Research map / Architectures and efficiency / Neural architecture search
NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search
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 NAS methods without parameter sharing rather than methods with parameter sharing | NAS-Bench-201 search on CIFAR-10 · CIFAR-10, CIFAR-100, and ImageNet-16-120 | accuracy | Helped |
| Use batch statistics for BN instead of accumulated running estimates (ablation) | parameter-sharing NAS methods · CIFAR-10 | one-shot validation accuracy correlation: 0.0039 -> 0.8141 | Helped |
| Use batch statistics for BN instead of accumulated running estimates (ablation) | parameter-sharing NAS methods · CIFAR-10 | one-shot validation accuracy correlation: 0.9049 -> 0.0862 | Helped |
| Use Gumbel-softmax sampling when optimizing the architecture encoding | GDAS compared with DARTS · CIFAR-10 | probability correlation with ground-truth accuracy: 0.0862 -> 0.2714 | Helped |
| Train shared parameters with uniform architecture sampling rather than joint optimization | parameter-sharing NAS · CIFAR-10 | one-shot validation accuracy correlation | Helped |
| Use the 12-epoch converged cosine schedule instead of evaluating at an early point on a 200-epoch cosine schedule (ablation) | architecture ranking for bandit-based NAS · CIFAR-10 validation accuracy correlated with full-training accuracy | correlation coefficient: 0.7767 -> 0.9110 | Helped |
| Continue training along the 200-epoch cosine schedule from 12 to 175 epochs (ablation) | architecture ranking · CIFAR-10 validation accuracy correlated with full-training accuracy | correlation coefficient: 0.7520 -> 0.9834 | Helped |
| Use the 200-epoch schedule rather than stopping at 175 epochs (ablation) | architecture ranking · CIFAR-10 validation accuracy correlated with full-training accuracy | correlation coefficient: 0.9834 -> 0.9993 | Hurt |
Closest papers by meaning
- NAS-Bench-101: Towards Reproducible Neural Architecture Search (ICML 2019 · similarity 0.95)
- NAS evaluation is frustratingly hard (ICLR 2020 · similarity 0.95)
- NAS-Bench-360: Benchmarking Neural Architecture Search on Diverse Tasks (NeurIPS 2022 · similarity 0.94)
- einspace: Searching for Neural Architectures from Fundamental Operations (NeurIPS 2024 · similarity 0.91)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.91)
- Insights from the Use of Previously Unseen Neural Architecture Search Datasets (CVPR 2024 · similarity 0.91)
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.