Research map / Architectures and efficiency / Neural architecture search
Global Convergence of MAML and Theory-Inspired Neural Architecture Search for Few-Shot Learning
Check 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 MetaNTK condition number instead of supervised NTK condition number in NAS (ablation) | TE-NAS-style search · mini-ImageNet, 5-way 1-shot and 5-shot, 5- and 8-cell networks | accuracy: 62.51 ± 0.42 -> 63.88 ± 0.81; 63.51 ± 0.16 -> 64.26 ± 0.14; 79.02 ± 0.35 -> 80.07 ± 0.45; 79.86 ± 0.42 -> 80.35 ± 0.12 | Helped |
| add the MetaNTK criterion to search using linear-region count (ablation) | Linear Region search · mini-ImageNet, 5-way 1-shot and 5-shot, 5- and 8-cell networks | accuracy: 62.96 ± 1.05 -> 63.88 ± 0.81; 63.99 ± 0.76 -> 64.26 ± 0.14; 79.13 ± 0.96 -> 80.07 ± 0.45; 79.88 ± 0.51 -> 80.35 ± 0.12 | Helped |
| add MetaNTK-guided search rather than randomly sampling cells (ablation) | random cell search · mini-ImageNet, 5-way 1-shot and 5-shot, 5- and 8-cell networks | accuracy: 62.55 ± 1.15 -> 63.88 ± 0.81; 62.76 ± 0.83 -> 64.26 ± 0.14; 78.90 ± 0.58 -> 80.07 ± 0.45; 79.18 ± 0.72 -> 80.35 ± 0.12 | Helped |
| replace MetaNTK condition number with NTK condition number (ablation) | MetaNTK-NAS · mini-ImageNet, 5-way 1-shot and 5-shot, 5- and 8-cell networks | accuracy: 63.88 ± 0.81 -> 62.51 ± 0.42; 64.26 ± 0.14 -> 63.51 ± 0.16; 80.07 ± 0.45 -> 79.02 ± 0.35; 80.35 ± 0.12 -> 79.86 ± 0.42 | Hurt |
| remove both kernel condition number and MetaNTK criterion, retaining linear-region count only (ablation) | MetaNTK-NAS · mini-ImageNet, 5-way 1-shot and 5-shot, 5- and 8-cell networks | accuracy: 63.88 ± 0.81 -> 62.96 ± 1.05; 64.26 ± 0.14 -> 63.99 ± 0.76; 80.07 ± 0.45 -> 79.13 ± 0.96; 80.35 ± 0.12 -> 79.88 ± 0.51 | Hurt |
| remove the search criteria and sample candidate cells randomly (ablation) | MetaNTK-NAS · mini-ImageNet, 5-way 1-shot and 5-shot, 5- and 8-cell networks | accuracy: 63.88 ± 0.81 -> 62.55 ± 1.15; 64.26 ± 0.14 -> 62.76 ± 0.83; 80.07 ± 0.45 -> 78.90 ± 0.58; 80.35 ± 0.12 -> 79.18 ± 0.72 | Hurt |
| use MetaNTK-NAS instead of MetaNAS for architecture search | RFS-trained 5-cell networks · mini-ImageNet and tiered-ImageNet, 5-way 1-shot and 5-shot | accuracy: 64.24±0.11 -> 63.88 ± 0.81; 79.75±0.13 -> 80.07±0.45; 70.16±0.09 -> 71.12±0.49; 84.99±0.22 -> 85.71±0.22 | Mixed |
| use MetaNTK-NAS instead of MetaNAS for architecture search | RFS-trained 8-cell networks · mini-ImageNet and tiered-ImageNet, 5-way 1-shot and 5-shot | accuracy: 63.88±0.23 -> 64.26 ± 0.14; 79.88±0.14 -> 80.35±0.12; 72.32±0.02 -> 72.37±0.79; 86.48±0.06 -> 86.43±0.53 | Mixed |
| use MetaNTK-NAS instead of MetaNAS for search | 5-cell architecture search on mini-ImageNet · mini-ImageNet | search cost: 168 hr -> 1.54 hr | Helped |
Closest papers by meaning
- Revisiting Neural Networks for Few-Shot Learning: A Zero-Cost NAS Perspective (ICML 2025 · similarity 0.92)
- Meta-Learning of Neural Architectures for Few-Shot Learning (CVPR 2020 · similarity 0.92)
- Rapid Model Architecture Adaption for Meta-Learning (NeurIPS 2022 · similarity 0.92)
- Repurposing Pretrained Models for Robust Out-of-domain Few-Shot Learning (ICLR 2021 · similarity 0.91)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.89)
- Neural Fine-Tuning Search for Few-Shot Learning (ICLR 2024 · similarity 0.88)
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.