Research map / Architectures and efficiency / Neural architecture search
Generalizing Few-Shot NAS with Gradient Matching
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 gradient-matching splitting instead of random splitting | DARTS-based NAS on NASBench-201 · NASBench-201, CIFAR-10 | accuracy: 70.47% -> 93.95% | Helped |
| Use gradient-matching splitting instead of exhaustive splitting | DARTS-based NAS on NASBench-201 · NASBench-201, CIFAR-10 | accuracy: 88.55% -> 93.95% | Helped |
| Apply gradient-matching supernet splitting | DARTS · NASBench-201, CIFAR-10 test | accuracy: 88.55 -> 93.72 | Helped |
| Apply gradient-matching supernet splitting | SNAS · CIFAR-10, DARTS Space test | test error: 2.70 ± 0.05 -> 2.55 ± 0.16 | Helped |
| Apply gradient-matching supernet splitting | ProxylessNAS · ImageNet, MobileNet Space, top-1 | test error: 24.1 -> 23.4 | Helped |
| Apply gradient-matching supernet splitting | OFA · ImageNet, MobileNet Space, top-1 | test error: 20.2 -> 19.7 | Helped |
| Replace cosine similarity with L2 distance (ablation) | GM-NAS · NASBench-201, CIFAR-10 | accuracy: 92.52±0.93 vs 93.95±0.08 | Hurt |
| Use per-filter cosine similarity instead of cosine similarity (ablation) | GM-NAS · NASBench-201, CIFAR-10 | accuracy: 93.87±0.11 vs 93.95±0.08 | No clear effect |
| Choose split edges randomly rather than by gradient-matching cut cost (ablation) | GM-NAS · NASBench-201, CIFAR-10 | accuracy: 93.58% -> 93.95% | Hurt |
| Do not restart sub-supernets after the final split (ablation) | GM-NAS · NASBench-201, CIFAR-10 | accuracy: 91.04%; 2.91% lower than the original GM-NAS | Hurt |
| Use gradient-matching splitting rather than Few-Shot exhaustive splitting | NASBench-201 supernets trained with RSPS · NASBench-201, top 1% architectures | Spearman correlation: 0.117 -> 0.532 | Helped |
| Add more search budget and use three splits (ablation) | GM OFA · ImageNet, MobileNet Space, top-1 | test error: 19.7 -> 19.4 | Helped |
Closest papers by meaning
- Few-Shot Neural Architecture Search (ICML 2021 · similarity 0.95)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.94)
- GreedyNAS: Towards Fast One-Shot NAS With Greedy Supernet (CVPR 2020 · similarity 0.93)
- HEP-NAS: Towards Efficient Few-shot Neural Architecture Search via Hierarchical Edge Partitioning (AAAI 2025 · similarity 0.92)
- Subnet-Aware Dynamic Supernet Training for Neural Architecture Search (CVPR 2025 · similarity 0.91)
- Revisiting Neural Networks for Few-Shot Learning: A Zero-Cost NAS Perspective (ICML 2025 · 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.