Unsupervised Domain Adaptation Based On Minimizing Maximum Mean Discrepancy

In recent years,deep learning has developed rapidly,and has achieved remarkable results in many fields,such as image recognition,face recognition,speech recognition,and natural language processing.But the great success of these applications relis on large scale labeled data.In fact,in the real wor-ld,there are many scenarios where large scale labeled data cannot be obtained,such as medical,geological and other fields.So usually choose a relevant field to get the labeled data,then train the model in the relevant domain(source domain),and then generalize to the target domain.However,when there is a discrepancy between the training data distribution(source domain)and the test data distribution(target domain),the generalization performance of the algorithm is greatly limited.For the problem of domain discrepancy between training sets and test sets,many previous methods based on maximum mean difference(MMD)learn domain invariant features through regular feature layers.Considering the finiteness of the transfer effect of the features learned by the regular feature layer,this paper proposes a novel domain adaptation method based on the MMD regular softmax prediction value.In this paper,we use the MMD regular softmax prediction value to obtain the maximum field alignment.At the same time,in order to avoid the degradation of the classification performance caused by the over-alignment of the softmax activation value,this paper also introduces the residual module,and the input of the residual module is the domain invariant feature learned by minimizing the MMD,and the output of the residual module is the final classification feature.The introduction of the residual module makes the domain invariant features and classification features independent and related to each other.The model ensures the same strong performance of classification labels while learning the strong domain invariant features.The method of this paper can be implemented in all deep networks with softmax classifiers.In order to compare with some recent deep domain adaptation methods,this paper implements the method of this paper on Alexnet,and makes an experimental evaluation on the two standard domain adaptation benchmark datasets of office31 and office-caltech,the method of this paper is the best overall performance,more than a large number of previous deep domain adaptation methods.