Research map / Label-efficient and robust learning / Noisy labels
CrossSplit: Mitigating Label Noise Memorization through Data Splitting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CrossSplit instead of UNICON | noisy-label training · CIFAR-10, symmetric 90% noise | test accuracy (%): 90.8 -> 91.3 | Helped |
| Use CrossSplit instead of UNICON | noisy-label training · CIFAR-100, symmetric 50% noise | test accuracy (%): 77.6 -> 75.7 | Hurt |
| Use CrossSplit instead of UNICON | noisy-label training · Tiny-ImageNet, symmetric noise | average test accuracy, last 10 epochs (%): 58.4 -> 58.8; 52.4 -> 52.0 | Mixed |
| Use CrossSplit instead of UNICON | noisy-label training · mini-WebVision | best test accuracy (%): 77.60 -> 78.48 | Helped |
| Remove data splitting (ablation) | CrossSplit · CIFAR-100, symmetric 90% noise | best test accuracy (%): 52.40 -> 14.19 | Hurt |
| Remove data splitting (ablation) | CrossSplit · CIFAR-100, asymmetric 40% noise | best test accuracy (%): 76.78 -> 72.12 | Hurt |
| Remove class-balancing normalization (ablation) | CrossSplit · CIFAR-100, symmetric 90% noise | best test accuracy (%): 52.40 -> 33.37 | Hurt |
| Remove class-balancing normalization (ablation) | CrossSplit · CIFAR-10, symmetric 50% noise | best test accuracy (%): 96.34 -> 96.73 | Helped |
| Remove cross-split label correction (ablation) | CrossSplit · CIFAR-100, symmetric 90% noise | best test accuracy (%): 52.40 -> 31.77 | Hurt |
| Remove cross-split label correction (ablation) | CrossSplit · CIFAR-100, asymmetric 40% noise | best test accuracy (%): 76.78 -> 69.61 | Hurt |
| Remove contrastive loss (ablation) | CrossSplit · CIFAR-100, symmetric 90% noise | best test accuracy (%): 52.40 -> 31.42 | Hurt |
| Remove contrastive loss (ablation) | CrossSplit · CIFAR-100, symmetric 50% noise | best test accuracy (%): 75.72 -> 75.68 | No clear effect |
Closest papers by meaning
- Mitigating Memorization of Noisy Labels via Regularization between Representations (ICLR 2023 · similarity 0.89)
- Subclass-Dominant Label Noise: A Counterexample for the Success of Early Stopping (NeurIPS 2023 · similarity 0.89)
- Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.87)
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.87)
- IOFM: Using the Interpolation Technique on the Over-Fitted Models to Identify Clean-Annotated Samples (AAAI 2024 · similarity 0.87)
- Combating Semantic Contamination in Learning with Label Noise (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.