본문 바로가기
  • Home

NA-Search: Differentiable Search of Normalization and Activation

  • Journal of The Korea Society of Computer and Information
  • Abbr : JKSCI
  • 2026, 31(1), pp.59~68
  • Publisher : The Korean Society Of Computer And Information
  • Research Area : Engineering > Computer Science
  • Received : November 25, 2025
  • Accepted : December 22, 2025
  • Published : January 30, 2026

Wangduk Seo 1

1경기대학교

Accredited

ABSTRACT

In this paper, we propose NA-Search, a novel differentiable neural architecture search framework that targets normalization-activation operations to achieve both efficiency and lightweight design in on-device environments. Conventional differentiable neural architecture search methods evaluate all candidate operations during search, resulting in computational and memory overhead that limits applicability on resource-constrained devices. To address this limitation, NA-Search reconstructs the search space around normalization-activation combinations and applies a -sampling that selects only  candidates, reducing the computational and memory cost of neural network evaluation. Experiments on CIFAR-10 dataset show that the proposed model achieves 89.75% accuracy with 0.124M parameters, outperforming widely used lightweight neural network in terms of accuracy-efficiency trade-off. Additional analyses show that adjusting the value of  during search contributes to improved stability and enhanced final performance, highlighting the effectiveness of the proposed sampling-based search design.

Citation status

* References for papers published after 2024 are currently being built.