Optimal transport for unsupervised domain adaptation based on deep learning

Abstract

This research has been done according to the concepts of machine learning, in unsupervised domain adaptation to predict the target domain labels and appropriate classification and learning in order to improve the performance. In order to obtain better domain adaptation results, the use of transfer matrix is suggested for mapping samples. In using the transfer matrix and mapping samples from the source domain to the target domain, a hidden space is used in the last layer of the CNN network to solve the transfer matrix as easily as possible. Also, for the convenience of calculations in larger data sets, the method of dividing data into smaller categories has been introduced. Following the introduction of these proposals and the type of research performance, the results of the proposed methods are more optimal than the previous methods. This model will give better results in new methods such as detecting traffic violations, the presence of diseases in people, the presence of goods in the commercial field, and in all areas that have very large datasets. Finally, in this research, with the help of CNN network, bigger problems have been investigated in order to provide a more optimal method for learning classification and feature space and prediction.

Keywords: Optimal transport, Domain adaptation, Deep learning, Joint distribution, Unsupervised lrarning