基于参数字典的多源域自适应学习算法

Multi-source Domain Adaptation Learning Based on Parameter Dictionary

领域自适应是机器学习算法研究中一个热点,多源域自适应旨在利用多个源域的相关知识辅助目标域进行学习,现有多源领域自适应方法仅关注各源域和目标域间的知识迁移,很少考虑各源域间的相关性和共享信息。为此,提出一种基于参数字典的多源域自适应学习算法(DL_MSDA),通过学习各源域模型参数的公共字典,挖掘源域间的共享知识,并将其迁移至目标域,指导目标域模型参数的学习,完成知识从多个源域到目标域的迁移。模型可利用交替迭代(ADMM)方法进行求解。实验选取经典的多源迁移学习算法DAM进行对比,并在多个迁移学习图像数据集上进行了充分的验证。实验结果表明,DL_MSDA能够有效挖掘多个源域间共有的信息,辅助目标域模型参数的学习,提升目标域的分类性能。

Domain adaptation is a hot issue in the study of machine learning.Multi-source domain adaptation aims to help the target domain learning by utilizing information from multiple source domains,it has become an important research direction in machine learning.The existing methods on multi-source domain adaptation mainly concentrate on knowledge transfer from each source domain to the target one,while pays no attention to the relationship between source domains.As a result,we propose a novel multi-source domain adaption method based on parameter dictionary learning(DL_MSDA),which learns the intrinsic relationship between multiple source domains through the common parameter dictionary,and transfers such knowledge to the target domain to guide its learning.The ADMM methods can be used to solve the optimization problem of DL_MSDA.Finally,the DL_MSDA is compared with DAM on several commonly used datasets.Experiment shows that DL_MSDA can effectively explore the common knowledge between multiple source domains to boost the performance of target domain.