Deep Domain Adaptation Based on Adversarial Network With Graph Regularization

Abstract

Although most transfer learning methods can reduce the difference of the feature distributions between the source and target domains effectively, some classes in the two domains may still be misaligned after domain adaptation, especially for the classes with similar features such as "bicycle" and "motorcycle". Therefore, a graph regularization based adversarial network model is proposed, whose innovations mainly include the following two aspects: First, a constraint function which is used to measure the difference between the features belonged to different classes is proposed, whose purpose is that not only the training accuracy is taken into account during supervised training, but also the difference between classes should be enlarged as much as possible; Then, a graph regularization constraint function is proposed, which makes all the classes have good local preserving properties after domain adaptation, and further reduces the possibility of all classes being misaligned. Experimental results on several cross-domain benchmark datasets show that our newly proposed approach outperforms state of the art methods.