Finding Robust Transfer Features for Unsupervised Domain Adaptation

Abstract

An insufficient number or lack of training samples is a bottleneck in traditional machine learning and object recognition. Recently, unsupervised domain adaptation has been proposed and then widely applied for cross-domain object recognition, which can utilize the labeled samples from a source domain to improve the classification performance in a target domain where no labeled sample is available. The two domains have the same feature and label spaces but different distributions. Most existing approaches aim to learn new representations of samples in source and target domains by reducing the distribution discrepancy between domains while maximizing the covariance of all samples. However, they ignore subspace discrimination, which is essential for classification. Recently, some approaches have incorporated discriminative information of source samples, but the learned space tends to be overfitted on these samples, because they do not consider the structure information of target samples. Therefore, we propose a feature reduction approach to learn robust transfer features for reducing the distribution discrepancy between domains and preserving discriminative information of the source domain and the local structure of the target domain. Experimental results on several well-known cross-domain datasets show that the proposed method outperforms state-of-the-art techniques in most cases.

Keywords

• unsupervised domain adaptation
• feature reduction
• generalized eigenvalue decomposition
• object recognition