电子学报 ›› 2020, Vol. 48 ›› Issue (2): 359-368.DOI: 10.3969/j.issn.0372-2112.2020.02.019

• 学术论文 • 上一篇    下一篇

基于张量表示的域适配的迁移学习中特征表示方法

赵鹏, 王美玉, 纪霞, 刘慧婷   

  1. 1. 安徽大学计算机智能与信号处理教育部重点实验室, 安徽合肥 230601;
    2. 安徽大学计算机科学与技术学院, 安徽合肥 230601
  • 收稿日期:2019-05-22 修回日期:2019-09-12 出版日期:2020-02-25 发布日期:2020-02-25
  • 作者简介:赵鹏 女,1976年生于安徽合肥.博士,副教授,硕士生导师,CCF会员,主要研究方向为机器学习、图像理解.E-mail:zhaopeng_ad@163.com;王美玉 女,1994年生于安徽黄山.硕士研究生,主要研究方向为迁移学习、图像标注;纪霞 女,1983年生于安徽宣城.博士,讲师,硕士生导师,CCF会员,主要研究方向为智能信息处理;刘慧婷 女,1978年生于安徽阜阳.博士,副教授,硕士生导师,主要研究方向为机器学习、智能推荐.
  • 基金资助:
    国家自然科学基金(No.61602004);安徽省高校自然科学研究重点项目(No.KJ2018A0013,No.KJ2017A011);安徽省自然科学基金(No.1908085MF188,No.1908085MF182);安徽省重点研究与开发计划项目(No.1804d08020309)

A Novel Feature Representation Based on Tensor and Domain Adaption for Transfer Learning

ZHAO Peng, WANG Mei-yu, JI Xia, LIU Hui-ting   

  1. 1. Key Laboratory of Intelligent Computing and Signal Processing of Ministry of Education, Anhui University, Hefei, Anhui 230601;
    2. School of Computer Science and Technology, Anhui University, Hefei, Anhui 230601, China
  • Received:2019-05-22 Revised:2019-09-12 Online:2020-02-25 Published:2020-02-25

摘要: 本文提出一种新的基于张量表示的域适配迁移学习中的特征表示方法,即融合联合域对齐和适配正则化的基于张量表示的迁移学习特征表示方法.当源域和目标域差异很大时,仅将源域对齐潜在共享空间,会造成数据扭曲过大.为缓解此问题,本文方法提出联合域对齐,即源域和目标域同时对齐共享子空间.并且本文方法将适配正则化引入张量表示空间求解.本文适配正则化包括动态分布对齐和图适配,以缩小域间分布差异和保留样本间流行一致性.最后融合联合域对齐,动态分布对齐和图适配,通过联合优化求解获得共享子空间表示.几个公共的跨域数据集上的大量实验结果表明了本文方法优于其它主流的迁移学习方法,验证了本文方法的有效性.

关键词: 张量表示, 迁移学习, 域适配

Abstract: A novel feature representation based on tensor and domain adaption for transfer learning is proposed, which combines joint domain alignment and adaptation regularization.When the difference between the source domain and the target domain is very large, only aligning the source domain to the potential shared subspace will result in too much data distortion.To alleviate this problem, this paper proposes joint domain alignment, which aligns the source domain and the target domain to the potential shared subspaces simultaneously.Furthermore, the adaption regularization is introduced into the subspace learning based on tensor.In the proposed method, adaptation regularization includes dynamic distribution alignment and graph adaptation to reduce the distribution differences among different domains and preserve the manifold consistency.Finally, the joint domain alignment, dynamic distributed alignment and graph adaptation are fused, and the joint optimization is utilized to obtain the feature representation.Extensive experiments on several common cross-domain datasets show that the proposed method outperforms the state-of-the-art on the tasks of transfer learning and the effectiveness of the proposed method is verified.

Key words: tensor representation, transfer learning, domain adaption

中图分类号: