Research map / Architectures and efficiency / Pruning and efficient tuning
Get More at Once: Alternating Sparse Training with Gradient Correction
Li Yang, Jian Meng, Jae-sun Seo, Deliang Fan
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 gradient correction (GC) to AST (si=0%) (ablation) | AST · CIFAR-10 wide ResNet-32, 98% sparsity | Top-1 accuracy: 92.44 -> 92.69 | Improved |
| Add gradient correction (GC) to AST (si=90%) (ablation) | AST · CIFAR-10 wide ResNet-32, 90% sparsity | Top-1 accuracy: 92.32 -> 92.90 | Improved |
| Add gradient correction to AST (si=0%) (ablation) | AST · CIFAR-100 wide ResNet-32, 90% sparsity | Top-1 accuracy: 73.12 -> 73.41 | Improved |
| Add gradient correction to AST (si=50%) (ablation) | AST · ImageNet ResNet-50, 80% prune ratio | Top-1 accuracy: 72.6 -> 73.2 | Improved |
| Add gradient correction to AST (si=50%) (ablation) | AST · ImageNet ResNet-50, 90% prune ratio | Top-1 accuracy: 72.3 -> 73.1 | Improved |
| Use non-disjoint (ND) sub-nets instead of completely subset (CS) (ablation) | AST · CIFAR-10 ResNet-32, 70% pruning | Accuracy: 93.69 -> 94.47 | Improved |
| Use non-disjoint instead of completely subset sub-nets (ablation) | AST · CIFAR-10 ResNet-32, 98% pruning | Accuracy: 92.49 -> 92.76 | Improved |
| AST+GC instead of Jointly-Trained S-Net | AST+GC · CIFAR-10 wide ResNet-32, 90% sparsity | Top-1 accuracy: 92.59 -> 93.88 | Improved |
| AST+GC instead of Jointly-Trained S-Net | AST+GC · CIFAR-100 wide ResNet-32, 90% sparsity | Top-1 accuracy: 70.40 -> 73.41 | Improved |
| AST+GC instead of individually trained GraNet (3x cost) | AST+GC · CIFAR-100 wide ResNet-32, 98% sparsity | Top-1 accuracy: 69.89 -> 68.42 | Hurt |
| AST+GC instead of GraNet | AST+GC · CIFAR-10 wide ResNet-32, 90% sparsity | Top-1 accuracy: 94.12 -> 93.88 | Hurt |
| Train more sub-nets (2 to 5) with AST instead of individual training (ablation) | AST · CIFAR-10 ResNet-18, 98% sparsity | Accuracy: 94.50 -> 94.18 | Mixed |
Closest papers by meaning
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.89)
- Dynamic Sparse Training with Structured Sparsity (ICLR 2024 · similarity 0.88)
- Sparse maximal update parameterization: A holistic approach to sparse training dynamics (NeurIPS 2024 · similarity 0.88)
- Navigating Extremes: Dynamic Sparsity in Large Output Spaces (NeurIPS 2024 · similarity 0.88)
- Dynamic Sparse Training of Diagonally Sparse Networks (ICML 2025 · similarity 0.88)
- Sparse Model Soups: A Recipe for Improved Pruning via Model Averaging (ICLR 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.
