Research map / Architectures and efficiency / Neural architecture search
NASTransfer: Analyzing Architecture Transferability in Large Scale Neural Architecture Search
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 |
|---|---|---|---|
| Search on CIFAR10 rather than CIFAR100 | DARTS transferred to ImageNet1K · ImageNet1K | Top-1 accuracy: 71.3±0.78 versus 63.37±4.99 | Helped |
| Use CIFAR10 as the search proxy | NAO transferred to ImageNet22K · ImageNet22K | Top-1 accuracy: 30.91 | Helped |
| Use a small unrelated dataset as proxy instead of the best target-sampled proxy | NAO transfer to ImageNet1K and ImageNet22K · ImageNet1K and ImageNet22K | Top-1 accuracy | Helped |
| Select a random subset of classes while retaining images per class rather than representing all classes with fewer examples | Search on ImageNet22K proxies · ImageNet22K | Top-1 accuracy | Helped |
| Search on CIFAR100 rather than CIFAR10 | NAS methods transferred to ImageNet1K · ImageNet1K | relative improvement over random sampling: DARTS 5.93 -> -18.71; ENAS -16.26 -> -11.75 | Hurt |
| Use ImageNet22K Proxy 1 instead of Proxy 2 or Proxy 3 | DARTS, ENAS, and NAO search · ImageNet22K | performance relative to random sampling | Helped |
| Change augmentation seed from 2 to 3 (ablation) | ENAS transfer from CIFAR10 to ImageNet1K · ImageNet1K | Top-1 accuracy: 56.58 -> 72.69 | Mixed |
| Change augmentation seed from 2 to 3 (ablation) | DARTS transfer from CIFAR10 to ImageNet1K · ImageNet1K | Top-1 accuracy: 71.58 -> 72.01 | No clear effect |
| Increase cell count from 20 to 40 (ablation) | DARTS transfer from CIFAR10 to ImageNet1K · ImageNet1K | Top-1 accuracy: 71.58 -> 63.75 | Hurt |
| Increase cell count from 20 to 60 (ablation) | DARTS transfer from CIFAR10 to ImageNet1K · ImageNet1K | Top-1 accuracy: 71.58 -> 53.98 | Hurt |
| Replace BASE with DAC augmentation (ablation) | DARTS direct training on ImageNet1K · ImageNet1K | Top-1 accuracy: 72.03±0.2 -> 69.13±0.45 | Hurt |
| Replace BASE with DAC augmentation (ablation) | NAO direct training on ImageNet1K · ImageNet1K | Top-1 accuracy: 63.64±1.78 -> 61.67±2.11 | Hurt |
Closest papers by meaning
- NAS-Bench-101: Towards Reproducible Neural Architecture Search (ICML 2019 · similarity 0.92)
- TransNAS-Bench-101: Improving Transferability and Generalizability of Cross-Task Neural Architecture Search (CVPR 2021 · similarity 0.92)
- NAS-Bench-360: Benchmarking Neural Architecture Search on Diverse Tasks (NeurIPS 2022 · similarity 0.91)
- EcoNAS: Finding Proxies for Economical Neural Architecture Search (CVPR 2020 · similarity 0.91)
- Insights from the Use of Previously Unseen Neural Architecture Search Datasets (CVPR 2024 · similarity 0.91)
- NAS evaluation is frustratingly hard (ICLR 2020 · similarity 0.91)
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.