Research map / Architectures and efficiency / Pruning and efficient tuning
Dynamic Sparsity Is Channel-Level Sparsity Learner
Lu Yin, Gen Li, Meng Fang, Li Shen, Tianjin Huang, Zhangyang Wang, Vlado Menkovski, Xiaolong Ma, Mykola Pechenizkiy, Shiwei Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add global parameter exploration to RigL (Chase Sc=0) (ablation) | RigL · ImageNet, ResNet-50, 80% sparsity | top-1 accuracy: 75.1 -> 75.87 | Improved |
| add soft memory bound (ablation) | Chase/RigL · CIFAR-100, ResNet-50, 98% sparsity | accuracy: +1.26% | Improved |
| add global parameter exploration on top of soft memory bound (ablation) | Chase · CIFAR-10/100, ResNet-50 and VGG-19 | accuracy | Improved |
| add gradual amenable channel pruning (50% channels) to RigL (ablation) | RigL · CIFAR-100, ResNet-50, 98% sparsity | accuracy: 74.84 -> 74.92 | Improved |
| add gradual amenable channel pruning (50% channels) to SET (ablation) | SET · CIFAR-100, ResNet-50, 98% sparsity | accuracy: 73.21 -> 73.86 | Improved |
| add gradual amenable channel pruning to RigL (ablation) | RigL · CIFAR-100, ResNet-50, 90% sparsity | accuracy: 77.18 -> 76.86 | Hurt |
| use UMM instead of random pruning for one-shot channel pruning | RigL · CIFAR-100, ResNet-50, 30% channel sparsity | accuracy: 3.79 -> 71.77 | Improved |
| prune with reversed UMM metric instead of UMM | RigL · CIFAR-100, ResNet-50, 10% channel sparsity | accuracy: 76.88 -> 3.18 | Hurt |
| add Chase to SNIP (static sparse training) | SNIP · CIFAR-100, ResNet-50, 90% sparsity | accuracy: 73.14 -> 73.44 | Improved |
| add Chase to GraSP (static sparse training) | GraSP · CIFAR-100, ResNet-50, 98% sparsity | accuracy: 62.12 -> 75.74 | Improved |
| increase channel sparsity from Sc=0 to Sc=0.4 (ablation) | Chase · ImageNet, ResNet-50, 80% sparsity | accuracy: 75.87 -> 75.27 | Hurt |
| also prune skip-connection layers (ablation) | Chase · ImageNet, ResNet-50, 90% sparsity, Sc=0.5 | accuracy / throughput: 73.39 -> 72.71; 78.19 -> 99.97 | Mixed |
Closest papers by meaning
- Navigating Extremes: Dynamic Sparsity in Large Output Spaces (NeurIPS 2024 · similarity 0.92)
- Dynamic Sparse Training with Structured Sparsity (ICLR 2024 · similarity 0.91)
- Dynamic Sparse Training of Diagonally Sparse Networks (ICML 2025 · similarity 0.91)
- Advancing Dynamic Sparse Training by Exploring Optimization Opportunities (ICML 2024 · similarity 0.89)
- Sparse Cocktail: Every Sparse Pattern Every Sparse Ratio All At Once (ICML 2024 · similarity 0.89)
- Sparse maximal update parameterization: A holistic approach to sparse training dynamics (NeurIPS 2024 · similarity 0.88)
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.
