Research Of Unsupervised Domain Adaptation Method Based Discrimination-aware

Machine learning is the core of artificial intelligence area,which aims to let the machine acquire knowledge from data independently,so as to train learning model and predict new unknown data.As an important branch of machine learning,domain adaptation focuses on the transfer of learned knowledge to new learning problems,which is a hot research direction.Domain adaptation usually aims to achieve good learning performance in the target domain by aligning the distribution among domains and minimizing the empirical risk of the source domain.However,it still has the following shortcomings: 1.Only considering the global alignment between domains and the transfer of discriminative knowledge of source domain classes,previous UDA methods ignore the discriminative knowledge of different categories across domains.2.Previous UDA methods ignore the semantic information contained in samples.For example,the knapsack features in the target domain may be mapped to the car features in the source domain,resulting in negative migration.Therefore,the main research content of this paper includes two aspects:First of all,Domain Adversarial Neural Network(DANN)is a transfer learning method which has attracted much attention in recent years.DANN uses a discriminator to distinguish the features of different domains,and the generator tries to confuse the domain features to make the distribution of domains aligned or consistent.Therefore,DANN encourages the global alignment of features between domains and does not consider the inter-class discrimination information between domains.Therefore,this paper proposes a Discrimination-Aware Domain Adversarial Neural Network(DA2NN)method,which introduces the differences of discriminative knowledge or instances cross domain classes into deep domain adaptation learning.DA2 NN uses multiple discriminators to realize similar knowledge transfer and heterogeneous discrimination perception between domains.Experiments show that DA2 NN has better transfer learning performance than DANN.Secondly,in domain adaptation learning,there are three kinds of important knowledge or information in data domain,i.e.,data semantic information,domain label and class label,and the existing domain adaptation methods usually use only one or two kinds of the above information.Different from the existing methods,this paper proposes an unsupervised domain adaptation model(SCAN)based on semantic information,which attempts to model data semantic information,domain tags and class tags in depth model to achieve domain adaptation.Firstly,In unsupervised domain adaptation depth model,clustering method is used to process the semantic information of samples,and then three alignment mechanisms are adopted for the proposed model,including sample semantic information alignment,domain alignment and class centroid alignment,so as to learn domain invariant features and semantics,and effectively reduce domain discrepancy.Experimental results show the good performance of the method.