CLDA: an adversarial unsupervised domain adaptation method with classifier-level adaptation

摘要

Domain adaptation is an active and important research field in transfer learning. Unsupervised domain adaptation, which is better in line with real-world scenarios than supervised and semi-supervised domain adaptation, has attracted much attention and research. Inspired by generative adversarial networks (GANs), adversarial unsupervised domain adaptation methods are proposed in recent years, which are shown to achieve state-of-the-art performance. Existing adversarial unsupervised domain adaptation methods generally adopt feature-level adaptation to reduce the cross-domain shifts, which is shown to have some limitations in related research. In this paper, we propose a classifier-level adaptation approach to further reducing the cross-domain shifts. The classifier-level adaptation uses two different but related classifiers for source domain and target domain, different from existing adversarial unsupervised domain adaptation methods. In addition, not only domain-invariant feature representations but also auxiliary information of class labels is used to exploit the joint distribution of category information and extracted features. Based on the above-mentioned approaches, a classifier-level domain adaptation (CLDA) method is proposed. Experimental results show that the proposed CLDA method outperforms state-of-the-art unsupervised domain adaptation methods on Digits and Office-31 datasets.