计算机工程 ›› 2020, Vol. 46 ›› Issue (12): 105-112,133.doi: 10.19678/j.issn.1000-3428.0058766

• 人工智能与模式识别 • 上一篇    下一篇

考虑多样性的深度神经网络结构搜索方法研究

杨晓梅, 郭文强, 张菊玲   

  1. 新疆财经大学 信息管理学院, 乌鲁木齐 830012
  • 收稿日期:2020-06-28 修回日期:2020-08-03 发布日期:2020-08-12
  • 作者简介:杨晓梅(1979-),女,副教授、硕士,主研方向为人工智能、大数据分析;郭文强,教授、博士;张菊玲,副教授、博士。
  • 基金项目:
    新疆维吾尔自治区自然科学基金(2019D01A27);新疆维吾尔自治区社会科学基金(19BXW085);教育部人文社会科学研究规划基金项目(17XJJAZH001)。

Research on Structure Search Method for Deep Neural Network Considering Diversity

YANG Xiaomei, GUO Wenqiang, ZHANG Juling   

  1. College of Information Management, Xinjiang University of Finance and Economics, Urumqi 830012, China
  • Received:2020-06-28 Revised:2020-08-03 Published:2020-08-12

摘要: 为提升网络结构的寻优能力,提出一种改进的深度神经网络结构搜索方法。针对网络结构间距难以度量的问题,结合神经网络的结构搜索方案,设计基于图的深度神经网络结构间距度量方式。对少量步数训练和充分训练2种情况下的网络结构性能进行分析,基于多样性解的优势,给出一种多样性最优网络结构搜索方法。实验结果表明,该方法能够有效提高解的质量,有助于寻找到更优的网络结构。

关键词: 深度神经网络, 结构搜索, 多样性, 网络对准, 节点相似性

Abstract: In order to improve the ability of network structure optimization,this paper proposes an improved deep neural network structure search method.Aiming at the problem that it is difficult to measure the distance between network structures,a graph-based method is proposed to measure the distance between deep neural network structures based on the structure search scheme of neural network.By analyzing the performance of network structure that is respectively trained by a small number of steps and fully trained,a method to find the diverse optimal network structure is proposed based on the advantages of diversity solutions.Experimental results show that the neural network structure search method can effectively improve the solution quality and help in finding a better network structure.

Key words: deep neural network, structure search, diversity, network alignment, node similarity

中图分类号: