A Hybrid Search Method for Accelerating Convolutional Neural Architecture Search

Research output: Chapters, Conference Papers, Creative and Literary WorksRGC 32 - Refereed conference paper (with host publication)peer-review

View graph of relations

Related Research Unit(s)

Detail(s)

Original languageEnglish
Title of host publicationICMLC 2023 - Proceedings of the 2023 15th International Conference on Machine Learning and Computing
PublisherAssociation for Computing Machinery
Pages177-182
ISBN (Print)9781450398411
Publication statusPublished - 2023

Publication series

NameACM International Conference Proceeding Series

Conference

Title15th International Conference on Machine Learning and Computing (ICMLC 2023)
LocationHybrid
PlaceChina
CityZhuhai
Period17 - 20 February 2023

Abstract

The performance evaluation of candidate architectures is a key step in evolution-based neural architecture search (ENAS). Generally, high-fidelity evaluation is desired for finding the optimal architecture but suffers from expensive computational cost while low-fidelity evaluation can reduce search cost but may influence the search performance. In this work, a hybrid search method, HybridNAS, is proposed by synergizing these two evaluation strategies to improve the search effectiveness of ENAS. This method consists of two stages: global exploration and local exploitation. The former applies low-fidelity evaluation to explore the promising architectures efficiently. As for the latter, it exploits the most promising ones identified by the former to generate the better-performancing architectures. Those two stages are iteratively repeated toward the optimal architecture. Extensive experiments on NAS-Bench-101 and NAS-Bench-201 reflect that HybridNAS can achieve comparable accuracy compared to existing popular NAS but at the expense of only 60%-80% computational cost. © 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM.

Research Area(s)

  • evaluation strategy, Neural architecture search, Hybrid search method

Citation Format(s)

A Hybrid Search Method for Accelerating Convolutional Neural Architecture Search. / Zhou, Xun; Liu, Songbai; Wong, KA-CHun et al.
ICMLC 2023 - Proceedings of the 2023 15th International Conference on Machine Learning and Computing. Association for Computing Machinery, 2023. p. 177-182 (ACM International Conference Proceeding Series).

Research output: Chapters, Conference Papers, Creative and Literary WorksRGC 32 - Refereed conference paper (with host publication)peer-review