Research map / Label-efficient and robust learning / Noisy labels
Learning Discriminative Dynamics with Label Corruption for Noisy Label Detection
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 DynaCor, which detects noisy labels by clustering learned representations of training dynamics, with label-corrupted augmented copies used to simulate noisy-label behavior.
Problem
Detecting incorrectly labeled instances in noisy training datasets; prior training-signal heuristics capture incomplete temporal behavior and do not generalize across noise types, and no clean/noisy supervision is available.
Approach
Create a corrupted copy of the dataset with randomly flipped labels to supply proxy noisy-label dynamics, then train a 1D-CNN dynamics encoder with a clustering loss and a cross-dataset alignment loss to form two clusters (clean vs noisy), plus an unsupervised validation metric for early stopping.
Key results
- DynaCor averages 87.7% F1 across nine CIFAR-10/100 settings with CLIP w/ MLP, +3.0% over the best baseline (SimiFeat-R, 84.7%).
- With ResNet34, DynaCor averages 85.2% F1, +2.8% over the best baseline (SimiFeat-V, 81.4%).
- Plugging DynaCor into DivideMix improves its noisy-label detection F1 and classification accuracy on CIFAR-100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DynaCor instead of the compared detection methods | Noisy label detection with CLIP w/ MLP · CIFAR-10 and CIFAR-100, nine noise settings | Average F1 score (%): +3.0% | Helped |
| Use DynaCor instead of the compared detection methods | Noisy label detection with ResNet34 · CIFAR-10 and CIFAR-100, nine noise settings | Average F1 score (%): +2.8% | Helped |
| Add the clustering loss to a randomly initialized dynamics encoder (ablation) | DynaCor · CIFAR-10 Asym. | F1 score (%): 93.8 ± 0.17 -> 93.2 ± 0.11 | Hurt |
| Add the clustering loss to a randomly initialized dynamics encoder (ablation) | DynaCor · CIFAR-10 Inst. | F1 score (%): 91.8 ± 0.39 -> 92.7 ± 0.36 | Helped |
| Add the clustering loss to a randomly initialized dynamics encoder (ablation) | DynaCor · CIFAR-10 Agg. | F1 score (%): 78.8 ± 0.37 -> 76.8 ± 0.83 | Hurt |
| Add the alignment loss to the clustering-loss variant (ablation) | DynaCor · CIFAR-10 Asym. | F1 score (%): 93.2 ± 0.11 -> 94.0 ± 0.15 | Helped |
| Add the alignment loss to the clustering-loss variant (ablation) | DynaCor · CIFAR-10 Inst. | F1 score (%): 92.7 ± 0.36 -> 92.3 ± 0.38 | Hurt |
| Add the alignment loss to the clustering-loss variant (ablation) | DynaCor · CIFAR-10 Agg. | F1 score (%): 76.8 ± 0.83 -> 79.6 ± 0.37 | Helped |
| Use the proposed validation metric instead of stopping at the maximum epoch (ablation) | DynaCor dynamics encoder · CIFAR-10 Inst. | F1 score (%): 86.7 ± 6.75 -> 92.3 ± 0.38 | Helped |
| Use the proposed validation metric instead of DBI (ablation) | DynaCor dynamics encoder · CIFAR-100 Inst. | F1 score (%): 60.0 ± 10.2 -> 81.7 ± 0.21 | Helped |
| Use the proposed validation metric rather than the optimal detection epoch (ablation) | DynaCor dynamics encoder · CIFAR-10 Inst. | F1 score (%): 92.3 ± 0.38; 92.6 ± 0.40 | No clear effect |
Surprising findings
- Higher corruption rate made training dynamics more distinguishable for instance-dependent noise with CLIP w/ MLP, suggesting symmetric corruption can ease detection.
Limitations
- Relies on the assumption that the diagonally dominant condition holds for the noise analysis.
- Detection performance varies with the corruption rate and the choice of backbone (ResNet34 more affected than CLIP).
- Some settings, e.g., CIFAR-10 Agg. and CIFAR-100 Human, remain near or below the best baseline.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N Clothing1M
Benchmarks: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N Clothing1M
Metrics: F1 score classification accuracy
Baselines: Avg.Encoder AUM CL CORES SimiFeat-V SimiFeat-R
Builds on: AUM Dividemix SimiFeat DEC (deep embedded clustering) memorization effect
Closest papers by meaning
- GradTrack: Detecting Noisy Labels via Temporal Trajectories of Class-wise Gradient Misalignment (NeurIPS 2026 · similarity 0.89)
- Learning from Training Dynamics: Identifying Mislabeled Data beyond Manually Designed Features (AAAI 2023 · similarity 0.88)
- Tackling Instance-Dependent Label Noise via a Universal Probabilistic Model (AAAI 2021 · similarity 0.87)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.87)
- Noisy Label Learning with Instance-Dependent Outliers: Identifiability via Crowd Wisdom (NeurIPS 2024 · similarity 0.86)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 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.