计算机科学与探索 ›› 2022, Vol. 16 ›› Issue (10): 2320-2329.DOI: 10.3778/j.issn.1673-9418.2101075

• 人工智能 • 上一篇    下一篇

快速多视角特权协同随机向量函数连接网络

吴天宇+(), 王士同   

  1. 江南大学 人工智能与计算机学院,江苏 无锡 214122
  • 收稿日期:2021-01-21 修回日期:2021-03-16 出版日期:2022-10-01 发布日期:2021-03-23
  • 通讯作者: + E-mail: 6191610014@jiangnan.edu.cn
  • 作者简介:吴天宇(1995—),男,甘肃酒泉人,硕士研究生,主要研究方向为人工智能、模式识别。
    王士同(1964—),男,江苏扬州人,教授,博士生导师,CCF会员,主要研究方向为人工智能、模式识别等。
  • 基金资助:
    国家自然科学基金(61972181)

Fast Multi-view Privileged Random Vector Function Link Network

WU Tianyu+(), WANG Shitong   

  1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Received:2021-01-21 Revised:2021-03-16 Online:2022-10-01 Published:2021-03-23
  • About author:WU Tianyu, born in 1995, M.S. candidate. His research interests include artificial intelligence and pattern recognition.
    WANG Shitong, born in 1964, professor, Ph.D. supervisor, member of CCF. His research interests include artificial intelligence, pattern recognition, etc.
  • Supported by:
    National Natural Science Foundation of China(61972181)

摘要:

现实情况中通常会针对同一对象从不同途径或层面获得特征数据,称这样获得的数据为多视角数据。对于多视角数据的挖掘利用具有研究价值,与传统的单视角学习相比表现出一定优势。多视角学习(MVL)中一个重要的问题是如何在满足视角间互补情况下同时保持视角之间一致性。为解决上述问题,基于多视角学习和特权信息学习(LUPI)概念,以随机向量函数连接网络(RVFL)为基础,提出了一种快速多视角特权协同随机向量函数连接网络(FMPRVFL)来有效地解决多视角分类任务。该方法的基本思想是在平均情况下相互利用冗余视角的附加信息作为特权信息监督当前视角的分类。在此基础上设计的FMPRVFL的目标函数可以利用解析解对目标函数进行优化,从而使FMPRVFL训练速度更快。理论分析表明,与经典的多视角学习方法相比,FMPRVFL可以提供额外的泛化能力。在64个数据集上的实验结果表明,FMPRVFL在平均测试精度和运行时间上都优于比较方法。

关键词: 多视角学习(MVL), 特权信息, 随机向量函数连接网络(RVFL)

Abstract:

In reality, feature data are usually obtained from different ways or levels for the same object, and the data obtained are called multi-view data. It is of research value to mine and utilize multi-view data and shows some advan-tages over traditional single-view learning. An important issue in multi-view learning (MVL) is how to meet the consistency between perspectives while maintaining complementarities between perspectives. In order to solve the above problem, based on the concepts including multi-view leaning and learning using privileged information (LUPI), a fast multi-view privileged RVFL (FMPRVFL) is proposed based on random vector function link network (RVFL) to accomplish multi-view classification tasks. The basic idea of FMPRVFL lies in the use of additional information from other views on average as the privileged information to supervise the classification in the current view. The objective function of FMPRVFL designed in such a mutually supervised manner can be optimized with analytical solutions, thus accelerating the training of FMPRVFL. It is revealed in the theoretical analysis that in contrast to classical multi-view learning, FMPRVFL can provide extra generalization capability. The results of an experiment on 64 datasets demonstrate that FMPRVFL outperforms other comparative methods both in average testing accuracy and running time.

Key words: multi-view learning (MVL), privileged information, random vector function link network (RVFL)

中图分类号: