Research map / Architectures and efficiency / Quantization
Constraint-Aware Deep Neural Network Compression
Changan Chen, Frederick Tung, Naveen Vedula, Greg Mori
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 a cooling schedule (exponential) to constraint-aware compression instead of unconstrained search | constraint-aware compression · DTD (60 ms) | top-1 accuracy: 56.3 -> 59.0 | Improved |
| add exponential cooling instead of one-step compression (ablation) | constraint-aware compression · DTD (60 ms) | top-1 accuracy: 53.9 ± 6.0% -> 59.0 ± 0.4% | Improved |
| use exponential cooling rather than linear cooling (ablation) | constraint-aware compression · DTD (60 ms) | top-1 accuracy: 57.8 ± 1.0% -> 59.0 ± 0.4% | Improved |
| add linear cooling instead of one-step compression (ablation) | constraint-aware compression · DTD (60 ms) | top-1 accuracy: 53.9 ± 6.0% -> 57.8 ± 1.0% | Improved |
| transfer only the final-step hyperparameters from exponential cooling into the one-step baseline (ablation) | one-step constraint-aware compression · DTD (60 ms) | top-1 accuracy: 59.0 -> 52.6 ± 2.4% | Hurt |
| also transfer the sparsity structure from exponential cooling to the one-step baseline (ablation) | one-step constraint-aware compression · DTD (60 ms) | top-1 accuracy: 52.6 -> 58.3 ± 1.2% | Improved |
| add one-epoch look-ahead fine-tuning in Bayesian optimization for the one-step baseline (ablation) | one-step constraint-aware compression · DTD (60 ms) | top-1 accuracy: 53.9% -> 55.9%, std 6.0% -> 1.1% | Improved |
| add constraint-aware compression with exponential cooling instead of unconstrained compression | AlexNet pruning · ImageNet (70 ms) | top-1 accuracy: 50.62 -> 53.70 | Improved |
| add exponential cooling instead of one-step (ablation) | constraint-aware compression · ImageNet (70 ms) | top-1 accuracy: 47.27 ± 3.31% -> 53.70 ± 0.15% | Improved |
| constraint-aware compression with latency target of 78.2 ms vs Guided Sparsity Learning | AlexNet pruning · ImageNet (GSL comparison) | top-1 accuracy: 57.5 -> 57.4 | No clear effect |
| constraint-aware compression vs Guided Sparsity Learning at aggressive latency targets | AlexNet pruning · DTD (GSL comparison) | top-1 accuracy: 60.9% at 74.2 ms vs 60.1% at 70 ms and 59.0% at 60 ms | No clear effect |
| switch constraint from latency to storage; layers pruned adapt automatically | constraint-aware compression · ImageNet (storage 5%) | top-1 accuracy: 57.41 -> 54.84% at 232.56 MB -> 11.33 MB | No clear effect |
Closest papers by meaning
- Bayesian Compression for Deep Learning (NeurIPS 2017 · similarity 0.89)
- Compression-aware Training of Deep Networks (NeurIPS 2017 · similarity 0.88)
- Discrete Model Compression With Resource Constraint for Deep Neural Networks (CVPR 2020 · similarity 0.87)
- A Bayesian Optimization Framework for Neural Network Compression (ICCV 2019 · similarity 0.87)
- Compressing Neural Networks: Towards Determining the Optimal Layer-wise Decomposition (NeurIPS 2021 · similarity 0.87)
- Auto-Compressing Networks (NeurIPS 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.
