Research map / Architectures and efficiency / Pruning and efficient tuning
Sparse Training via Boosting Pruning Plasticity with Neuroregeneration
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 |
|---|---|---|---|
| Add gradient-based zero-cost neuroregeneration after gradual magnitude pruning | GraNet versus GMP · CIFAR-100, VGG-19, 90%, 95%, and 98% sparsity | accuracy | Helped |
| Add sparse-to-sparse gradual pruning with neuroregeneration, starting at s_i=0.5 | VGG-19 versus RigL · CIFAR-10 and CIFAR-100, 98% sparsity | accuracy: +1.40% and 2.22% | Helped |
| Begin sparse-to-sparse training from a denser model and gradually prune it | GraNet versus RigL · ImageNet, ResNet-50, 80% and 90% sparsity | Top-1 accuracy: 0.9% and 1.5% | Helped |
| Regenerate pruned connections using gradients | pruned networks during training · CIFAR-10, ResNet-20; pruning plasticity | pruning plasticity | Helped |
| Prune at a lower learning rate after learning-rate drops | during-training pruning · CIFAR-10, ResNet-20; pruning plasticity | pruning plasticity | Hurt |
| Use an extreme one-shot pruning rate such as 0.98 instead of a mild rate | during-training pruning · CIFAR-10, ResNet-20; pruning plasticity | pruning plasticity | Hurt |
| Set GraNet initial sparsity to 0.5 rather than 0.0 (ablation) | GraNet · ImageNet, ResNet-50, final sparsity 0.9 | Top-1 accuracy: 74.2 -> 74.5 | Helped |
| Increase initial sparsity from 0.5 to 0.8 (ablation) | GraNet · ImageNet, ResNet-50, final sparsity 0.9 | Top-1 accuracy: 74.5 -> 74.1 | Hurt |
| Extend training steps by 2 times instead of using the standard GraNet schedule (ablation) | GraNet (s_i=0) · CIFAR-10 and CIFAR-100, VGG-19 and ResNet-50 | accuracy | Helped |
| Apply Lottery Ticket Hypothesis retraining to GraNet subnetworks (ablation) | GraNet (s_i=0) · CIFAR-10 and CIFAR-100, VGG-19 and ResNet-50 | accuracy | Hurt |
| Apply learning-rate rewinding to GraNet subnetworks (ablation) | GraNet (s_i=0) · CIFAR-10 and CIFAR-100, VGG-19 and ResNet-50 | accuracy | No clear effect |
| Randomly reinitialize weights while retaining the pruned mask (ablation) | GraNet-discovered subnetworks · CIFAR-10, VGG-19 | accuracy | Hurt |
Closest papers by meaning
- Pruning neural networks without any data by iteratively conserving synaptic flow (NeurIPS 2020 · similarity 0.91)
- When To Prune? A Policy Towards Early Structural Pruning (CVPR 2022 · similarity 0.90)
- The Unreasonable Effectiveness of Random Pruning: Return of the Most Naive Baseline for Sparse Training (ICLR 2022 · similarity 0.90)
- Pruning Neural Networks at Initialization: Why Are We Missing the Mark? (ICLR 2021 · similarity 0.90)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.88)
- Improving the Sparse Structure Learning of Spiking Neural Networks from the View of Compression Efficiency (ICLR 2025 · 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.