Research map / Architectures and efficiency / Neural architecture search
Zero-Shot Neural Architecture Search: Challenges, Solutions, and Opportunities
Guihong Li, Duc Hoang, Kartikeya Bhardwaj, Ming Lin, Zhangyang Wang, Radu Mărculescu
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 #Params instead of Zen-score to pick architecture | zero-shot NAS · NASBench-201 ImageNet16-120 | best accuracy (%): 40.77 -> 41.44 | Improved |
| use #Params instead of Synflow | zero-shot NAS · NASBench-201 CIFAR-100 | best accuracy (%): 62.22 -> 71.11 | Improved |
| use Zen-score instead of #Params | zero-shot NAS · NATS-Bench ImageNet16-120 | best accuracy (%): 44.73 -> 46.27 | Improved |
| restrict correlation to top 5% architectures | existing proxies · NASBench-201 and NATS-Bench | rank correlation | Hurt |
| use Jacob_cov or Zen-score instead of #Params on non-classification tasks | zero-shot proxies · TransNAS-Bench-101 | correlation | Improved |
| use #Params instead of Fisher for best architecture | zero-shot NAS · TransNAS-Bench-101 autoencoding | SSIM: 0.49 -> 0.46 | Hurt |
| use one-shot NAS instead of zero-shot #Params | NAS · ImageNet-1K ProxylessNAS space | top-1 accuracy (%): 72.87 -> 74.39 | Improved |
| use NTK_Cond instead of #Params | zero-shot NAS · ImageNet-1K ProxylessNAS space | top-1 accuracy (%): 72.87 -> 73.63 | Improved |
| use #Params instead of NTK_Cond | zero-shot NAS · ImageNet-1K ProxylessNAS space | search cost (GPU hours): 37 -> 0.03 | Improved |
| use zero-shot #Params instead of one-shot NAS | NAS · ImageNet-1K ProxylessNAS space | search cost (GPU hours): 200 -> 0.03 | Improved |
| use #FLOPs instead of #Params | zero-shot NAS · COCO detection | mAP: 0.26 -> 0.28 | Improved |
| use proxies within ResNet and Wide-ResNet families | SNIP, Zen-score, #Params, #FLOPs, NN-Mass · ImageNet-1K ResNet family | Spearman rho: rho > 0.9 | Improved |
Closest papers by meaning
- ZiCo: Zero-shot NAS via inverse Coefficient of Variation on Gradients (ICLR 2023 · similarity 0.93)
- ParZC: Parametric Zero-Cost Proxies for Efficient NAS (AAAI 2025 · similarity 0.91)
- L-SWAG: Layer-Sample Wise Activation with Gradients Information for Zero-Shot NAS on Vision Transformers (CVPR 2025 · similarity 0.91)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.90)
- Revisiting Neural Networks for Few-Shot Learning: A Zero-Cost NAS Perspective (ICML 2025 · similarity 0.89)
- AZ-NAS: Assembling Zero-Cost Proxies for Network Architecture Search (CVPR 2024 · similarity 0.89)
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.
