《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (10): 3065-3075.DOI: 10.11772/j.issn.1001-9081.2025091152

• 人工智能 •    

可微神经架构搜索方法综述

王涵, 于瑞祺, 任爽()   

  1. 北京交通大学 计算机科学与技术学院,北京 100044
  • 收稿日期:2025-09-29 修回日期:2025-12-09 接受日期:2025-12-12 发布日期:2025-12-26 出版日期:2026-10-10
  • 通讯作者: 任爽
  • 作者简介:王涵(2001—),男,山西大同人,硕士研究生,主要研究方向:人工智能、量子计算
    于瑞祺(2002—),男,黑龙江鹤岗人,博士研究生,主要研究方向:量子计算、机器学习
  • 基金资助:
    国家自然科学基金资助项目(62072025)

Survey on differentiable neural architecture search methods

Han WANG, Ruiqi YU, Shuang REN()   

  1. School of Computer Science and Technology,Beijing Jiaotong University,Beijing 100044,China
  • Received:2025-09-29 Revised:2025-12-09 Accepted:2025-12-12 Online:2025-12-26 Published:2026-10-10
  • Contact: Shuang REN
  • About author:WANG Han, born in 2001, M. S. candidate. His research interests include artificial intelligence, quantum computing.
    YU Ruiqi, born in 2002, Ph. D. candidate. His research interests include quantum computing, machine learning.
  • Supported by:
    National Natural Science Foundation of China(62072025)

摘要:

神经架构搜索(NAS)通过自动化设计神经网络结构,突破传统手动设计的试错局限,显著降低了深度学习模型开发对专业经验的依赖,且大幅提升了模型表现。在众多方法中,可微神经架构搜索(DNAS)借助one-shot框架与梯度下降优化大幅缩短了搜索时间,已成为当前研究的主流方向。然而,现有研究仍缺乏对可微方法及其发展脉络的系统化梳理。为了填补相关空白并为相关研究提供参考,对近年来出现的DNAS方法进行归纳分析。首先,介绍DNAS算法的基本理论和基准数据集,并梳理DNAS的现存挑战;其次,对现有的DNAS算法进行层次化分析;最后,展望DNAS未来的发展方向。

关键词: 可微神经架构搜索, one-shot架构, 超网, 梯度下降, 架构崩溃, 硬件延迟

Abstract:

By automatically designing neural network structures, Neural Architecture Search (NAS) overcomes the trial-and-error limitations of traditional manual design, thereby substantially reducing the reliance of deep learning model development on professional experience, while enhancing the model performance significantly. Among various methods, Differentiable Neural Architecture Search (DNAS), which leverages the one-shot framework and gradient descent optimization, has drastically shortened the search time and has emerged as a main research direction in current research. However, a systematic review of differentiable methods and their developmental trajectories is still absent. To bridge this gap and provide a reference for related studies, recent DNAS methods were synthesized and analyzed. First, the fundamental theories and benchmark datasets of DNAS algorithms were introduced, followed by a survey of the existing challenges of DNAS. Then, a hierarchical analysis of current DNAS algorithms was conducted. Finally, the future research directions of DNAS were discussed.

Key words: Differentiable Neural Architecture Search (DNAS), one-shot architecture, supernet, gradient descent, architecture collapse, hardware latency

中图分类号: