Research map / Architectures and efficiency / Pruning and efficient tuning
Pruning neural networks without any data by iteratively conserving synaptic flow
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 |
|---|---|---|---|
| Use data-agnostic iterative SynFlow pruning instead of single-shot pruning | Pruning at initialization · VGG and ResNet; CIFAR-10, CIFAR-100, and Tiny ImageNet; high-compression regimes | Top-1 test accuracy | Helped |
| Use data-agnostic iterative SynFlow pruning instead of SNIP or GraSP | Pruning at initialization · VGG and ResNet; CIFAR-10, CIFAR-100, and Tiny ImageNet; low-compression regimes | Top-1 test accuracy | Mixed |
| Iteratively re-evaluate SNIP scores (ablation) | SNIP · VGG-16 on CIFAR-100 | Top-1 test accuracy | Helped |
| Continue iterative SNIP pruning at high compression (ablation) | SNIP · VGG-16 on CIFAR-100 | Top-1 test accuracy | Hurt |
| Invert SynFlow scores so the most important connections are pruned first (ablation) | SynFlow · VGG-16 on CIFAR-100 | Top-1 test accuracy | Hurt |
| Invert SNIP scores (ablation) | SNIP · VGG-16 on CIFAR-100 | Top-1 test accuracy | Hurt |
| Invert GraSP scores (ablation) | GraSP · VGG-16 on CIFAR-100 | Top-1 test accuracy | No clear effect |
| Reinitialize weights after SynFlow pruning (ablation) | SynFlow · Convolutional models; low-compression regimes | Top-1 test accuracy | No clear effect |
| Shuffle the SynFlow mask within each layer (ablation) | SynFlow · Convolutional models; low-compression regimes | Top-1 test accuracy | No clear effect |
| Use a linear instead of exponential compression schedule (ablation) | SynFlow · VGG-19 on ImageNet | Maximum prune-size-to-cut-size ratio | Hurt |
Closest papers by meaning
- Pruning Neural Networks at Initialization: Why Are We Missing the Mark? (ICLR 2021 · similarity 0.92)
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization (ICML 2024 · similarity 0.91)
- Sparse Training via Boosting Pruning Plasticity with Neuroregeneration (NeurIPS 2021 · similarity 0.91)
- Pruning via Sparsity-indexed ODE: a Continuous Sparsity Viewpoint (ICML 2023 · similarity 0.91)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.90)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · 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.