文献类型：会议论文

英文题名：Gaussian Process for Transfer Learning through Minimum Encoding

作者：Shao, Hao[1];Xu, Rui[2];Tao, Feng[3]

机构：[1]Shanghai Univ Int Business & Econ, Sch WTO Res & Educ, Shanghai, Peoples R China;[2]Chinese Univ Sci & Technol, Sch Comp Sci & Technol, Hefei, Peoples R China;[3]E China Univ Sci & Technol, Sch Business, Shanghai 200237, Peoples R China

会议论文集：14th International Conference on Intelligent Data Engineering and Automated Learning (IDEAL)

会议日期：OCT 20-23, 2013

会议地点：Hefei, PEOPLES R CHINA

语种：英文

外文关键词：Transfer Learning; Gaussian Process; MDLP

摘要：In real applications, labeled instances are often deficient which makes the classification problem on the target task difficult. To solve this problem, transfer learning techniques are introduced to make use of existing knowledge from the source data sets to the target data set. However, due to the discrepancy of distributions between tasks, directly transferring knowledge will possibly lead to degenerated performance which is also called negative trasnfer. In this paper, we adopted the Gaussian process to alleviate this problem by directly evaluating the distribution differences, with the parameter-free Minimum Description Length Principle (MDLP) for encoding. The proposed method inherits the good property of solid theoretical foundation as well as noise-tolerance. Extensive experiments results show the effectiveness of our method.