杭文龙2,,,
冯伟2,
刘学军2
1.南京邮电大学地理与生物信息学院 ??南京 ??210023
2.南京工业大学计算机科学与技术学院 南京 211816
基金项目:国家自然科学基金(61802177),江苏省高校自然科学研究面上项目(18KJB520020),南京邮电大学引进人才科研启动基金(NY219034),江苏省重点研发计划(BE2015697)
详细信息
作者简介:梁爽:女,1987年生,讲师,研究方向为机器学习、信号处理
杭文龙:男,1988年生,讲师,研究方向为机器学习、模式识别
冯伟:男,1995年生,硕士生,研究方向机器学习、模式识别
刘学军:男,1970年生,教授,硕士生导师,研究方向为数据挖掘、大数据分布式处理
通讯作者:杭文龙 wlhang@njtech.edu.cn
中图分类号:TP181计量
文章访问数:1705
HTML全文浏览量:987
PDF下载量:36
被引次数:0
出版历程
收稿日期:2018-11-20
修回日期:2019-04-30
网络出版日期:2019-05-16
刊出日期:2019-11-01
Adaptive Knowledge Transfer Based on Classification-error Consensus Regularization
Shuang LIANG1,Wenlong HANG2,,,
Wei FENG2,
Xuejun LIU2
1. School of Geographic and Biologic Information, Nanjing University of Posts and Telecommunications, Nanjing 210023, China
2. Department of Computer and Technology, Nanjing Tech University, Nanjing 211816, China
Funds:The National Nature Science Foundation of China (61802177), The Natural Science Foundation of the Higher Education Institutions of Jiangsu Province (18KJB520020), NUPTSF (NY219034), Key Research and Development Program of Jiangsu Province (BE2015697)
摘要
摘要:目前大多数迁移学习方法在利用源域数据辅助目标域数据建模时,通常假设源域中的数据均与目标域数据相关。然而在实际应用中,源域中的数据并非都与目标域数据的相关程度一致,若基于上述假设往往会导致负迁移效应。为此,该文首先提出分类误差一致性准则(CCR),对源域与目标域分类误差的概率分布积分平方误差进行最小化度量。此外,该文提出一种基于CCR的自适应知识迁移学习方法(CATL),该方法可以快速地从源域中自动确定出与目标域相关的数据及其权重,以辅助目标域模型的构建,使其能在提高知识迁移效率的同时缓解负迁移学习效应。在真实图像以及文本数据集上的实验结果验证了CATL方法的优势。
关键词:迁移学习/
负迁移/
概率分布/
分类误差一致性规则
Abstract:Most current transfer learning methods are modeled by utilizing the source data with the assumption that all data in the source domain are equally related to the target domain. In many practical applications, however, this assumption may induce negative learning effect when it becomes invalid. To tackle this issue, by minimizing the integrated squared error of the probability distribution of the source and target domain classification errors, the Classification-error Consensus Regularization (CCR) is proposed. Furthermore, CCR-based Adaptive knowledge Transfer Learning (CATL) method is developed to quickly determine the correlative source data and the corresponding weights. The proposed method can alleviate the negative transfer learning effect while improving the efficiency of knowledge transfer. The experimental results on the real image and text datasets validate the advantages of the CATL method.
Key words:Transfer learning/
Negative transfer/
Probability distribution/
Classification-error Consensus Regularization (CCR)
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=cdac6fb1-83a7-49eb-9f13-f822b48a7122