Multiple-source domain adaptation with generative adversarial nets

Abstract

Current unsupervised domain adaptation (UDA) methods based on GAN (Generative Adversarial Network) architectures assume that source samples arise from a single distribution. These methods have shown compelling results by finding the transformation between source and target domains to reduce the distribution divergence. However, the one-to-one assumption renders the existing GAN-based UDA methods ineffective in a more realistic scenario that source samples are typically collected from diverse sources. In this paper, we present a novel GAN-enabled framework, which we call Multi-Source Adaptation Network (MSAN), for multiple-source domain adaptation (MDA) to mitigate the domain shifts between multiple source domains and the target domain. The proposed framework consists of multiple GAN architectures to learn bidirectional transformations between the source domains and the target domain efficiently and simultaneously. Technically, we introduce a joint feature space to guide the multi-level consistency constraints across all the transformations, in order to preserve the domain-invariant pattern and endow the discriminative power for the unlabeled target samples simultaneously during the adaptation. Moreover, the proposed model can naturally be used to enlarge the target dataset by utilizing the synthetic target images (with ground-truth labels from different source domains) and the pseudo-labeled target images, thereby allowing constructing the target-specific classifier in an unsupervised manner. Experiments demonstrate that our models exceed state-of-the-art results for MDA tasks on several benchmark datasets.