Research map / Architectures and efficiency / Pruning and efficient tuning
SPDY: Accurate Pruning with Speedup Guarantees
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 SPDY-generated layer sparsity targets instead of the strongest uniform or GMP profile | gradual pruning of ResNet50, MobileNetV1, YOLOv5, and BERT · all speedup settings in Table: Gradual Pruning Results | accuracy or mAP@0.5: 76.01 -> 76.39; 61.33 -> 71.38; 50.90 -> 53.10; 88.22 -> 88.53 | Helped |
| Regenerate the SPDY profile from an already gradually pruned intermediate model (ablation) | BERT SQuAD pruning · 4.00× speedup | F1: 86.37 -> 87.14 | Helped |
| Replace learned error scores with layer-wise loss scores (ablation) | DP pruning profiles · ResNet34 and BERT SQuAD, 2.50× and 3.50× speedups | accuracy: 68.23 -> 65.92; 51.68 -> 46.11; 74.59 -> 30.59; 18.83 -> 6.13 | Hurt |
| Remove the reconstruction database from SPDY profile search (ablation) | SPDY one-shot pruning · ResNet50, 2.00× and 3.00× speedups | accuracy: 71.17 -> 64.59; 26.81 -> 00.67 | Hurt |
| Use global AdaPrune after standard AdaPrune | post-training 2:4 pruning of YOLOv5 · YOLOv5 models | mAP@0.5: 13.40 -> 37.10; 40.70 -> 51.60; 54.00 -> 61.20; 59.80 -> 65.40 | Helped |
| Use SPDY profiles instead of uniform profiles before quantization-aware training | ResNet50 with pruning and QAT · 2× pruning speedup followed by QAT | accuracy: 73.33 vs. 71.94; 75.00 vs. 74.33 | Helped |
| Use iterative AdaPrune reconstruction instead of no-reconstruction layer-wise loss (ablation) | DP-generated profiles · ResNet34, 2.00× speedup | accuracy: 54.51 -> 65.92 | Helped |
| Use iterative AdaPrune reconstruction instead of no-reconstruction layer-wise loss (ablation) | DP-generated profiles · ResNet50, 3.00× speedup | accuracy: 08.85 -> 20.22 | Helped |
| Use SPDY learned scores instead of squared-weight scores (ablation) | DP-generated profiles · ResNet34 and BERT SQuAD, tested speedups | accuracy: 54.18 -> 68.23; 31.59 -> 51.68; 71.20 -> 74.59; 07.88 -> 18.83 | Mixed |
| Use SPDY's DP-reparameterized local search instead of direct constrained search (ablation) | ResNet50 profile generation · 2.5× speedup | calibration loss: ≈ 10× fewer evaluations | Helped |
| Use SPDY's local search instead of the tested genetic algorithm (ablation) | ResNet50 profile generation · 2.5× speedup | calibration loss | Helped |
Closest papers by meaning
- Computation and Memory-Efficient Model Compression with Gradient Reweighting (NeurIPS 2025 · similarity 0.92)
- Learnable Sparsity for Vision Generative Models (ICLR 2026 · similarity 0.90)
- Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision (ICML 2025 · similarity 0.90)
- AutoCompress: An Automatic DNN Structured Pruning Framework for Ultra-High Compression Rates (AAAI 2020 · similarity 0.90)
- Dynamic Model Pruning with Feedback (ICLR 2020 · similarity 0.90)
- Structured Pruning of Neural Networks With Budget-Aware Regularization (CVPR 2019 · similarity 0.90)
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.