The $$\ell _{2,1}$$-Norm Stacked Robust Autoencoders via Adaptation Regularization for Domain Adaptation

Abstract

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. How to reconstruct more robust and high-level feature space to reduce the discrepancy between source and target domains is the crucial problem in domain adaptation. Recently, deep learning methods have shown promising results on learning new representations. However, in most of the previous work, the local geometry structure of data is not taken into account, which is benefit for constructing more robust feature space. Therefore, we propose a novel algorithm termed \(\ell _{2,1}\)-norm stacked robust autoencoders via adaptation regularization (SRAAR) to learn better feature representations for domain adaptation. More specifically, we incorporate an effective manifold regularization into the objective function to preserve the local geometry structure of data. In addition, label information is used to compute the maximum mean discrepancy (MMD) in order to reduce the distance between source and target domains. Experimental results show the effectiveness of proposed approach.