Research map / Architectures and efficiency / Neural architecture search
Neural Architecture Transfer
Zhichao Lu, Gautam Sreekumar, Erik D. Goodman, Wolfgang Banzhaf, Kalyanmoy Deb, Vishnu Naresh Boddeti
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 |
|---|---|---|---|
| add crossover operator to the evolutionary search (ablation) | NAT evolutionary search · ImageNet, CIFAR-10, Flowers102, Stanford Cars | top-1 accuracy (%) | Improved |
| raise crossover probability (vs lowering from 0.9 to 0.2) (ablation) | NAT evolutionary search · four datasets | top-1 accuracy (%) | Improved |
| lower mutation probability (ablation) | NAT evolutionary search · four datasets | top-1 accuracy (%) | Improved |
| use low mutation index eta_m (more distant offspring) (ablation) | NAT evolutionary search · four datasets | top-1 accuracy (%) | Improved |
| adapt the supernet instead of adapting each subnet (ablation) | NAT · CIFAR-10, CIFAR-100, Flowers102, STL-10 | top-1 accuracy (%) | Improved |
| additional 150-epoch fine-tuning of subnets vs supernet adaptation (ablation) | subnet adaptation · CIFAR-10 (4,000 per class) vs Flowers102 (20 per class) | top-1 accuracy (%) | Mixed |
| use RBF ensemble instead of standalone RBF accuracy predictor (ablation) | accuracy predictor · ImageNet | Spearman rank correlation: 0.705 -> 0.866 | Improved |
| use RBF ensemble instead of standalone RBF accuracy predictor (ablation) | accuracy predictor · CIFAR-10 | Spearman rank correlation: 0.969 -> 0.959 | No clear effect |
| use NAT evolutionary search instead of random search or NSGANet (ablation) | NAT pipeline · three datasets, bi-objective | hypervolume sample efficiency: 3x - 5x | Improved |
| use reference point based selection instead of domination based selection (ablation) | many-objective EA · DTLZ1 (3 to 15 objectives) | IGD | Improved |
| transfer architecture and weights with NAT instead of search on OnceForAll supernet | NAT · ImageNet | top-1 accuracy (%): 76.9 -> 77.5 | Improved |
| NAT-M1 vs MobileNetV3 at similar MAdds | NAT-M1 · ImageNet-1K | top-1 accuracy (%): 75.2 -> 77.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DARTS — DARTS-: Robustly Stepping out of Performance Collapse Without Indicators (ICLR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TransNAS-Bench-101: Improving Transferability and Generalizability of Cross-Task Neural Architecture Search (CVPR 2021 · similarity 0.92)
- Efficient Transfer Learning via Joint Adaptation of Network Architecture and Weight (ECCV 2020 · similarity 0.91)
- Towards Fast Adaptation of Neural Architectures with Meta Learning (ICLR 2020 · similarity 0.91)
- Boosting Order-Preserving and Transferability for Neural Architecture Search: a Joint Architecture Refined Search and Fine-tuning Approach (CVPR 2024 · similarity 0.90)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.89)
- Subnet-Aware Dynamic Supernet Training for Neural Architecture Search (CVPR 2025 · 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.
