Research On Unsupervised Domain Adaptation Based On Adversarial Networks

With the development of deep neural networks,supervised deep learning has been able to achieve good results in the field with a large number of labeled data.But in fact,there are only a large number of unlabeled data in many fields.Manual labeling of a large number of data is time-consuming and costly,and the quality of labeling is directly related to the effect of the model.So how to use a large number of unlabeled data to train model has become a key problem in the development of deep learning.And domain adaptation is an effective way to solve this problem.At present,the domain adaptation based on adversarial training has achieved good results.This kind of method uses the idea in training of generative adversarial networks,optimizes the domain classification loss through adversarial training,and aligns the feature distribution of source domain and target domain.By reducing the distribution differences between the two domain features,the model trained with source domain data can be applied to target domain data.The existing domain adaptation method is to train the model on the features after domain adaptation and does not make full use of the original information of the target domain data.When the data distribution of the two domains is quite different,the key information of the target domain data may be lost in the process of domain adaptation and reduce the intra-domain discriminability of the target domain.In view of the shortcomings of the existing methods,this paper makes improvements in optimizing the process of adversarial training and improve the discriminability of target domain features based on the adversarial domain adaptation,and improves the effect of the model through ensemble learning.(1)Improve the discriminability of target domain data features.The target domain data is labelled with pseudo-label by clustering and used to constrain the target domain feature encoder training through supervised learning in domain adaptation stage and improve the discriminability of target domain features by using the original distribution information of the target domain data.(2)Ensemble learning.Correct the pseudo-labeled data obtained by clustering and train classifiers according to the corrected samples.The classifier trained by clustering results and the classifier trained by domain adaptation will be used for ensemble learning,train the model with high confidence samples to improve the final effect of the model.The classifier trained by clustering results use raw data and will not cause the loss of features due to domain adaptation,and the two classifiers of clustering and domain adaptation have a better diversity,which is helpful to ensemble learning.Finally,we experiment on four standard data sets of the domain adaptation method and compared with several advanced domain adaptation methods.The results show that the improved method proposed in this paper has better experimental results than other methods.Through the visualization analysis of the results,it is proved that this paper improves both the effect of feature alignment and the discriminability of the target domain features.