Research map / Architectures and efficiency / Pruning and efficient tuning
The Generalization-Stability Tradeoff In Neural Network Pruning
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 |
|---|---|---|---|
| lower pruning stability (more destabilizing pruning rate/target) | magnitude pruning VGG11 · CIFAR-10 | final test accuracy: Pearson r=-.73 | Helped |
| lower pruning stability | magnitude pruning ResNet18 · CIFAR-10 | final test accuracy: Pearson r=-.43 | Helped |
| prune largest-magnitude weights instead of smallest (ablation) | magnitude pruning · CIFAR-10 | test accuracy: 87.72 -> 88.03 | Helped |
| raise iterative pruning rate up to one-shot for Prune_S/Prune_R (ablation) | magnitude pruning VGG11 · CIFAR-10 | test accuracy | Helped |
| raise iterative pruning rate for Prune_L (already unstable) (ablation) | magnitude pruning VGG11 · CIFAR-10 | test accuracy | Hurt |
| raise total pruning percentage 46% to 59% for Prune_L 13% (ablation) | magnitude pruning ResNet18 · CIFAR-10 | test accuracy: 88.03 -> 87.48 | Hurt |
| raise total pruning percentage 46% to 59% for Prune_S 13% (ablation) | magnitude pruning ResNet18 · CIFAR-10 | test accuracy: 87.72 -> 87.95 | Helped |
| raise total pruning percentage 46% to 59% for Prune_S 1% (ablation) | magnitude pruning ResNet18 · CIFAR-10 | test accuracy: 87.65 -> 87.52 | No clear effect |
| temporarily zero filters for 50 batches instead of permanent pruning (ablation) | Prune_L VGG11 · CIFAR-10 | final test accuracy | No clear effect |
| add Gaussian noise to weights for 1105 batches (ablation) | VGG11 · CIFAR-10 | test accuracy | Helped |
| pruning 10% of each layer vs 6% under traditional training (ablation) | ResNet18 · CIFAR-10 | test accuracy: 94.25 -> 94.41 | Mixed |
| lower pruning stability increases flatness | magnitude pruning VGG11 · CIFAR-10 | flatness measures | Helped |
Closest papers by meaning
- Pruning Neural Networks at Initialization: Why Are We Missing the Mark? (ICLR 2021 · similarity 0.91)
- Pruning’s Effect on Generalization Through the Lens of Training and Regularization (NeurIPS 2022 · similarity 0.91)
- A Three-regime Model of Network Pruning (ICML 2023 · similarity 0.90)
- Simplicity Bias and Optimization Threshold in Two-Layer ReLU Networks (ICML 2025 · similarity 0.88)
- GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning (NeurIPS 2024 · similarity 0.88)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.88)
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.