Research map / Architectures and efficiency / Pruning and efficient tuning
Coarsening the Granularity: Towards Structurally Sparse Lottery Tickets
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 |
|---|---|---|---|
| Post-process IMP masks by regrouping weights into flexible dense blocks | IMP subnetworks · CIFAR and Tiny-ImageNet classification | GPU running time savings: up to 64.93% | Helped |
| Use channel-wise refilling rather than unstructured IMP alone | IMP subnetworks · CIFAR and Tiny-ImageNet classification | GPU running time savings: 57.53%~61.79% | Helped |
| Use IMP-Regroup instead of IMP-Refill(+) at high sparsity | Structural tickets · CIFAR and Tiny-ImageNet | accuracy at high sparsity | Helped |
| Add extra channels through refilling+ (ablation) | IMP-Refill · Tiny-ImageNet | accuracy-efficiency trade-off | Hurt |
| Use channel-weight L1 norm to rank channels for refilling (ablation) | IMP-Refill · ResNet-18, CIFAR-10 | testing accuracy | Helped |
| Use 50% rewound weights for retraining (ablation) | Structural tickets · CIFAR-10, ResNet-18 | accuracy | Helped |
| Use rewound weights rather than random re-initialization (ablation) | IMP-Refill and IMP-Regroup · CIFAR-10, ResNet-18 | testing accuracy: 94.45% vs 94.20%; 94.75% vs 93.42% | Helped |
| Integrate regrouping into every IMP pruning round (ablation) | Group-aware IMP · Structural ticket experiments | performance | Hurt |
| Use IMP or OMP rather than other unstructured-mask sources before regrouping (ablation) | Regrouping · Structural ticket experiments | train-from-scratch accuracy | Helped |
| Replace rewound weights with random initialization (ablation) | Structural tickets · Sparsity levels >30% | accuracy | Hurt |
| Use channel-wise refilling on ImageNet | ResNet-50 · ImageNet | structural winning-ticket performance | Hurt |
| Use finer-grained group-wise structures rather than channel-wise structures | Structural tickets · ImageNet, ResNet-50 | GPU running time savings: 53.75% / 60.23% | Helped |
Closest papers by meaning
- The Elastic Lottery Ticket Hypothesis (NeurIPS 2021 · similarity 0.94)
- Find A Winning Sign: Sign Is All We Need to Win the Lottery (ICLR 2025 · similarity 0.92)
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (ICLR 2019 · similarity 0.92)
- Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win (ICML 2025 · similarity 0.89)
- On the Sparsity of the Strong Lottery Ticket Hypothesis (NeurIPS 2024 · similarity 0.89)
- Fast Track to Winning Tickets: Repowering One-Shot Pruning for Graph Neural Networks (AAAI 2025 · 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.