Research map / Architectures and efficiency / Neural architecture search
PASHA: Efficient HPO and NAS with Progressive Resource Allocation
Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, Giovanni Zappella
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 |
|---|---|---|---|
| dynamically increase maximum resources only when successive rung rankings remain inconsistent | ASHA-based HPO/NAS · CIFAR-10 | Runtime: 3.0h ± 0.6h -> 1.3h ± 0.6h | Improved |
| dynamically increase maximum resources only when successive rung rankings remain inconsistent | ASHA-based HPO/NAS · CIFAR-100 | Accuracy (%): 71.69 ± 1.05 -> 71.84 ± 1.41 | No clear effect |
| dynamically increase maximum resources only when successive rung rankings remain inconsistent | ASHA-based HPO/NAS · ImageNet16-120 | Runtime: 8.8h ± 2.2h -> 2.9h ± 1.7h | Improved |
| dynamically increase maximum resources only when successive rung rankings remain inconsistent | ASHA-based HPO · WMT | Runtime: 43.7h ± 37.2h -> 2.8h ± 0.6h | Improved |
| dynamically increase maximum resources only when successive rung rankings remain inconsistent | ASHA-based HPO · ImageNet | Accuracy (%): 75.10 ± 2.03 -> 73.37 ± 2.71 | Hurt |
| combine progressive resource allocation with Bayesian optimization search | MOBSTER · CIFAR-100 | Runtime: 5.7h ± 1.4h -> 1.6h ± 0.5h | Improved |
| replace automatically estimated soft ranking with direct ranking (ablation) | PASHA · CIFAR-100 | Runtime: 0.9h ± 0.4h -> 2.8h ± 0.7h | Hurt |
| fix soft-ranking ε at 2.5% instead of estimating ranking noise (ablation) | PASHA · CIFAR-100 | Accuracy (%): 71.84 ± 1.41 -> 71.41 ± 1.15 | No clear effect |
| use soft-ranking ε=2σ instead of estimating ranking noise (ablation) | PASHA · CIFAR-100 | Runtime: 0.9h ± 0.4h -> 1.9h ± 0.7h | Hurt |
| replace soft ranking with Rank Biased Overlap (ablation) | PASHA · CIFAR-100 | Runtime: 0.9h ± 0.4h -> 2.4h ± 0.7h | Hurt |
| replace soft ranking with Reciprocal Rank Regret (ablation) | PASHA · CIFAR-100 | Runtime: 0.9h ± 0.4h -> 1.2h ± 0.5h | Improved |
Closest papers by meaning
- NASOA: Towards Faster Task-Oriented Online Fine-Tuning With a Zoo of Models (ICCV 2021 · similarity 0.87)
- HPO-B: A Large-Scale Reproducible Benchmark for Black-Box HPO based on OpenML (NeurIPS 2021 · similarity 0.86)
- Per-Architecture Training-Free Metric Optimization for Neural Architecture Search (NeurIPS 2025 · similarity 0.86)
- Speedy Performance Estimation for Neural Architecture Search (NeurIPS 2021 · similarity 0.86)
- Neural Architecture Generator Optimization (NeurIPS 2020 · similarity 0.86)
- Frugal Optimization for Cost-related Hyperparameters (AAAI 2021 · similarity 0.85)
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.
