Transfer Learning in Hierarchical Feature Spaces

Transfer learning provides an approach to solve target tasks more quickly and effectively by using previously acquired knowledge learned from source tasks. As one category of transfer learning approaches, feature-based transfer learning approaches aim to find a latent feature space shared between source and target domains. The issue is that the sole feature space can't exploit the relationship of source domain and target domain fully. To deal with this issue, this paper proposes a transfer learning method that uses deep learning to extract hierarchical feature spaces, so knowledge of source domain can be exploited and transferred in multiple feature spaces with different levels of abstraction. In the experiment, the effectiveness of transfer learning in multiple feature spaces is compared and this can help us find the optimal feature space for transfer learning.

I. Intruduction

Although machine learning technologies have made great achievements in many research areas, most of these technologies work under assumption that source domain and target domain have the same feature space and distribution. It means that if the feature space or/and distribution of the target data change, the prediction model trained using source data can't be used for the target tasks, so new model should be built using adequate labeled target data, which is time-consuming and sometimes unavailable. In real world situations, very few labeled target data can be obtained, and collecting new labeled data and constructing a new prediction model for target tasks is impossible. If knowledge exploited from similar but not identical source domain with plenty of labeled data can be utilized to target tasks, building a well prediction model for target task becomes possible.