Research Of Unsupervised Domain Adaptation Algorithm Based On Adversarial Learning

Unsupervised domain adaptation is a popular research branch of transfer learning.By alleviating the domain shift between labeled source domain and unlabeled target domain,unsupervised domain adaptation intends to boost the generalization performance of machine learning models on target domain.Existing unsupervised domain adaptation methods can be roughly divided into three categories,namely cross-domain distance measurement based unsupervised domain adaptation,adversarial learning based unsupervised domain adaptation,and selfsupervised learning based unsupervised domain adaptation.With excellent performance,adversarial learning based unsupervised domain adaptation have been favored in recent years.In this paper,two adversarial learning based unsupervised domain adaptation methods are proposed.These two methods aim to alleviate the mode collapse problem in the adversarial learning based unsupervised domain adaptation and the error accumulation problem in the self-supervised based unsupervised domain adaptation.So the generalization performance of models on target domain can be improved.This thesis firstly proposes an unsupervised domain adaptation algorithm based on cross-domain features’ distance consistency constraint.By introducing classifier’s categories discriminative information into the domain adaptation task,the proposed algorithm alleviates the mode collapse problem caused by the neglect of target domain’s categories discriminative information in the adversarial learning based unsupervised domain adaptation.Specifically,the algorithm takes classifier weight as the category prototypes in the feature space,and calculates the distance between the features and the categories prototypes to obtain the category distance vectors.Then,the consistency constraint adversarial training is performed on the category distance vectors to alleviate the mode collapse problem.The comparative experiments and analysis experiments on the commonly used public datasets verify the effectiveness of the algorithm.In order to better learn the discriminability structure of target domain,selfsupervised learning has been introduced into the domain adaptation task in recent years.However,due to the distribution discrepancy between domains,the pseudolabels of target domain instances assigned by self-supervised learning algorithm usually cannot accurately represent the true category information of these instances,resulting in the error accumulation problem.To address that,this thesis further proposes an unsupervised domain adaptation algorithm based on dynamic decision boundary and decoupling optimization.The algorithm estimates the semantic uncertainty of instances by dynamically shifting decision boundaries,and then adopts decoupling optimization strategy according to the degree of target domain instances’ semantic uncertainty.Furthermore,the participation of target domain instances with high semantic uncertainty in the process of cross-domain distribution alignment can be reduced,and the error accumulation problem can be alleviated.By comparing and analyzing the proposed algorithm with existing domain adaptation algorithm on the datasets commonly used in the domain adaptation task,the results show that the proposed algorithm has good performance.