Research map / Label-efficient and robust learning / Noisy labels
Leveraging Unlabeled Data to Track Memorization
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 susceptibility to filter highly memorizing models before ranking by training accuracy | Model selection on CIFAR-10 with 50% label noise · CIFAR-10 50% label noise | correlation between training and test accuracy: \rho=0.608 -> \rho=0.951 | Helped |
| Select using both high training accuracy and low susceptibility rather than high training accuracy alone (ablation) | Model selection on CIFAR-10 with 50% label noise · CIFAR-10 50% label noise | test accuracy: 70.59\% -> 76\% | Helped |
| Use susceptibility with training accuracy rather than oracle clean/noisy labels | Model selection on CIFAR-10 with 50% label noise · CIFAR-10 50% label noise | test accuracy: 77.93_{\pm 4.68}\% vs 76.15_{\pm 6.32}\% | No clear effect |
| Estimate model quality with susceptibility instead of noisy validation accuracy on a small held-out set | Model selection on CIFAR-10 with 50% label noise · CIFAR-10 50% label noise, set size 10 | Kendall τ correlation with test accuracy: 0.244 -> \mathbf{0.792} | Helped |
| Use susceptibility rather than noisy validation accuracy on 128 held-out examples | Model selection on CIFAR-10 with 50% label noise · CIFAR-10 50% label noise, set size 128 | Kendall τ correlation with test accuracy: 0.458 -> \mathbf{0.890} | Helped |
| Select with high training accuracy as well as low susceptibility, rather than low susceptibility alone (ablation) | Noisy-label model selection · Clothing-1M | test accuracy: 57.893\% -> 61.799\% | Helped |
| Select with high training accuracy as well as low susceptibility, rather than low susceptibility alone (ablation) | Noisy-label model selection · Animal-10N | test accuracy: 92.508\% -> 96.371\% | Helped |
| Select with high training accuracy as well as low susceptibility, rather than low susceptibility alone (ablation) | Noisy-label model selection · CIFAR-10N | test accuracy: 85.44\% -> 87.2\% | Helped |
| Track memorization with susceptibility ζ computed on randomly labeled examples | Neural networks trained on noisy CIFAR-10 · CIFAR-10 50% label noise | Pearson correlation with Train ACC Noisy: \rho=0.884 | Helped |
| Use overall training-accuracy increase as a signal of noisy-label memorization (ablation) | Neural networks trained on noisy CIFAR-10 · CIFAR-10 50% label noise | correlation with accuracy on noisy training subset: \rho=0.636 | Hurt |
| Use the training accuracy drop after injecting random labels to track memorization (ablation) | Neural networks trained on noisy CIFAR-10 · CIFAR-10 50% label noise | correlation with accuracy on noisy training subset | Hurt |
| Use the GMM mean-difference early-stopping metric to compare memorization across settings (ablation) | Neural networks trained on noisy CIFAR-10 · CIFAR-10 50% label noise | correlation with accuracy on noisy training subset | Hurt |
Closest papers by meaning
- Unveiling Memorization–Generalization Coexistence: A Case Study on Arithmetic Tasks with Label Noise (NeurIPS 2026 · similarity 0.88)
- Towards Memorization Estimation: Fast, Formal and Free (ICML 2025 · similarity 0.88)
- Memorization Through the Lens of Curvature of Loss Function Around Samples (ICML 2024 · similarity 0.87)
- To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets (ICLR 2024 · similarity 0.86)
- The Cost of Robustness: Tighter Bounds on Parameter Complexity for Robust Memorization in ReLU Nets (NeurIPS 2025 · similarity 0.86)
- Learning the Signature of Memorization in Autoregressive Language Models (NeurIPS 2026 · 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.