Domain adaptation, and transfer learning more generally, seeks to remedy the problem created when training and testing datasets are generated by different distributions. In this work, we introduce a new unsupervised domain adaptation algorithm for when there are multiple sources available to a learner. Our technique assigns a rough labeling on the target samples, then uses it to learn a transformation that aligns the two datasets before final classification. In this article we give a convenient implementation of our method, show several experiments using it, and compare it to other methods commonly used in the field.
View on arXiv