Gradient-Free Versus Gradient-Based Optimization in Sparse Neural Architecture Search: A Benchmarking Study Across Heterogeneous Hardware-Constrained Deployment Environments

Authors

  • Pat Walker Professor
  • Adrian White Associate Professor
  • Dana Lewis PhD
  • Drew Miller D.Sc

Keywords:

Neural Architecture Search, gradient-free optimization, differentiable NAS, hardware-aware inference, sparse model compression, Bayesian hyperparameter optimization, edge AI deployment, automated machine learning, latency-accuracy Pareto optimization

Abstract

Neural Architecture Search (NAS) has emerged as a foundational paradigm in automated machine learning, yet the trade-offs between gradient-based differentiable NAS (DNAS) and gradient-free evolutionary or Bayesian strategies remain insufficiently characterized under hardware-constrained deployment conditions. This study presents a systematic benchmarking analysis of six representative NAS algorithms—including DARTS, GDAS, SNAS, CMA-ES-NAS, BOHB, and REA—evaluated across three edge-inference hardware platforms and four sparse-model complexity regimes. Employing standardized search space normalization and latency-aware proxy metrics, we assess convergence efficiency, Pareto-front quality, and generalization fidelity on CIFAR-100, ImageNet-1K, and a domain-specific industrial defect classification corpus. Results indicate that gradient-free Bayesian approaches consistently achieve superior hardware-accuracy trade-offs under strict latency budgets below 15 ms, while DNAS retains advantages in high-parameter regimes. Practical deployment guidelines are derived from empirical cross-platform evidence.

Author Biographies

Pat Walker, Professor

Professor
Korea Advanced Institute of Science and Technology
291 Daehak-ro, Yuseong-gu, Daejeon 34141, Republic of Korea

Adrian White, Associate Professor

Associate Professor
Technical University of Munich
Arcisstraße 21, 80333 Munich, Bavaria, Germany

Dana Lewis, PhD

PhD
University of Waterloo
200 University Avenue West, Waterloo, Ontario N2L 3G1, Canada

Drew Miller, D.Sc

D.Sc
The University of Tokyo
7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8654, Japan

References

Kumar, N., & Kataria, V. Enhanced Sentiment Classification using a Multi-layered Stacked Ensemble Architecture.

Kumar, Nitin, and Vipin Kataria. "Enhanced Sentiment Classification using a Multi-layered Stacked Ensemble Architecture."

Kumar, N., & Kataria, V. (2023). Enhanced Sentiment Classification using a Multi-layered Stacked Ensemble Architecture. International Journal of Intelligent Systems and Applications in Engineering, 11(4s), 304–311.

Semeniuk, V. (2024). Аналіз та оптимізація продуктивності баз даних у розподілених системах. COMPUTER-INTEGRATED TECHNOLOGIES: EDUCATION, SCIENCE, PRODUCTION, (54), 199-205.

D. Satyanarayana, Sathyashree “A Three Level Architecture for Wireless Communication using LiFi”, in Recent Advances in Information and Communication Technology, Vol: 566, pp: 212- 221, Springer International Publishing, 2018 book. Print ISBN: 978-3-319-60662-0; Electronic ISBN: 978-3-319-60663-7; Book Series.

Satyanarayana, D. and Rao, S. V. (2011), “Fault-tolerant spanners for ad hoc networks” International Journal of Network Management, Vol: 22, Issue: 4, Pages: 311-329, July-2012, Wiley Publishers. https://doi.org/10.1002/nem.807

D. Satyanarayana, Sathyashree. S “Mobility Management in Voronoi based Cell Structure for Wireless networks”, IEEE proceedings for the International Conferences on Currents trends in Information Technology (CTIT), Pages: 204-208, Dubai, December-2013. DOI: 10.1109/CTIT.2013.6749504; INSPEC Accession Number: 14131032

D Satyanarayana, Perspective Chapter: Postquantum Security’, Coding Theory - Advances in Communications Engineering and Information Security. IntechOpen, Nov. 19, 2024. doi: 10.5772/intechopen.1007829.

D. Satyanarayana and J. M. H. Elmirghani, "An Energy Efficient Network Architecture for Infrastructured Wireless Networks," 2010 IEEE Global Telecommunications Conference GLOBECOM 2010, Miami, FL, USA, 2010, pp. 1-6, doi: 10.1109/GLOCOM.2010.5683234.

Alkalbani, A.S., Mantoro, T., Degala, S. (2016). Energy-Distance Aware Clustering Scheme (E-DACS) for Wireless Sensor Networks. In: Meesad, P., Boonkrong, S., Unger, H. (eds) Recent Advances in Information and Communication Technology 2016. Advances in Intelligent Systems and Computing, vol 463. Springer, Cham. https://doi.org/10.1007/978-3-319-40415-8_27

Rakhmanov, N. R., Kurbatsky, V. G., Guliyev, H. B., Tomin, N. V., & Rakhmanov, R. N. (2018). Analysis and modeling of harmonic distortions in consumer’s distribution networks, containing a powerful non-linear load. In E3S Web of Conferences (Vol. 58, p. 03015). EDP Sciences.

Hashimov, A. M., Rahmanov, N. R., Guliyev, H. B., Ibrahimov, F. S., & Tabatabaei, N. M. (2020). Probabilistic evaluatio n of voltage stability limit of power system under the conditions of accidental emergency outages of lines and generators. International Journal on Technical and Physical Problems of Engineering, 12(2), 40-45.

Guliyev, H. B. (2024). Management modes of reactive power compensation facilities in networks with renewable energy sources with distorting loads. IJTPE Journal International Journal on Technical and physical problems of engineering (IJTPE), 16(1), 14-20.

Published

2024-12-24

Issue

Section

Articles