Representation learning via serial autoencoders for domain adaptation

摘要

Domain adaption aims to promote the learning tasks in target domain by using the knowledge from source domain whose data distribution is different from target domain. The crucial problem in domain adaptation is learning more robust and higher-level feature representations to reduce the distribution divergences. Recently, deep learning methods based on autoencoders have been successfully applied in representation learning for domain adaptation. However, most existing methods of autoencoders rely on the single autoencoder model, which poses challenges for learning different characteristics of data. In this paper, we propose a new representation learning framework via serial autoencoders (SEAE), which extracts richer feature representations by serially connecting two different types of autoencoders. This framework consists of two encoding stages. In the first encoding stage, marginalized denoising autoencoder (MDAE) is applied to learn domain invariant features which occurred in both domains frequently. With the result of the first stage, stacked robust autoencoder via graph regularization (SRAEG) is used in the second encoding stage to improve the quality of feature representations. Additionally, SRAEG model can be computed in a closed form with less time cost. Experimental results demonstate the effectiveness of our proposed approach compared with several state-of-the-art baseline methods. (C) 2019 Elsevier B.V. All rights reserved.