Research map / Architectures and efficiency / Neural architecture search
Speedy Performance Estimation for 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 |
|---|---|---|---|
| Replace inherited-weight validation accuracy with TSE computed over additional minibatches | RandNAS, FairNAS, and MultiPaths · NB201-CIFAR10 and DARTS | rank correlation: 0.44 (0.15) -> 0.70 (0.02); 0.56 (0.17) -> 0.84 (0.01); 0.67 (0.05) -> 0.83 (0.01); 0.11(0.04) -> 0.30(0.04) | Helped |
| Replace inherited-weight validation accuracy with TSE computed over additional minibatches | RandNAS, FairNAS, and MultiPaths · NB201-CIFAR10 and DARTS | top-10 average test accuracy: 91.47 (0.31) -> 92.67 (0.12; 91.73 (0.21) -> 92.7 (0.1); 91.77 (0.78) -> 92.63 (0.12); 93.20(0.04) -> 93.64(0.04) | Helped |
| Use TSE-EMA instead of early-stopping validation accuracy | Query-based NAS (RE and BO) · NB201 on CIFAR10, CIFAR100, and ImageNet-16-120 | search speed to find high-performing architectures | Helped |
| Replace validation-loss gradients with TSE gradients | DARTS and DrNAS · NB201-CIFAR10 and DARTS-NB301 | test accuracy during search | Helped |
| Use training-loss sums rather than validation-loss sums (ablation) | TSE versus SoVL · NB201 on CIFAR10, CIFAR100, and ImageNet-16-120 | rank correlation with final test accuracy | Helped |
| Use the sum of training losses instead of a single minibatch training loss (ablation) | TSE versus TLmini · DARTS on CIFAR10 under varied NAS set-ups | rank correlation | Helped |
| Restrict TSE-E to the most recent epoch (E=1) (ablation) | TSE-E · NB201 and DARTS datasets | rank correlation | Helped |
| Vary the TSE-EMA decay rate among 0.9, 0.95, 0.99, and 0.999 (ablation) | TSE-EMA · NB201 and DARTS datasets | rank correlation | No clear effect |
| Weight recent training losses more heavily while retaining earlier epochs (ablation) | TSE-EMA versus TSE-E · Various NAS search spaces, especially at small training budgets | rank correlation | Helped |
| Use TSE-E or TSE-EMA after architectures begin to overfit | TSE-E and TSE-EMA · NB201-CIFAR10 and NB201-CIFAR100 at later training epochs | rank correlation | Hurt |
Closest papers by meaning
- Rethinking Performance Estimation in Neural Architecture Search (CVPR 2020 · similarity 0.92)
- Evaluating Efficient Performance Estimators of Neural Architectures (NeurIPS 2021 · similarity 0.92)
- Neural Architecture Generator Optimization (NeurIPS 2020 · similarity 0.92)
- Contrastive Neural Architecture Search With Neural Architecture Comparators (CVPR 2021 · similarity 0.92)
- Robustifying and Boosting Training-Free Neural Architecture Search (ICLR 2024 · similarity 0.91)
- MOTE-NAS: Multi-Objective Training-based Estimate for Efficient Neural Architecture Search (NeurIPS 2024 · similarity 0.90)
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.