Research map / Architectures and efficiency / Pruning and efficient tuning
Pruning via Sparsity-indexed ODE: a Continuous Sparsity Viewpoint
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 |
|---|---|---|---|
| Replace magnitude/SNIP/etc. one-shot pruning with one-shot PSO (SpODE path) | PSO one-shot unstructured pruning · CIFAR-100, ResNet-20/VGG16-bn/WRN-20 | top-1 accuracy | Mixed |
| Iterative pruning scheme added to PSO | PSO · CIFAR-100, WRN-20 98% sparsity | top-1 accuracy: 59.92 to 63.17 | Helped |
| Increase SpODE batch size from 64 to 128 (with longer finetuning) (ablation) | PSO on ImageNet ResNet-50 · ImageNet ResNet-50 90% sparsity | top-1 accuracy: 73.19 to 74.20 | Helped |
| Use exponential Δt schedule instead of linear or inverse-exponential (ablation) | PSO SpODE discretization · CIFAR-100, ResNet-20 | top-1 accuracy | Helped |
| Replace explicit SpODE update with regularized (penalty) counterpart (ablation) | PSO · CIFAR-100, VGG16-bn | top-1 accuracy | Hurt |
| Reduce SpODE discretization steps to 20 (ablation) | PSO · CIFAR-100, VGG16-bn 95% sparsity | top-1 accuracy: about 2.5% drop with 20 steps | No clear effect |
| Use Gaussian/quantile/one-hot empirical polarizers with ε<0.3, α<0.1 (ablation) | PSO polarizer · CIFAR-100, ResNet-20 95% sparsity | top-1 accuracy | No clear effect |
| One-shot PSO vs magnitude pruning on ImageNet ResNet-50 | PSO one-shot unstructured pruning · ImageNet ResNet-50 90% and 95% sparsity | top-1 accuracy: +0.12 at 90%, +0.06 at 95% | Helped |
| Iterative PSO vs iterative baselines on ImageNet ResNet-50 at 90% | PSO iterative pruning · ImageNet ResNet-50 90% sparsity | top-1 accuracy: 75.10 vs best baseline WoodFisher 75.21 is higher; PSO beats STR, Powerprop, ProbMask | Helped |
Closest papers by meaning
- Probabilistic Neural Pruning via Sparsity Evolutionary Fokker-Planck-Kolmogorov Equation (ICLR 2025 · similarity 0.93)
- The Unreasonable Effectiveness of Random Pruning: Return of the Most Naive Baseline for Sparse Training (ICLR 2022 · similarity 0.91)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.91)
- SAFE: Finding Sparse and Flat Minima to Improve Pruning (ICML 2025 · similarity 0.90)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.89)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · 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.