Research On Key Issue Of Generative Adversarial Domain Adaptation

In recent years,with the development of deep learning,it has been widely used in various fields.But in practice,deep learning still has many problems.First,it requires large-scale labeled samples for training to obtain a well-performing deep learning model.But the process of labeling samples is a waste of time and labor.Secondly,it's better to train the model using the data in the current scenario,but when there is a difference in distribution between the test data and the training data,the generalization ability of the deep learning model is significantly reduced,which leads to spending a lot of manpower and material resources to relabel the samples and retrain the model.Therefore,it has become a hot topic about how to use the knowledge learned in related fields to transfer and reuse the data of the current scene.In response to the large demand and the differences in distribution for training data,researchers have proposed the concept of domain adaptation.Domain adaptation is a branch of transfer learning,which aims to solve the transfer problem when the dataset distribution is different but the tasks are the same or similar.The generative adversarial domain adaptation method has achieved good results.Feature distribution of the target domain is aligned with that of the source domain through the approach of adversarial training.However,since the discriminator in the generative adversarial network is only a binary classifier which output is only 0 or 1,and the category feature information of the source domain and the target domain is not fully considered.It is easy to cause mismatches of category features in the process of overall feature alignment.As a result,negative transform occurs.Aiming at the key problem of category feature mismatching,our paper proposes a variational domain adaptive algorithm,which consists of a pre-training framework and a generative adversarial domain adaptation framework.First,the autoencoder-classifier structure is trained with labeled source domain data,and then the mean and variance of each category sample in the source domain are extracted through the Gaussian mixture model to generate the probability distribution of the source domain category features.Secondly,in the domain adaptation stage,on the one hand,the generative adversarial network is used to reduce the overall feature difference between the source domain and the target domain;on the other hand,the category alignment problem is converted into a problem that the posterior probability is close to the prior probability.The distribution of the source category feature is used as the prior probability of the target domain feature.In this way,the category information carried by the source domain is assigned to the generation process of target domain feature,and the target domain carries the category feature information and approximates the source domain to achieve category alignment.Finally,the source domain's classifier is used to classify the target domain samples.Through this method,the mismatch between the source domain and the target domain category is reduced,the negative migration problem in the domain adaptation process is alleviated,and the classification accuracy is improved.The algorithm proposed in this paper has an average accuracy rate of 97.8% on the digits data set,and an average accuracy rate of 94.5% on Office-31,which is higher than the mainstream algorithms.