Research map / Architectures and efficiency / Neural architecture search
NAS evaluation is frustratingly hard
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 |
|---|---|---|---|
| Use advanced training tricks and longer training instead of the Base protocol (ablation) | DARTS search-space architectures · CIFAR10 | top-1 test accuracy: over 3 percentage points | Helped |
| Train sampled architectures with the best protocol rather than a weaker protocol (ablation) | DARTS search-space architectures · CIFAR10 | top-1 test accuracy: 98.15% | Helped |
| Replace DARTS operations with plain convolutions and max pooling (ablation) | DARTS search space · CIFAR10 | top-1 test accuracy: 0.18 accuracy points | Hurt |
| Change the training seed (ablation) | DARTS-sampled architectures · CIFAR10 | Kendall tau ranking correlation: 0.48 | Hurt |
| Change the training seed (ablation) | DARTS-sampled architectures · CIFAR10 | test accuracy change: 0.13% ± 0.08 (max change is 0.39%) | Mixed |
| Change the number of cells between training and evaluation (ablation) | DARTS-sampled architectures · CIFAR10 | Kendall tau ranking correlation: 0.54 | Hurt |
| Change the number of cells between training and evaluation (ablation) | DARTS-sampled architectures · CIFAR10 | architecture rank shift: up to 18 positions (out of 32) | Hurt |
| Sample different architectures within the DARTS search space (ablation) | DARTS search space · CIFAR10 | top-1 test accuracy: 97.03 ± 0.23; 96.18 to 97.56 | No clear effect |
| Use the searched architecture rather than the random-sampling baseline | NAS methods in the benchmark · CIFAR10, CIFAR100, SPORT8, MIT67, FLOWERS102 | relative improvement (%): 0.32; 0.23; -0.13; 0.10; 0.25 (DARTS) | Mixed |
Closest papers by meaning
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search (ICLR 2020 · similarity 0.95)
- Understanding Architectures Learnt by Cell-based Neural Architecture Search (ICLR 2020 · similarity 0.93)
- Evaluating The Search Phase of Neural Architecture Search (ICLR 2020 · similarity 0.93)
- Insights from the Use of Previously Unseen Neural Architecture Search Datasets (CVPR 2024 · similarity 0.93)
- einspace: Searching for Neural Architectures from Fundamental Operations (NeurIPS 2024 · similarity 0.92)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · 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.