Research map / Architectures and efficiency / Pruning and efficient tuning
Sparse Double Descent: Where Network Pruning Aggravates Overfitting
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 |
|---|---|---|---|
| Increasing magnitude-based pruning sparsity (moderate sparsity, Critical Phase) under label noise | LeNet-300-100 on MNIST, ResNet-18 on CIFAR-10/100 with symmetric label noise · MNIST / CIFAR-10 / CIFAR-100 with 20-40% label noise | test accuracy | Hurt |
| Increasing sparsity to high levels (Sweet Phase) under label noise | Pruned ResNet-18 / LeNet-300-100 with label noise · CIFAR-100 eps=40%, MNIST eps=20% | test accuracy | Helped |
| Pruning to extreme sparsity (Collapsed Phase) | Pruned networks under label noise · MNIST, CIFAR-10, CIFAR-100 | train and test accuracy | Hurt |
| Random pruning instead of magnitude-based pruning | ResNet-18 on CIFAR-100 with 40% label noise · CIFAR-100 eps=40% | test accuracy | No clear effect |
| Retraining from random reinitialization (scratch/reinit) instead of original initialization (lottery ticket rewinding) for sparse models (ablation) | ResNet-18 on CIFAR-10 and CIFAR-100 with label noise · CIFAR-10 eps=20% | test accuracy | Mixed |
| Removing label noise (zero noise) from training data (ablation) | ResNet-18 on CIFAR-10/100 · CIFAR-10 / CIFAR-100 no label noise | test accuracy | Hurt |
| Re-dense training (recover pruned weights and retrain) after sparse training at high sparsity | Sparse solutions of ResNet-18 on CIFAR-100 eps=40% · CIFAR-100 eps=40% | test accuracy | Hurt |
| Learning distance (l2 from initialization) as explanatory measure versus minima flatness for sparse models | Sparse models across sparsities · LeNet-300-100 MNIST eps=20%; ResNet-18 CIFAR | correlation with test accuracy curve | Helped |
| Using best test accuracy across epochs instead of last-epoch test accuracy | Sparse models across sparsities · CIFAR-100 eps=40% | best test accuracy | No clear effect |
Closest papers by meaning
- DSD²: Can We Dodge Sparse Double Descent and Compress the Neural Network Worry-Free? (AAAI 2024 · similarity 0.91)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.88)
- Unraveling the Enigma of Double Descent: An In-depth Analysis through the Lens of Learned Feature Space (ICLR 2024 · similarity 0.88)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.87)
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization (ICML 2024 · similarity 0.87)
- Asymmetric Scaling Laws from Sparse Features (NeurIPS 2026 · 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.