Research map / Architectures and efficiency / Pruning and efficient tuning
Exposing and Exploiting Fine-Grained Block Structures for Fast and Accurate Sparse Training
Peng Jiang, Lihan Hu, Shihui Song
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 |
|---|---|---|---|
| remove model adaptation (static shuffled block, SSB-16) at 0.5 sparsity (ablation) | DSB-16 · ImageNet ResNet50 | top-1 accuracy: 75.83 -> 76.33 | Improved |
| remove model adaptation (SSB-16) at 0.5 sparsity (ablation) | DSB-16 · ImageNet ResNet18 | top-1 accuracy: 68.52 -> 69.84 | Improved |
| remove model adaptation (SSB-16) at 0.75 sparsity (ablation) | DSB-16 · ImageNet ResNet50 | top-1 accuracy: 72.76 -> 74.04 | Improved |
| remove model adaptation (SSB-16) at 0.75 sparsity (ablation) | DSB-16 · ImageNet ResNet18 | top-1 accuracy: 64.49 -> 65.72 | Improved |
| DSB-16 at 0.5 sparsity versus dense training | DSB-16 · ImageNet ResNet50 | top-1 accuracy: 76.13 -> 76.33 | Improved |
| DSB-16 at 0.75 sparsity versus SSLT at 0.7 sparsity | DSB-16 · ImageNet ResNet50 | top-1 accuracy: 71.5 -> 74.04 | Improved |
| DSB-16 at 0.5 sparsity versus SSLT at 0.51 sparsity | DSB-16 · ImageNet ResNet50 | top-1 accuracy: 75.65 -> 76.33 | Improved |
| use block structure (DSB) instead of unstructured RigL (block size 1) | DSB · CIFAR10 | test accuracy | Hurt |
| remove shuffling (nonshuffled block sparsity, DNSB) (ablation) | DSB · CIFAR10/100 | test accuracy | Improved |
| remove model adaptation (SSB) on CIFAR10/100 (ablation) | DSB · CIFAR10/100 | test accuracy | Improved |
| block size 4 or 8 or 16 at sparsity up to 0.8 (ablation) | DSB · CIFAR10 | test accuracy | No clear effect |
| increase block size to 32 (ablation) | DSB · CIFAR10 | test accuracy | Hurt |
Closest papers by meaning
- Dynamic Sparse Training with Structured Sparsity (ICLR 2024 · similarity 0.88)
- SequentialAttention++ for Block Sparsification: Differentiable Pruning Meets Combinatorial Optimization (NeurIPS 2024 · similarity 0.87)
- Sparse Cocktail: Every Sparse Pattern Every Sparse Ratio All At Once (ICML 2024 · similarity 0.87)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.87)
- Sparse maximal update parameterization: A holistic approach to sparse training dynamics (NeurIPS 2024 · similarity 0.87)
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.87)
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.
