Research map / Architectures and efficiency / Neural architecture search
On Network Design Spaces for Visual Recognition
Ilija Radosavovic, Justin Johnson, Saining Xie, Wan-Yen Lo, Piotr Dollar
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 |
|---|---|---|---|
| compare design spaces by distributions instead of minimum error (point estimate) | ResNet design space, 100 vs 1000 samples · CIFAR-10 | conclusion reliability: In 90% of cases M has lower minimum than B; KS D=0.079, p=0.60 | Improved |
| normalize error distributions by model complexity (params or flops) | ResNeXt-A vs ResNeXt-B comparison · CIFAR-10 | error EDF gap | Improved |
| add residual connections (ResNet vs Vanilla design space) (ablation) | Vanilla design space · CIFAR-10 | fraction of models with error under 8%: ~15% -> >80% | Improved |
| use ResNet instead of Vanilla design space for random search (ablation) | random search · CIFAR-10 | random search efficiency | Improved |
| increase number of sampled models for EDF estimation (10 to 100 to 1000) (ablation) | ResNet design space EDF · CIFAR-10 | KS statistic D vs full 25k sample: D about 0.1 at 100 samples; saturates at 1000 | Improved |
| use newer NAS design spaces (DARTS) instead of NASNet and Amoeba | NAS design spaces · CIFAR-10 | error EDF: max KS D=0.51 between DARTS and NASNet | Improved |
| use ResNeXt-B design space instead of DARTS (params normalized) | standard vs NAS design space · CIFAR-10 (normalized for params) | error EDF: KS D=0.09 | No clear effect |
| use DARTS instead of ResNeXt-B (flops normalized) | standard vs NAS design space · CIFAR-10 (normalized for flops) | error EDF | Improved |
| use enhanced training setup (600 epochs, deep supervision, Cutout, DropPath) | ResNet-110 · CIFAR-10 | test error (%): 5.91 -> 3.65 | Improved |
| use enhanced training setup | ResNeXt* · CIFAR-10 | test error (%): 4.90 -> 2.75 | Improved |
| use enhanced training setup | DARTS* · CIFAR-10 | test error (%): 5.21 -> 2.63 | Improved |
| average error over 1 to 3 training runs (ablation) | error EDFs of 5k models · CIFAR-10 | KS statistic D: D<0.009 | No clear effect |
Closest papers by meaning
- einspace: Searching for Neural Architectures from Fundamental Operations (NeurIPS 2024 · similarity 0.88)
- Exploring Randomly Wired Neural Networks for Image Recognition (ICCV 2019 · similarity 0.88)
- NAS evaluation is frustratingly hard (ICLR 2020 · similarity 0.88)
- TransNAS-Bench-101: Improving Transferability and Generalizability of Cross-Task Neural Architecture Search (CVPR 2021 · similarity 0.88)
- Teacher Guided Architecture Search (ICCV 2019 · similarity 0.88)
- Insights from the Use of Previously Unseen Neural Architecture Search Datasets (CVPR 2024 · similarity 0.87)
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.
