基于最小包含球的领域迁移学习新方法

摘要

Traditional machine learning methods assume that different learning tasks have nothing with each other,but in fact there are some links between them.Transfer learning attempts to use these links and even past learning experiences between different tasks to accelerate the learning for new tasks.This paper integrated the MEB (Minimum enclosing ball algorithm together with Parzen windows probability estimation to develop a new transfer learning method named MEBTL (Minimum enclosing ball Transfer learning).We also used CVM (Core Vector Machines) theory to develop its fast version of the proposed algorithm CCMEBTL for large domain adaptation.The experimental results about "WIFI indoor positioning" and "face detection" indicate the effectiveness of the proposed algorithm.%传统机器学习方法认为不同的学习任务彼此无关,但事实上不同的学习任务常常相互关联.迁移学习试图利用任务之间的联系以及过去的学习经验加速对于新任务的学习.将最小包含球(Minimum Enclosing Ball,MEB)算法与Parzen Windows概率估计公式相结合,提出了一种新的迁移学习算法MEBTL((Minimum Enclosing Ball Transfer Learning).该算法同时结合CVM(Core Vector Machines)理论提出了CCMEBTL(Center Constrained Minimum Enclosing Ball Transfer Learning)算法,其可以在不同领域之间完成大样本的迁移学习.作为验证,将其应用在WIFI数据的室内定位、人脸识别检测上,并取得了较好的效果.