When to transfer: a dynamic domain adaptation method for effective knowledge transfer

Transfer learning has achieved a lot of success recently in saving training samples. However, most of the existing methods only focus on what and how to transfer, but ignore when is the proper transfer time. In the study, we find that transfer useful knowledge at proper time is also significant for the performance. To address this issue, we propose a dynamic domain adaptation approach based on the particle swarm optimization evolutionary algorithm, which searches transfer opportunity automatically for different data domains and training stages. We evaluate the proposed method on various deep learning network structures, and find that the transfer coefficient has large variance in the first several training epochs, and becomes smaller later. This indicates that the features learned in the first several epochs are not stable and is not suitable for static transfer. In addition, the proposed method is not sensitive to the hyper-parameters generated, and it searches suitable transfer coefficients dynamically and automatically instead of conventional manual way. Extensive experiments conducted on various datasets and network structures demonstrate the superiority of the proposed method.