Research Of Heterogeneous Transfer Learning Algorithm

Transfer learning is so far a hot research field in machine learning,which takes advantage of the related source domain knowledge to assist the learning of target domain.Due to the different feature spaces between source and target domains,heterogeneous transfer learning(HDA)accordingly has more challenges in transferring from source to target domain.How to align different feature spaces,and adaptively transfer relevant knowledge is crucial for HDA.There still exist some drawbacks in current methods of HDA: 1)The previous methods only focus on aligning feature spaces of source and target domains,and yet overlook the discriminative information of source and target domains.2)The direct transfer of source domain knowledge may lead to negative transfer.3)The previous methods aim at finding the common features for the source and target domains,and pay no attention to the unique features.As a result,the researches in this paper mainly include the following two aspects,First,considering that direct transfer of source domain knowledge can lead to negative transfer,a new adaptive method called Adaptive teacher-student Heterogeneous Domain Adaptive method(Ats HDA)is proposed.It treats source domain knowledge as a teacher to assist the learning of target domain,and target domain adaptively learns target classifier under the guidance of the teacher to avoid negative transfer.Experimental results show that,adaptive transfer of Ats HDA method achieves improvement in performance.Next,Domain Adversarial Separation Network(DASN)is proposed.In DASN,a semi-supervised domain adversarial network is utilized.Specifically,DASN not only uses new semi-supervised domain adversarial network,but also introduces a novel objective function of sample similarity,tojointly adjust the representation of sample to different target domain tasks.Experimental results show that performance improvement can be achieved by the DASN method.