Research map / Learning theory and algorithms / Sample complexity
What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation
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 |
|---|---|---|---|
| remove unique high-influence training examples (ablation) | ResNet50 · CIFAR-100 test set | accuracy: 76.06±0.28% -> 73.52±0.25% | Hurt |
| remove unique high-influence training examples (ablation) | ResNet50 · CIFAR-100 highly-influenced test examples | accuracy: 72.14±1.32% -> 45.38±1.45% | Hurt |
| use a frozen full-data penultimate-layer representation and train only a linear classifier (ablation) | ResNet50 CIFAR-100 influence and memorization estimation · CIFAR-100 | training examples with memorization estimate above 0.25: 18,099 -> 38 | Hurt |
| use a frozen full-data penultimate-layer representation and train only a linear classifier (ablation) | ResNet50 · CIFAR-100 test set (70% training data) | accuracy: 75.8±0.1% vs. 72.3±0.3% | Mixed |
| remove memorized examples rather than an equal-size random subset | ResNet50 · ImageNet test set | marginal utility: ≈ 2.6% -> ≈ 3.4% | Helped |
| train only the last layer over representations from a model trained on the full dataset (ablation) | CIFAR-100 influence estimation · CIFAR-100 | training time: 720 | Helped |
Closest papers by meaning
- How Spurious Features are Memorized: Precise Analysis for Random and NTK Features (ICML 2024 · similarity 0.86)
- Rethinking Benign Overfitting in Two-Layer Neural Networks (ICML 2025 · similarity 0.86)
- The Pitfalls of Memorization: When Memorization Hurts Generalization (ICLR 2025 · similarity 0.85)
- Memorization Through the Lens of Curvature of Loss Function Around Samples (ICML 2024 · similarity 0.84)
- To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets (ICLR 2024 · similarity 0.84)
- Memorization in Self-Supervised Learning Improves Downstream Generalization (ICLR 2024 · similarity 0.84)
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.