Research map / Architectures and efficiency / Pruning and efficient tuning
DARB: A Density-Adaptive Regular-Block Pruning for Deep Neural Networks
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 density-adaptive regular blocks instead of uniform block pruning | DARB · LSTM decoder hardware | decoding efficiency: 103 -> 1478 | Helped |
| Use density-adaptive regular blocks instead of uniform block pruning | DARB · LSTM decoder hardware | pruning ratio: 7.48× -> 13.14× | Helped |
| Select the maximum-magnitude weight in each block | BMWM · LSTM weight salience | relative difference: 45.3% -> 8.8% | Helped |
| Adapt block size to each row's density | DARB · PTB | pruning ratio at similar perplexity: 3.03× -> 13.14× | Helped |
| Prune the full model, including the embedding layer | DARB · PTB | pruning ratio: 3.03× -> 13.14× | Helped |
| Use DARB rather than BBS at 8× pruning | LSTMP · TIMIT | PER degradation: 0.25% -> 0.1% | Helped |
| Use DARB rather than C-LSTM at 16× pruning | LSTMP · TIMIT | PER degradation: 1.32% -> 0.2% | Helped |
| Use DARB rather than block-circulant pruning | GRU · TIMIT | PER degradation: 0.6% -> no degradation | Helped |
| Use DARB-I rather than PermDNN | AlexNet FC6-FC8 · ImageNet | pruning ratio: 9.0× -> 21.3× | Helped |
| Increase DARB pruning from DARB-I to DARB-II | AlexNet-BN FC6-FC8 · ImageNet | Top-5 accuracy: 83.4% -> 83.2% | Hurt |
| Allow more pruning with DARB-II | AlexNet-BN FC6-FC8 · ImageNet | pruning ratio: 21.3× -> 25.0× | Helped |
| Allow different row densities rather than enforcing uniform row density | DARB · Neural network weight matrices | model accuracy | Helped |
Closest papers by meaning
- Structured Pruning of Neural Networks With Budget-Aware Regularization (CVPR 2019 · similarity 0.90)
- “Learning-Compression” Algorithms for Neural Net Pruning (CVPR 2018 · similarity 0.89)
- Data-Efficient Structured Pruning via Submodular Optimization (NeurIPS 2022 · similarity 0.89)
- Dynamic Structure Pruning for Compressing CNNs (AAAI 2023 · similarity 0.89)
- Computation and Memory-Efficient Model Compression with Gradient Reweighting (NeurIPS 2025 · similarity 0.89)
- Learning to Prune Deep Neural Networks via Layer-wise Optimal Brain Surgeon (NeurIPS 2017 · 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.