Research map / Architectures and efficiency / Pruning and efficient tuning
Rethinking the Value of 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 |
|---|---|---|---|
| Train predefined structured-pruned architectures from random initialization rather than fine-tuning inherited weights | L1-norm filter pruning · CIFAR-10, VGG-16-A | accuracy: 93.41 ($\pm$0.12) -> 93.78 ($\pm$0.15) | Helped |
| Train predefined structured-pruned architectures from random initialization with matched computation budget | L1-norm filter pruning · ImageNet, ResNet-34-A | accuracy: 72.56 -> 73.03 | Helped |
| Train automatically discovered structured-pruned architectures from random initialization with matched computation budget | Network Slimming · CIFAR-10, VGG-19 | accuracy: 93.60 ($\pm$0.16) -> 93.81 ($\pm$0.14) | Helped |
| Train a pruned structure from scratch using matched computation budget | Sparse Structure Selection · ImageNet, ResNet-41 | accuracy: 75.44 -> 76.17 | Helped |
| Train the unstructured-pruned network from scratch on CIFAR-10 | Magnitude-based pruning · CIFAR-10 | accuracy: 93.51 ($\pm$0.05) -> 93.71 ($\pm$0.09); 95.06 ($\pm$0.05) -> 94.84 ($\pm$0.07) | Mixed |
| Train the unstructured-pruned network from scratch on ImageNet | Magnitude-based pruning · ImageNet | accuracy: 73.68 -> 73.42 | Hurt |
| Use matched computation budget instead of the standard epoch budget for scratch training (ablation) | Pruned models · CIFAR-10, VGG-16-A | accuracy: 93.62 ($\pm$0.11) -> 93.78 ($\pm$0.15) | Helped |
| Train the pruned model with original winning-ticket initialization instead of random initialization (ablation) | Structured L1-norm filter pruning · CIFAR-10, ResNet-56-A, learning rate 0.1 | accuracy: 92.75 ($\pm$0.26) -> 92.72 ($\pm$0.10) | No clear effect |
| Train the pruned model with original winning-ticket initialization instead of random initialization (ablation) | Structured L1-norm filter pruning · CIFAR-10, ResNet-56-A, learning rate 0.01 | accuracy: 89.87 ($\pm$0.25) -> 90.00 ($\pm$0.15) | No clear effect |
| Use winning-ticket initialization rather than random initialization for unstructured pruning (ablation) | One-shot magnitude pruning · CIFAR-10, learning rate 0.01 | accuracy: 92.52 ($\pm$0.15) -> 92.78 ($\pm$0.11) | Helped |
| Use winning-ticket initialization rather than random initialization for unstructured pruning (ablation) | One-shot magnitude pruning · CIFAR-10, learning rate 0.1 | accuracy: 93.79 ($\pm$0.11) -> 93.66 ($\pm$0.20) | No clear effect |
| Rescale random initialization standard deviation according to remaining nonzero weights (ablation) | Unstructured magnitude-pruned models · CIFAR and ImageNet | accuracy | No clear effect |
Closest papers by meaning
- Pruning from Scratch (AAAI 2020 · similarity 0.94)
- What Makes a Good Prune? Maximal Unstructured Pruning for Maximal Cosine Similarity (ICLR 2024 · similarity 0.92)
- When To Prune? A Policy Towards Early Structural Pruning (CVPR 2022 · similarity 0.92)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.91)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.90)
- GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning (NeurIPS 2024 · similarity 0.90)
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.