Transfer Learning in Non-stationary Environments

Abstract

The fields of transfer learning and learning in non-stationary environments are closely related. Both look into the problem of training and test data that come from different probability distributions. However, these two fields have evolved separately. Transfer learning enables knowledge to be transferred between different domains or tasks in order to improve predictive performance in a target domain and task. It has no notion of continuing time. Learning in non-stationary environments concerns with updating learning models over time in such a way to deal with changes that the underlying probability distribution of the problem may suffer. It assumes that training examples arrive in the form of data streams. Very little work has investigated the connections between these two fields. This chapter provides a discussion of such connections and explains two existing approaches that perform online transfer learning in non-stationary environments. A brief summary of the results achieved by these approaches in the literature is presented, highlighting the benefits of integrating these two fields. As the first work to provide a detailed discussion of the relationship between transfer learning and learning in non-stationary environments, this chapter opens up the path for future research in the emerging area of transfer learning in non-stationary environments.