Research map / Architectures and efficiency / Neural architecture search
K-shot NAS: Learnable Weight-Sharing for NAS with K-shot Supernets
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 |
|---|---|---|---|
| replace one shared weight set with K-shot supernets and architecture-customized simplex codes | NAS supernet path evaluation · NAS-Bench-201 CIFAR-10 | Kendall's Tau: 55.00% -> 62.64% | Helped |
| add architecture-customized simplex-net codes to K-shot supernets (ablation) | K-shot NAS · ImageNet 420M FLOPs | Top-1 accuracy: 77.2% -> 77.5% | Helped |
| add channel-width search to K-shot NAS (ablation) | K-shot NAS with simplex-net · ImageNet 420M FLOPs | Top-1 accuracy: 77.5% -> 77.6% | Helped |
| add non-parametric channel-width regularization (ablation) | K-shot NAS with channel search · ImageNet 420M FLOPs | Top-1 accuracy: 77.6% -> 77.9% | Helped |
| use channel-width search with ordinary one-shot sharing (ablation) | One-shot NAS · ImageNet 420M FLOPs | Top-1 accuracy: 77.0% -> 77.1% | Helped |
| replace learned codes with randomly initialized, fixed codes (ablation) | K-shot supernets · ImageNet 330M and 420M FLOPs | Top-1 accuracy: 76.2 -> 77.2; 77.1 -> 77.6 | Hurt |
| replace learned codes with fixed 1/K codes (ablation) | K-shot supernets · ImageNet 330M and 420M FLOPs | Top-1 accuracy: 76.4 -> 77.2; 77.2 -> 77.6 | Hurt |
| increase K from 1 to 8 (ablation) | K-shot NAS supernets · NAS-Bench-201 CIFAR-10 | Kendall's Tau: 55.02 ± 0.27 -> 62.64 ± 0.17 | Helped |
| increase K from 8 to 12 (ablation) | K-shot NAS supernets · NAS-Bench-201 CIFAR-10 | Kendall's Tau: 62.64 ± 0.17 -> 62.13 ± 0.13 | Hurt |
| jointly optimize supernets and simplex-net rather than train them iteratively (ablation) | K-shot NAS · ImageNet | Top-1 accuracy | Hurt |
Closest papers by meaning
- PA&DA: Jointly Sampling Path and Data for Consistent NAS (CVPR 2023 · similarity 0.89)
- GreedyNAS: Towards Fast One-Shot NAS With Greedy Supernet (CVPR 2020 · similarity 0.89)
- Subnet-Aware Dynamic Supernet Training for Neural Architecture Search (CVPR 2025 · similarity 0.87)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.87)
- HEP-NAS: Towards Efficient Few-shot Neural Architecture Search via Hierarchical Edge Partitioning (AAAI 2025 · similarity 0.86)
- AZ-NAS: Assembling Zero-Cost Proxies for Network Architecture Search (CVPR 2024 · similarity 0.84)
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.