A Novel Weight Update Rule of Online Transfer Learning
Xingda Wang, Xiaoping Wang, Zhigang Zeng
Abstract
Transfer learning aims to enhance performance in a target domain by exploiting useful information from related source domains. Transfer learning is important for many applications where the target domain instances are difficult to obtain while many data are available in source domains. Online transfer learning is established as an effective technology in many applications where the target data is received in an online manner. However, most existing methods ignore that target domain model does not have any prior knowledge when combined with source domain models. It is inappropriate that using conventional weight update rules to combine target domain model with source domain models. In this paper, we put forward a novel weight update rule for online transfer learning. Specifically, we prove the shortcomings of the previous methods and propose an effective improvement method which does not require a large amount of computation. Extensive experiments verify that our method can outperform several state-of-the-art methods on real-world data sets.