Research On Partial Adversarial Domain Adaptation Based On Pseudo-Margin

In recent years,with the dramatic development in the computing field,deep learning has been increasingly used,but its main weakness lies in the need for large amounts of labeled training data.Yet the ground truth is that obtaining enough labeled data is very challenging.To solve this critical problem,transfer learning has been widely studied and emerged.Domain Adaptation(DA)is one of the key areas of research in transfer learning.In this thesis,we focus on a more specific scenario of domain adaptation,Partial Domain Adaptation(PDA).2018 saw the introduction of the concept of partial domain adaptation,which specifically refers to a more realistic scenario where the label space of the target domain is true subspace of the source domain label space,where the target domain data is completely unlabeled,while the source domain has fully labeled data.However,the challenging point is that since the target domain does not contain labeling information,it is unknown which classes are common to the source and target domains and which classes are private to the source domain.For this reason,this thesis proposes a new approach to solve partial domain adaptation,Partial Adversarial Domain Adaptation Network based on Pseudo-margin.The essence of this method is to use the concept of pseudo-margin to measure the probability that the source class belongs to the common class of the two domains,and use this metric to weight the source class,while the method uses adversarial learning to adapt the distribution of the source and target domains,and quantifies the distribution distance between the common class data of the two domains by a similarity measure function,and minimizes this target loss.Compared with existing algorithms,this thesis improves on two main points.First,the pseudo margin is used on the framework of the adversarial network to weight the source domain class data,thus eliminating the possible negative migration of outlier source classes to the adaptation process.In contrast to existing partial domain adaptation networks,this thesis applies pseudo-margin to it and domain aligns the public class data in the source and target domains by introducing a pseudo-margin vector(PMV).The Target Pseudo-margin Register(TPMR)is used to continuously optimize the pseudo-margin vector to effectively avoid negative migration.In addition,the similarity metric is used to further reduce the distribution difference between the source and target domains common classes.Finally,inspired by the entropy minimization principle,the target entropy is added to the objective function in this thesis to achieve low-density separation.The method in this thesis is robust to outlier classes in the source domain because the method in this thesis applies a class-based weighting mechanism in the source domain.In this thesis,comparison experiments are conducted with other good domain adaptation and partial domain adaptation methods on a standard dataset,from which the advantages and disadvantages of different methods can be visualized.For example,on the Office-31 dataset,the accuracy of this thesis is up to 97.82%,exceeding the accuracy of the best comparison algorithm by 1.09%;on the Vis DA-2017 dataset,the accuracy of this thesis is up to 75.39%,exceeding the accuracy of the best comparison algorithm by 7.14%.In addition,additional complementary experiments are provided to analyze the adaptability,compactness,and robustness of the methods in this thesis.The experimental results fully demonstrate that the method in this thesis works well in various partial domain adaptation scenarios and outperforms other domain adaptation and partial domain adaptation methods by a large margin.In conclusion,the proposed partial domain adaptation method in this thesis can improve the classification accuracy of the target domain as well as the generalization ability of the classification model on standard datasets.Finally,this thesis summarizes this work and analyzes its shortcomings,and provides an outlook on future research directions.