Research map / Architectures and efficiency / Neural architecture search
Zen-NAS: A Zero-Shot NAS for High-Performance Image Recognition
Ming Lin, Pichao Wang, Zhenhong Sun, Hesen Chen, Xiuyu Sun, Qi Qian, Hao Li, Rong Jin
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 Zen-Score instead of FLOPs as NAS proxy (ablation) | Zen-NAS evolutionary search · CIFAR-100 | top-1 accuracy: 64.7% -> 80.1% | Improved |
| use Zen-Score instead of synflow as NAS proxy (ablation) | Zen-NAS evolutionary search · CIFAR-100 | top-1 accuracy: 75.9% -> 80.1% | Improved |
| use Zen-Score instead of synflow as NAS proxy (ablation) | Zen-NAS evolutionary search · CIFAR-10 | top-1 accuracy: 95.1% -> 96.2% | Improved |
| use Zen-Score instead of TE-Score as NAS proxy (ablation) | Zen-NAS evolutionary search · CIFAR-10 | top-1 accuracy: 96.1% -> 96.2% | Improved |
| use Zen-Score instead of NASWOT as NAS proxy (ablation) | Zen-NAS evolutionary search · CIFAR-100 | top-1 accuracy: 77.5% -> 80.1% | Improved |
| use FLOPs or gradient-norm as proxy instead of random search (ablation) | NAS proxy · CIFAR-10 | top-1 accuracy: 93.5% -> 93.1% | Hurt |
| use Zen-Score instead of TE-Score for scoring speed | ResNet-50 scoring · time for 16 images at 224x224 | time (s): 0.77 -> 0.037 | Improved |
| use Zen-Score instead of NASWOT for scoring speed | ResNet-18 scoring · time for 16 images at 224x224 | time (s): 0.040 -> 0.012 | Improved |
| search with Zen-NAS instead of EfficientNet scaling (accuracy vs cost) | ImageNet ZenNet-1.2ms · ImageNet-1k, V100 FP16 batch 64 | top-1 accuracy / search cost: 83.6 at 0.5 GPU days vs 84.4 at 3800 GPU days | Improved |
| add enhanced training recipe to ResNet-152 | ResNet-152 · ImageNet-1k val | top-1 accuracy: 79.2% -> 82.3% | Improved |
| add BN layers back to deep vanilla network without rescaling (ablation) | Phi-score · networks of varying depth/width | score discrimination | Hurt |
| add BN-variance rescaling (Zen-Score) to the BN-network score (ablation) | Phi-score · deep vanilla networks | numerical stability: overflow after 30 layers without BN | Improved |
Closest papers by meaning
- Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective (ICLR 2021 · similarity 0.89)
- Zero-Cost Operation Scoring in Differentiable Architecture Search (AAAI 2023 · similarity 0.89)
- Zero-Shot Neural Architecture Search: Challenges, Solutions, and Opportunities (TPAMI 2024 · similarity 0.89)
- ParZC: Parametric Zero-Cost Proxies for Efficient NAS (AAAI 2025 · similarity 0.87)
- NEAR: A Training-Free Pre-Estimator of Machine Learning Model Performance (ICLR 2025 · similarity 0.87)
- L-SWAG: Layer-Sample Wise Activation with Gradients Information for Zero-Shot NAS on Vision Transformers (CVPR 2025 · 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.
