特征迁移中投影子空间的分析与构造

Subspace projection techniques in feature transfer learning

中文关键词  迁移学习  子空间投影  线性判别分析  最大均值差异

英文关键词  transfer learning  subspace projection  linear discriminant analysis  maximum mean discrepancy

中文摘要

      近年来, 针对实际应用场景中可匹配的训练数据不足的问题, 科研人员发展出了迁移学习的概念, 希望通过提取源域数据的特征信息进行迁移, 从而提升目标域的学习效果. 本文根据迁移学习所处理的不同数据类型, 构造了两种典型的模型：单类别投影基构造模型与监督多类别投影模型. 由于子空间投影可以在一定程度上反映原始样本空间的特征性质. 因此, 我们应用线性判别分析的技巧以及最大均值差异的思想，分别构造了上述模型的求解算法并对相应的非线性核方法进行了推广.

英文摘要

      In recent years, researchers have developed the concept of transfer learning for lack of effective training data in practice, hoping to improve the performance in target domain by learning the feature information in source domain. According to different types of data in transfer learning, we build two typical models in this paper, which is referred to as one-class projective base construction and supervised multi-class projection. Since the essential properties of the raw data space can be characterized by taking projection in a proper subspace, we study the above models using techniques from linear discriminant analysis (LDA) and maximum mean discrepancy (MMD). Also, we extend those methods to the corresponding nonlinear kernel cases.