Unsupervised domain adaptation via Bregman Divergence minimization and Adaptive Classifier learning

In pattern recognition and image classification, the common assumption that the training set (source domain) and test set (target domain) share the same distribution is often violated in real-world applications. In this case, traditional learning models may not generalize well on test sets. To tackle this problem, domain adaptation try to exploit training data with same distribution from other related source domain to generalize model for target domain.This paper presents a domain adaptation method which learns to adapt the data distribution of the source domain to that of the target domain where no labeled data of the target domain is available. Our method jointly learns a low dimensional representation space and an adaptive classifier. In fact, we try to find a representation space and an adaptive classifier on this representation space such that the distribution gap between the two domains is minimized and the risk of the adaptive classifier is also minimized.In this paper, we propose a novel solution to tackle unsupervised domain adaptation for classification. In the unsupervised scenario where no labeled samples from the target domain is available, our model transforms data such that the source and target distributions become similar. To compare two distributions, our approaches make use of Bregman divergence. However, this does not suffice to generalize the model. Here, we propose to make better use model matching along with representation learning to tackle distribution mismatch across domains. The framework extends classification model by adding an adaptive classifier, which generalizes the target classifier far from the source data. Then this framework guarantees the target classifier minimizes the empirical risk in target domain and maximize manifold consistency with source data structure. Our empirical study on multiple open data sets validates that our proposed approach can consistently improve the classification accuracy compared to the basic machine learning and state-of-the-art transfer learning methods.