Research map / Architectures and efficiency / Neural architecture search
Distilling Optimal Neural Networks: Rapid Search in Diverse Spaces
Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, Tijmen Blankevoort
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 |
|---|---|---|---|
| use the DONNA linear accuracy predictor instead of the DNA ranking model | DNA ranking · DONNA search space | Kendall-Tau: 0.75 -> 0.91 | Improved |
| initialize finetuning from blockwise distillation weights instead of training from scratch | DONNA architectures · ImageNet | training epochs: 450 -> 50 | Improved |
| increase architecture library from 10 to 30 targets (ablation) | DONNA predictor · DONNA search space | Kendall-Tau: 0.87 -> 0.9 | Improved |
| increase architecture library from 10 to 40 targets (ablation) | DONNA predictor · MobileNetV3 (1.2x) search space | Kendall-Tau: 0.79 -> 0.82 | Improved |
| use NSR quality metric instead of network-level top-1 accuracy (ablation) | DONNA predictor · DONNA search space | MSE (top-1 %): 0.44 -> 0.19 | Improved |
| use Ridge regression instead of LARS as predictor (ablation) | DONNA predictor · DONNA search space | Kendall-Tau: 0.16 -> 0.91 | Improved |
| use Ridge regression instead of SVM as predictor (ablation) | DONNA predictor · DONNA search space | MSE (top-1 %): 1.84 -> 0.15 | Improved |
| allow optimal placement of SE/Swish attention in the search space (E vs D) (ablation) | DONNA search space · V100 latency exploration | accuracy: +1% at 26ms | Improved |
| add SE-attention optimally (C vs D) in the search space (ablation) | DONNA search space · V100 latency exploration | latency: 25% speedup at 76% accuracy | Improved |
| apply FP-trained predictor to INT8-quantized networks (ablation) | DONNA predictor · MobileNetV3 (1.2x) | Kendall-Tau: 0.85 -> 0.86 | No clear effect |
| use OFA progressive-shrinking weights instead of from-scratch weights for the detection backbone (ablation) | OFA backbone · COCO EfficientDet-D0 | mAP: +1.4% top-1 -> +0.1% mAP | No clear effect |
| search DONNA backbones directly on COCO instead of transferring from ImageNet | DONNA backbones · COCO EfficientDet-D0 | mAP | Improved |
Closest papers by meaning
- DA-NAS: Data Adapted Pruning for Efficient Neural Architecture Search (ECCV 2020 · similarity 0.91)
- Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective (ICLR 2021 · similarity 0.91)
- DDPNAS: Efficient Neural Architecture Search via Dynamic Distribution Pruning (IJCV 2023 · similarity 0.90)
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search (ICLR 2020 · similarity 0.90)
- Block-Wisely Supervised Neural Architecture Search With Knowledge Distillation (CVPR 2020 · similarity 0.90)
- MiLeNAS: Efficient Neural Architecture Search via Mixed-Level Reformulation (CVPR 2020 · 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.
