Towards Understanding Transfer Learning Algorithms Using Meta Transfer Features

Abstract

Transfer learning, which aims to reuse knowledge in different domains, has achieved great success in many scenarios via minimizing domain discrepancy and enhancing feature discriminability. However, there are seldom practical determination methods for measuring the transferability among domains. In this paper, we bring forward a novel meta-transfer feature method (MetaTrans) for this problem. MetaTrans is used to train a model to predict performance improvement ratio from historical transfer learning experiences, and can consider both the Transferability between tasks and the Discriminability emphasized on targets. We apply this method to both shallow and deep transfer learning algorithms, providing a detail explanation for the success of specific transfer learning algorithms. From experimental studies, we find that different transfer learning algorithms have varying dominant factor deciding their success, so we propose a multi-task learning framework which can learn both common and specific experience from historical transfer learning results. The empirical investigations reveal that the knowledge obtained from historical experience can facilitate future transfer learning tasks.