Deep Transfer Learning via Restricted Boltzmann Machine for Document Classification

Abstract

Transfer learning aims to improve a targeted learning task using other related auxiliary learning tasks and data. Most current transfer-learning methods focus on scenarios where the auxiliary and the target learning tasks are very similar: either (some of) the auxiliary data can be directly used as training examples for the target task or the auxiliary and the target data share the same representation. However, in many cases the connection between the auxiliary and the target tasks can be remote. Only a few features derived from the auxiliary data may be helpful for the target learning. We call such scenario the deep transfer-learning scenario and we introduce a novel transfer-learning method for deep transfer. Our method uses restricted Boltzmann machine to discover a set of hierarchical features from the auxiliary data. We then select from these features a subset that are helpful for the target learning, using a selection criterion based on the concept of kernel-target alignment. Finally, the target data are augmented with the selected features before training. Our experiment results show that this transfer method is effective. It can improve classification accuracy by up to more than 10%, even when the connection between the auxiliary and the target tasks is not apparent.