Research map / Architectures and efficiency / Neural architecture search
NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size
Xuanyi Dong, Lu Liu, Katarzyna Musiał, Bogdan Gabryś
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 REA instead of random search | multi-trial NAS · CIFAR-100 (topology space) | test accuracy: 71.55 -> 72.23 | Improved |
| use REA instead of random search | multi-trial NAS · ImageNet-16-120 (size space) | test accuracy: 45.42 -> 45.94 | Improved |
| use second-order instead of first-order DARTS | DARTS · CIFAR-10 (topology space) | test accuracy: 59.84 -> 65.38 | Improved |
| use second-order instead of first-order DARTS | DARTS · ImageNet-16-120 (topology space) | test accuracy: 37.88 -> 36.79 | Hurt |
| search on CIFAR-100 instead of CIFAR-10 | DARTS · topology space | test accuracy: 59.84 -> 61.26 | Improved |
| sample candidates with shared weights and select best validation instead of argmax over architecture parameters | weight-sharing NAS (RSPS/ENAS/SETN) · CIFAR-10 (topology space) | test accuracy: 59.84 -> 93.76 | Improved |
| use channel-wise interpolation instead of masking + sampling | size-space weight-sharing search · ImageNet-16-120 (size space) | test accuracy: 45.11 -> 47.17 | Improved |
| use channel-wise interpolation instead of masking + sampling | size-space weight-sharing search · CIFAR-10 (size space) | test accuracy: 92.78 -> 93.40 | Improved |
| use 200 epochs instead of 12 epochs for trial feedback (ablation) | REA · CIFAR-10 (topology space) | test accuracy of discovered architecture: after 1.2e6 seconds still worse than 12 epochs | Hurt |
| use 90 epochs instead of 12 epochs for trial feedback (ablation) | REA · CIFAR-10 (size space) | test accuracy of discovered architecture: outperforms after about 8e4 seconds | Mixed |
| use weight sharing instead of multi-trial search | NAS algorithms · CIFAR-100 (topology space) | test accuracy: 72.23 -> 68.26 | Hurt |
| use REA search instead of largest candidate | size space · ImageNet-16-120 | test accuracy: 47.17 -> 45.94 | Hurt |
Closest papers by meaning
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search (ICLR 2020 · similarity 0.95)
- NAS-Bench-360: Benchmarking Neural Architecture Search on Diverse Tasks (NeurIPS 2022 · similarity 0.94)
- NAS-Bench-Suite: NAS Evaluation is (Now) Surprisingly Easy (ICLR 2022 · similarity 0.93)
- NAS-Bench-101: Towards Reproducible Neural Architecture Search (ICML 2019 · similarity 0.93)
- Surrogate NAS Benchmarks: Going Beyond the Limited Search Spaces of Tabular NAS Benchmarks (ICLR 2022 · similarity 0.92)
- NAS-Bench-Graph: Benchmarking Graph Neural Architecture Search (NeurIPS 2022 · similarity 0.92)
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.
