HPE-DARTS: Hybrid Pruning and Proxy Evaluation in Differentiable Architecture Search

Hung-I. Lin, Lin-Jing Kuo, Sheng-De Wang

2025

Abstract

Neural architecture search (NAS) has emerged as a powerful methodology for automating deep neural network design, yet its high computational cost limits practical applications. We introduce Hybrid Pruning and Proxy Evaluation in Differentiable Architecture Search (HPE-DARTS), integrating soft and hard pruning with a proxy evaluation strategy to enhance efficiency. A warm-up phase stabilizes network parameters, soft pruning via NetPerfProxy accelerates iteration, and hard pruning eliminates less valuable operations to refine the search space. Experiments demonstrate HPE-DARTS reduces search time and achieves competitive accuracy, addressing the reliance on costly validation. This scalable approach offers a practical solution for resource-constrained NAS applications.

Paper Citation

in Harvard Style

Lin H., Kuo L. and Wang S. (2025). HPE-DARTS: Hybrid Pruning and Proxy Evaluation in Differentiable Architecture Search. In Proceedings of the 17th International Conference on Agents and Artificial Intelligence - Volume 2: ICAART; ISBN 978-989-758-737-5, SciTePress, pages 252-263. DOI: 10.5220/0013148700003890

in Bibtex Style

@conference{icaart25,
author={Hung-I. Lin and Lin-Jing Kuo and Sheng-De Wang},
title={HPE-DARTS: Hybrid Pruning and Proxy Evaluation in Differentiable Architecture Search},
booktitle={Proceedings of the 17th International Conference on Agents and Artificial Intelligence - Volume 2: ICAART},
year={2025},
pages={252-263},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013148700003890},
isbn={978-989-758-737-5},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 17th International Conference on Agents and Artificial Intelligence - Volume 2: ICAART
TI - HPE-DARTS: Hybrid Pruning and Proxy Evaluation in Differentiable Architecture Search
SN - 978-989-758-737-5
AU - Lin H.
AU - Kuo L.
AU - Wang S.
PY - 2025
SP - 252
EP - 263
DO - 10.5220/0013148700003890
PB - SciTePress