Research map / Architectures and efficiency / Pruning and efficient tuning
AutoCompress: An Automatic DNN Structured Pruning Framework for Ultra-High Compression Rates
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 |
|---|---|---|---|
| combine filter and column pruning rather than use filter pruning alone (ablation) | ADMM-based pruning on VGG-16 · CIFAR-10 | CONV Params Rt.: 9.3× -> 44.3× | Helped |
| combine filter and column pruning rather than use filter pruning alone (ablation) | ADMM-based pruning on ResNet-18 · CIFAR-10 | CONV Params Rt.: 5.2× -> 43.3× | Helped |
| use combined filter-and-column pruning instead of filter pruning, with the same ADMM and SA setup (ablation) | VGG-16 · ImageNet | Params Rt.: 3.8× -> 6.4× | Helped |
| add enhanced SA hyperparameter search instead of manual hyperparameter determination (ablation) | combined structured pruning on VGG-16 · CIFAR-10 | CONV Params Rt.: 44.3× -> 52.2× | Helped |
| replace DRL hyperparameter determination with enhanced SA (ablation) | combined structured pruning on ResNet-18 · CIFAR-10 | CONV Params Rt.: 11.8× -> 54.2× | Helped |
| add purification after ADMM regularization by removing low-norm columns and filters (ablation) | ADMM-based structured pruning · CIFAR-10 and ImageNet | accuracy | No clear effect |
| train from scratch using the combined-pruned structure instead of retaining pruned weights (ablation) | combined structured pruning · CIFAR-10 | accuracy: 93.21% -> 91.4%; 93.81% -> 91.88% | Hurt |
| train from scratch using the filter-pruned structure instead of retaining pruned weights (ablation) | filter-only pruning · CIFAR-10 | accuracy: 93.22% -> 93.19%; 93.91% -> 93.89% | No clear effect |
| optimize FLOPs reduction rather than parameter reduction (ablation) | Full AutoCompress on VGG-16 · CIFAR-10 | CONV FLOPs Rt.: 8.8× -> 10.8× | Helped |
| optimize FLOPs reduction rather than parameter reduction (ablation) | Full AutoCompress on VGG-16 · CIFAR-10 | accuracy: 93.21% -> 92.65% | Hurt |
| increase compression beyond the reported 54.2× parameter-reduction setting (ablation) | Full AutoCompress on ResNet-18 · CIFAR-10 | CONV Params Rt.: 54.2× -> 80.8× | Helped |
| increase compression beyond the reported 54.2× parameter-reduction setting (ablation) | Full AutoCompress on ResNet-18 · CIFAR-10 | accuracy: 93.81% -> 92.98% | Hurt |
Closest papers by meaning
- AutoPrune: Automatic Network Pruning by Regularizing Auxiliary Parameters (NeurIPS 2019 · similarity 0.91)
- Computation and Memory-Efficient Model Compression with Gradient Reweighting (NeurIPS 2025 · similarity 0.90)
- SPDY: Accurate Pruning with Speedup Guarantees (ICML 2022 · similarity 0.90)
- Numerical Pruning for Efficient Autoregressive Models (AAAI 2025 · similarity 0.89)
- Automatic Joint Structured Pruning and Quantization for Efficient Neural Network Training and Compression (CVPR 2025 · similarity 0.88)
- Auto-Train-Once: Controller Network Guided Automatic Network Pruning from Scratch (CVPR 2024 · similarity 0.88)
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.