IOFM scores | DivideMix · CIFAR10, symmetric noise 0.9 | test accuracy (%): 71.34 -> 81.20 | Helped |
| Replace DivideMix per-sample losses with IOFM scores | DivideMix · CIFAR10, asymmetric noise 0.4 | test accuracy (%): 93.36 -> 93.41 | Helped |
Surprising findings
- The memorization effect fails in imbalanced settings: during early training the model memorizes majority-class samples before clean minority samples, so small-loss selection misidentifies minority samples as noisy.
Limitations
- Gains on CIFAR10-N Worst are not best (SOP+ slightly higher), and some baselines are cited from papers rather than re-run.
- Theoretical result (Theorem 1) guarantees existence of a suitable DNN but not that gradient training finds it; relies on regularity assumptions such as separable smooth decision boundary and majority-clean data.
Experimental setup
Datasets: CIFAR10 CIFAR100 Mini-ImageNet Clothing1M CIFAR10-N CIFAR100-N two-moon
Benchmarks: CIFAR10 CIFAR100 CIFAR10-N CIFAR100-N Mini-ImageNet Clothing1M
Metrics: AUC (clean/noisy classification) test accuracy
Baselines: Loss (small-loss) Ens-Loss AUM Margin DivideMix ELR+ CORES SOP+ ROBOT NLS
Builds on: MixUp DivideMix AUM memorization effect
Closest papers by meaning
- CrossSplit: Mitigating Label Noise Memorization through Data Splitting (ICML 2023 · similarity 0.87)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.87)
- Learning from Training Dynamics: Identifying Mislabeled Data beyond Manually Designed Features (AAAI 2023 · similarity 0.87)
- Mitigating Memorization of Noisy Labels via Regularization between Representations (ICLR 2023 · similarity 0.87)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · similarity 0.86)
- Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · 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.