Research On Domain Adaptation Based On Adversarial And Regularized Method

In the production and life of real society,with the rapid growth of data volume,the demand for analysis and utilization of images and other types of data is growing rapidly.In order to analyze various types of data tasks,we have made rapid progress in Deep Learning,which brings convenience to our life and greatly improves the quality of life.Although Deep Learning has been successful in many applications,its outstanding performance is largely dependent on a large number of labelled data,and the process of labelling data is generally time-consuming and labor-consuming.In order to reduce the high dependence on labelled data,one solution is to use the information of other data domain with existing labels to apply to the unlabelled data domain.Therefore,it is necessary to solve the data distribution mismatch between other data domain and this data domain.In the past,researchers have proposed a large number of domain adaptation methods and proved their effectiveness in solving the problem of data distribution mismatch between domains in experiments.However,the current domain adaptation method does not consider the stability of the prediction results of the model in the unlabelled target domain,so in the more difficult unsupervised domain adaptation scenario where the target domain is missing the label,it is often unable to have a good prediction effect in the target domain.In view of the above shortcomings,this paper proposes two novel domain adaptation methods based on adversarial and regularization methods,which are summarized as follows:(1)We propose a Weighted Temporal-ensembling Domain Adversarial Network(W-TeDAN).On one hand,it makes the model learn the invariant general features between domains and reduces the distance between domains.On the other hand,it regularizes the prediction results of the model in the unlabelled data domain and effectively uses the information of unlabelled data.In addition,this method uses the conditional distribution based on label information to represent the domain distribution.Compared with the marginal distribution only,the information of label and data will be considered simultaneously in the domain adversarial network,so that the purpose of domain adaptation is more refined and accurate.We also prove the effectiveness of this method through experiments.(2)In the process of domain adaptation,there may be some intrinsic implicit features that are easier to transfer between domains than other intrinsic implicit features.The domain adaptation network constructed by us is better able to learn this general feature and make it transfer between domains.For this purpose,we propose a Weighted Temporal-ensembling Domain Adversarial Network based on Meta Learning(W-TeDAN-meta).We try to introduce the idea of meta learning to make the whole model quickly learn some experience in each sub task,without over fitting the sub task and then synthesizing the experience learned by each sub task to make the target task learn better.In addition,model ensemble can make prediction results more stable.Experiments show that this method can further improve the prediction effect.