Research map / Architectures and efficiency / Pruning and efficient tuning
Progressive Skeletonization: Trimming more fat from a network at initialization
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 |
|---|---|---|---|
| Progressively prune while recalculating connection sensitivity | SNIP-style pruning · CIFAR-10, ResNet50 and VGG19, 99.9% pruning | test accuracy: nearly 10% -> more than 80% | Helped |
| Average SNIP connection sensitivity over multiple batches | SNIP · CIFAR-10, 99.0% sparsity | accuracy: 10% -> almost 90% | Helped |
| Use one-shot SNIP instead of progressive pruning | CIFAR-10, ResNet50 and VGG19, 99.0% sparsity · CIFAR-10, 99.0% sparsity | accuracy: 10% | Hurt |
| Allow pruned weights to resurrect during iterative pruning (ablation) | Iterative SNIP, forming FORCE · CIFAR-10, ResNet architectures | test accuracy | Helped |
| Use FORCE instead of Iterative SNIP | CIFAR-10, MobileNet-v2 · CIFAR-10, high sparsity | test accuracy | Hurt |
| Increase the number of pruning iterations (ablation) | FORCE and Iterative SNIP · CIFAR-10, 99.5% sparsity | test accuracy: 10% acc; up to 90% for FORCE and 89% for Iter SNIP | Helped |
| Use only a few pruning iterations rather than more iterations (ablation) | FORCE · CIFAR-10, 99.5% sparsity | test accuracy | Mixed |
| Apply GRASP iteratively using its Taylor approximation (ablation) | GRASP · CIFAR-10 pruning experiments | trainability | Hurt |
| Use the paper's gradient approximation for iterative gradient-norm pruning (ablation) | GRASP, forming Iterative GRASP · CIFAR-10, high sparsity | test accuracy | Helped |
| Use FORCE rather than GRASP | ImageNet, ResNet50, 95% sparsity · ImageNet, ResNet50, 95% sparsity | Top-1 accuracy: 46.2 -> 59.0 | Helped |
| Use FORCE rather than GRASP | ImageNet, VGG19, 95% sparsity · ImageNet, VGG19, 95% sparsity | Top-1 accuracy: 67.6 -> 65.8 | Hurt |
| Use FORCE rather than GRASP | ImageNet, ResNet50, 90% sparsity · ImageNet, ResNet50, 90% sparsity | Top-1 accuracy: 65.4 -> 64.9 | Hurt |
Closest papers by meaning
- When To Prune? A Policy Towards Early Structural Pruning (CVPR 2022 · similarity 0.89)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.89)
- Rethinking the Value of Network Pruning (ICLR 2019 · similarity 0.88)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.87)
- SNIP: SINGLE-SHOT NETWORK PRUNING BASED ON CONNECTION SENSITIVITY (ICLR 2019 · similarity 0.87)
- SAFE: Finding Sparse and Flat Minima to Improve Pruning (ICML 2025 · similarity 0.86)
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.