Research map / Architectures and efficiency / Pruning and efficient tuning
Variance-Based Pruning for Accelerating and Compressing Trained Networks
Uranik Berisha, Jens Mehnert, Alexandru Paul Condurache
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 mean-shift compensation to variance-based pruning (ablation) | VBP · ImageNet DeiT-B 50% | accuracy retained before fine-tuning: 26.04 -> 66.40 (retention); 80.62 -> 80.99 (final) | Improved |
| use variance-based selection instead of random-like selection (keep mean-shift) (ablation) | VBP · ImageNet DeiT-B 50% | top-1 accuracy: 80.23 -> 80.99 (final); 55.19 -> 66.40 (retention) | Improved |
| measure variance after the activation instead of before (ablation) | VBP · ImageNet DeiT-B 50% | top-1 accuracy: 77.92 -> 80.99 (final); 0.43 -> 66.40 (retention) | Improved |
| use variance criterion instead of SNIP | structured one-shot pruning · ImageNet DeiT-B 50% | top-1 accuracy: 80.40 -> 80.99 (final); 53.24 -> 66.40 (retention) | Improved |
| use variance criterion instead of Magnitude | structured one-shot pruning · ImageNet DeiT-B 50% | top-1 accuracy: 78.88 -> 80.99 (final); 0.37 -> 66.40 (retention) | Improved |
| use variance criterion instead of SNIP | structured one-shot pruning · ImageNet DeiT-T 50% | top-1 accuracy: 69.10 -> 70.61 (final); 24.89 -> 39.58 (retention) | Improved |
| apply VBP on top of ToMe with smaller token reduction instead of larger ToMe reduction | ToMe on DeiT-B · ImageNet | top-1 accuracy: 75.74 -> 80.09 (final) at about 2x speed-up | Improved |
| use VBP instead of NViT with the same 1-epoch pruning | NViT Base · ImageNet | top-1 accuracy: 81.92 -> 82.32 (final); 69.10 -> 72.37 (retention) | Improved |
| use VBP with 1 epoch instead of NViT with 50 epochs of pruning | NViT Base · ImageNet | top-1 accuracy: 82.18 -> 82.32 (final) | Improved |
| lower DeiT-B pruning rate from 55% to 20% (ablation) | VBP · ImageNet DeiT-B | top-1 accuracy before fine-tuning: 57.58 -> 80.87 (retention) | Improved |
| apply VBP to ConvNeXt instead of transformers | VBP · ImageNet | accuracy retained before fine-tuning: 16.80 (ConvNeXt-T) vs 64.44 (DeiT-S) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SNIP — SNIP: SINGLE-SHOT NETWORK PRUNING BASED ON CONNECTION SENSITIVITY (ICLR 2019) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Network Pruning via Performance Maximization (CVPR 2021 · similarity 0.91)
- Pruning from Scratch (AAAI 2020 · similarity 0.91)
- Connectivity Matters: Neural Network Pruning Through the Lens of Effective Sparsity (JMLR 2023 · similarity 0.90)
- Learning Pruning-Friendly Networks via Frank-Wolfe: One-Shot, Any-Sparsity, And No Retraining (ICLR 2022 · similarity 0.90)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.89)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 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.
