Research map / Architectures and efficiency / Neural architecture search
Towards Accurate and Compact Architectures via Neural Architecture Transformer
Yong Fang Guo, Yin Zheng, Mingkui Tan, Qi Chen, Zhipeng Li, Jian Chen, Peilin Zhao, Junzhou Huang
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 |
|---|---|---|---|
| enlarge transition space from NAT to NAT++ (two-level type and kernel transitions) | NAT on VGG16 · CIFAR-10 | accuracy (%): 96.04 -> 96.16 | Improved |
| apply NAT++ to ResNet56 | ResNet56 · CIFAR-10 | accuracy (%): 93.21 -> 95.47 | Improved |
| apply NAT++ to DARTS architecture | DARTS · ImageNet | top-1 accuracy (%): 73.1 -> 73.9 | Improved |
| apply NAT++ to MobileNetV2 (already efficient ops) | MobileNetV2 · ImageNet | top-1 accuracy (%): 72.0 -> 72.5 | No clear effect |
| enlarge transition space from NAT to NAT++ on random architectures (ablation) | NAT · CIFAR-10 (20 random architectures) | test accuracy (%): 96.56 -> 96.79 | Improved |
| enlarge transition space from NAT to NAT++ on random architectures (ablation) | NAT · CIFAR-10 (20 random architectures) | #Params (M): 4.67 -> 3.66 | Improved |
| increase m from 1 to 10 in NAT (ablation) | NAT · CIFAR-10 (20 random architectures) | accuracy (%): 96.56 -> 96.61 | No clear effect |
| increase m from 1 to 30 in NAT (ablation) | NAT · CIFAR-10 (20 random architectures) | search cost (GPU hour): 5.3 -> 122.7 | Hurt |
| increase n from 1 to 30 in NAT++ (ablation) | NAT++ · CIFAR-10 (20 random architectures) | accuracy (%): 96.79 -> 96.84 | No clear effect |
| increase n from 1 to 30 in NAT++ (ablation) | NAT++ · CIFAR-10 (20 random architectures) | search cost (GPU hour): 5.7 -> 86.7 | Hurt |
| use 1-layer GCN instead of 2-layer GCN (ablation) | NAT/NAT++ controller · CIFAR-10 (20 random architectures) | accuracy | Hurt |
| use 5 or 10 GCN layers instead of 2 (ablation) | NAT/NAT++ controller · CIFAR-10 (20 random architectures) | accuracy | Hurt |
Closest papers by meaning
- NAT: Neural Architecture Transformer for Accurate and Compact Architectures (NeurIPS 2019 · similarity 0.97)
- Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective (ICLR 2021 · similarity 0.89)
- Learning Versatile Neural Architectures by Propagating Network Codes (ICLR 2022 · similarity 0.88)
- Evaluating The Search Phase of Neural Architecture Search (ICLR 2020 · similarity 0.88)
- NetAdaptV2: Efficient Neural Architecture Search With Fast Super-Network Training and Architecture Optimization (CVPR 2021 · similarity 0.88)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 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.
