Transfer Learning Realized With Nonlinearly Transformed Input Space

This study is concerned with the design of a nonlinear transformation strategy to address the issues caused by the mismatch of marginal probability distributions between the source and target domains in transfer learning. In the process of transfer learning, the existing model (which is regarded as the available source of knowledge) is constructed in the source domain with the assumption that data points are independent and identically distributed. However, this assumption does not hold when the existing model is transferred to a new target domain. As a consequence, the performance of the existing model in the target domain deteriorates. Through mapping the original target space to a new space of the same dimensionality by using nonlinear transformations, the distributions of the data in the source and target domains could well matched each other, which significantly facilitates the transfer of accumulated knowledge to the new domain. No prior knowledge of the data distributions or the detailed information of the existing model is required. Experiments involving both synthetic dataset and real-world datasets are performed to demonstrate the effectiveness of the proposed approach in improving classification accuracy of the exiting model in the target domain.