Research map / Architectures and efficiency / Pruning and efficient tuning
Pruning Neural Networks at Initialization: Why Are We Missing the Mark?
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 SNIP, GraSP, SynFlow, or magnitude pruning at initialization instead of random pruning | Neural network pruning at initialization · Image classification at matching sparsities | accuracy | Helped |
| Prune at initialization rather than after training with magnitude pruning | SNIP, GraSP, SynFlow, and magnitude pruning · Image classification at matching sparsities | accuracy | Hurt |
| Replace each method's specific within-layer mask positions with a random shuffle, preserving layerwise sparsity (ablation) | SNIP, GraSP, SynFlow, and magnitude pruning at initialization · Image classification | accuracy | No clear effect |
| Randomly reinitialize the unpruned weights from the same initialization distribution (ablation) | SNIP, GraSP, SynFlow, and magnitude pruning at initialization · Image classification | accuracy | No clear effect |
| Shuffle SynFlow's layerwise pruning mask at extreme sparsities (ablation) | SynFlow · ResNet-20 and ResNet-18 | accuracy | Helped |
| Prune the weights SNIP, magnitude, or SynFlow scores as most important, retaining the least important (ablation) | SNIP, magnitude pruning, and SynFlow · Image classification at initialization | accuracy | Hurt |
| Invert GraSP's score selection by pruning the weights scored as most important (ablation) | GraSP · Image classification at initialization | accuracy | No clear effect |
| Prune using SNIP, SynFlow, or magnitude after training has begun rather than at initialization | SNIP, SynFlow, and magnitude pruning · ResNet-20, VGG-16, and ResNet-18 | accuracy | Helped |
| Use GraSP scores with the opposite selection direction, pruning the lowest-scored weights (ablation) | GraSP · Image classification | accuracy | Helped |
Closest papers by meaning
- Robust Pruning at Initialization (ICLR 2021 · similarity 0.92)
- “Learning-Compression” Algorithms for Neural Net Pruning (CVPR 2018 · similarity 0.92)
- A Signal Propagation Perspective for Pruning Neural Networks at Initialization (ICLR 2020 · similarity 0.92)
- Pruning neural networks without any data by iteratively conserving synaptic flow (NeurIPS 2020 · similarity 0.92)
- Pruning from Scratch (AAAI 2020 · similarity 0.91)
- The Generalization-Stability Tradeoff In Neural Network Pruning (NeurIPS 2020 · similarity 0.91)
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.