Research map / Architectures and efficiency / Pruning and efficient tuning
Neural Pruning via Growing Regularization
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 growing L2 regularization before removing the selected filters instead of removing them one-shot | ResNet56 · CIFAR10, pruning ratio 90% | accuracy: 87.34±0.21 -> 89.49±0.23 | Helped |
| Use growing L2 regularization before removing the selected filters instead of removing them one-shot | VGG19 · CIFAR100, pruning ratio 90% | accuracy: 51.36±0.11 -> 57.09±0.03 | Helped |
| Use growing regularization with the same randomly selected filters instead of one-shot removal | ResNet56 · CIFAR10, pruning ratio 90%, random filters | accuracy: 86.85±0.43 -> 89.08±0.23 | Helped |
| Use growing regularization with the same randomly selected filters instead of one-shot removal | VGG19 · CIFAR100, pruning ratio 70%, random filters | accuracy: 65.04±0.29 -> 66.79±0.21 | Helped |
| Use growing-regularization pruning schedule instead of one-shot OBD pruning for the same selected weights | ResNet56 · CIFAR10, 6.87× speedup | accuracy: 86.90 (0.03) -> 88.69 (0.06) | Helped |
| Use GReg-2 rather than GReg-1 | CIFAR pruning · ResNet56 + CIFAR10, 2.55× speedup | pruned accuracy: 93.18 -> 93.36 | Helped |
| Use GReg-2 rather than GReg-1 | CIFAR pruning · VGG19 + CIFAR100, 8.84× speedup | pruned accuracy: 67.55 -> 67.75 | Helped |
| Use GReg-2 rather than GReg-1 | ImageNet pruning · ResNet50, 2.31× speedup | pruned top-1 accuracy: 75.16 -> 75.36 | Helped |
| Set GReg-1 update interval Ku to 1 instead of the default 10 (ablation) | GReg-1 · ResNet56 + CIFAR10, pruning ratio 90% | accuracy: 89.49±0.23 -> 89.40±0.04 | No clear effect |
| Set GReg-1 update interval Ku to 15 instead of the default 10 (ablation) | GReg-1 · ResNet56 + CIFAR10, pruning ratio 90% | accuracy: 89.49±0.23 -> 89.69±0.05 | No clear effect |
| Set GReg-1 update interval Ku to 20 instead of the default 10 (ablation) | GReg-1 · VGG19 + CIFAR100, pruning ratio 70% | accuracy: 67.35±0.15 -> 66.93±0.22 | No clear effect |
Closest papers by meaning
- Structured Pruning of Neural Networks With Budget-Aware Regularization (CVPR 2019 · similarity 0.91)
- Computation and Memory-Efficient Model Compression with Gradient Reweighting (NeurIPS 2025 · similarity 0.91)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.90)
- SequentialAttention++ for Block Sparsification: Differentiable Pruning Meets Combinatorial Optimization (NeurIPS 2024 · similarity 0.90)
- Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision (ICML 2025 · similarity 0.90)
- Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty (ICML 2025 · similarity 0.89)
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.