Conditional Probability Distribution Divergence Reduction in Visual Domain Adaptation

The rapid evolution of data has challenged traditional machine learning methods and leads to the failure of many learning models. As a possible solution to the lack of sufficient labeled data, transfer learning aims to exploit the accumulated knowledge in the auxiliary domain to develop new predictive models. This article studies a specific type of transfer learning called domain adaptation, which works based on subspace learning in order to minimize the distance between class conditional probability distributions of source and target domains and to preserve source discriminative information. Efficient classifiers trained on source domain data have been used to predict target domain data labels to facilitate subspace learning. In this work, subspace learning is formulated as an optimization problem and experiments have been carried out on real-world datasets. The results of experiments indicate that the proposed method outperforms several existing methods in terms of accuracy on three datasets: Office-Caltech10, Office, and SS5 datasets.