Research map / Architectures and efficiency / Pruning and efficient tuning
Practical Network Acceleration With Tiny Sets: Hypothesis, Theory, and Algorithm
Guo-Hua Wang, Jianxin Wu
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 |
|---|---|---|---|
| drop blocks instead of pruning filters (BP baseline) | ResNet-34 finetuned with cross-entropy · ImageNet-1k, 16% latency, 50 images | Top-1 accuracy (%): 39.0 -> 66.5 | Improved |
| drop blocks instead of pruning filters (BP baseline) | ResNet-34 finetuned with cross-entropy · ImageNet-1k, 22.1% latency, 50 images | Top-1 accuracy (%): 24.2 -> 60.6 | Improved |
| use Practise (recoverability-based block selection plus feature mimicking) instead of MiR | ResNet-34 · ImageNet-1k, 16% latency, 50 images | Top-1 accuracy (%): 64.1 -> 70.3 | Improved |
| use Practise instead of MiR at larger latency reduction | ResNet-34 · ImageNet-1k, 22.1% latency, 500 images | Top-1 accuracy (%): +3.6% | Improved |
| use Practise instead of MiR at small latency reduction | ResNet-34 · ImageNet-1k, 8.2% latency, 500 images | Top-1 accuracy (%): +1.1% | Improved |
| use Practise instead of MiR on MobileNetV2 | MobileNetV2 · ImageNet-1k, ~16% latency, 500 images | Top-1 accuracy (%): 67.6 -> 69.3 | Improved |
| replace filter-pruned data-free baseline ADI* with Practise on synthesized images | ResNet-50 · ImageNet-1k, data-free | Top-1 accuracy (%): 73.5 -> 74.8 | Improved |
| replace filter-pruned data-free baseline ADI* with Practise on synthesized images | MobileNetV2 · ImageNet-1k, data-free | Top-1 accuracy (%): 62.8 -> 68.0 | Improved |
| add recoverability as block-selection criterion (vs CURL) with 5 dropped blocks (ablation) | Practise · ImageNet-1k, ResNet-34, 500 images | Top-1 accuracy (%): 64.97 -> 69.35 | Improved |
| add recoverability as block-selection criterion (vs L2 distance) with 4 dropped blocks (ablation) | Practise · ImageNet-1k, ResNet-34, 500 images | Top-1 accuracy (%): 68.61 -> 70.86 | Improved |
| add recoverability criterion; gap to other criteria with 1 dropped block (ablation) | Practise · ImageNet-1k, ResNet-34, 500 images | Top-1 accuracy (%): 72.33 -> 73.02 | Improved |
| increase out-of-domain training images from 50 to all (Place365) | Practise · ImageNet-1k, ResNet-50 pruned 21% | Top-1 accuracy (%): 72.80 -> 75.21 | Improved |
Closest papers by meaning
- Practical Network Acceleration With Tiny Sets (CVPR 2023 · similarity 0.97)
- The Unreasonable Effectiveness of Random Pruning: Return of the Most Naive Baseline for Sparse Training (ICLR 2022 · similarity 0.88)
- Winning the Lottery Ahead of Time: Efficient Early Network Pruning (ICML 2022 · similarity 0.88)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.87)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.87)
- Preserving Deep Representations in One-Shot Pruning: A Hessian-Free Second-Order Optimization Framework (ICLR 2025 · similarity 0.86)
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.
