智能科学与技术学报 ›› 2024, Vol. 6 ›› Issue (2): 272-280.doi: 10.11959/j.issn.2096-6652.202417

• 学术论文 • 上一篇    

基于自适应平滑度策略的三维模型分类神经架构搜索

周鹏1, 杨军1,2()   

  1. 1.兰州交通大学电子与信息工程学院,甘肃 兰州 730070
    2.兰州交通大学测绘与地理信息学院,甘肃 兰州 730070
  • 收稿日期:2024-01-02 修回日期:2024-05-14 出版日期:2024-06-15 发布日期:2024-07-31
  • 通讯作者: 杨军 E-mail:yangj@mail.lzjtu.cn
  • 作者简介:周鹏(1982- ),男,兰州交通大学电子与信息工程学院博士生,主要研究方向为三维模型处理、神经架构搜索。
    杨军(1973- ),男,博士,兰州交通大学教授、博士生导师,主要研究方向为模式识别、三维模型的空间分析、遥感影像分析与处理等。
  • 基金资助:
    国家自然科学基金项目(42261067);甘肃省自然科学基金项目(22JR11RA157)

Neural architecture search for 3D model classification based on adaptive smoothness strategy

Peng ZHOU1, Jun YANG1,2()   

  1. 1.School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
    2.Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou 730070, China
  • Received:2024-01-02 Revised:2024-05-14 Online:2024-06-15 Published:2024-07-31
  • Contact: Jun YANG E-mail:yangj@mail.lzjtu.cn
  • Supported by:
    The National Natural Science Foundation of China(42261067);The Natural Science Foundation of Gansu Province(22JR11RA157)

摘要:

针对人工设计三维模型分类网络架构过度依赖专家经验且泛化能力较差的问题,提出了一种自适应平滑度策略的神经架构搜索方法。首先,使用改进候选操作选择策略和连续松弛化方法将离散的搜索空间连续化,并利用权重共享机制提高搜索效率。其次,在损失函数中添加自适应平滑度策略的正则化,由温度参数控制损失函数的平滑程度。最后,使用指数归一化方法计算损失函数,以避免损失值溢出。在三维点云数据集和蛋白质间相互作用数据集上的实验结果表明,在相同的训练样本和迭代次数下,自适应平滑度策略的神经架构搜索方法的分类准确率更高,性能更稳定。

关键词: 正则化, 神经架构搜索, 搜索空间, 点云, 分类

Abstract:

Aiming at the problem of poor generalization ability in hand-crafted architectures that overly rely on expert experience, a neural network architecture search method with an adaptive smoothness strategy was proposed. Firstly, an improved candidate operation selection strategy and a continuous relaxation method were used to convert discrete search space into continuous space, and a weight-sharing mechanism was employed to enhance search efficiency. Secondly, a regularization operation with an adaptive smoothness strategy was added to the loss function, whose smoothness degree was controlled by a temperature parameter. Finally, the loss function was calculated using an exponential normalization method to avoid loss value overflow. Experimental results on 3D point cloud datasets and protein-protein interaction datasets showed that the proposed method achieved higher classification accuracy and more stable performance under the same training samples and iterations.

Key words: regularization, neural architecture search, search space, point cloud, classification

中图分类号: 

No Suggested Reading articles found!