No abstract

In general, deep neural networks trained on a given labeled dataset are expected to produce equivalent results when tested on a new unlabeled dataset. However, data are generally collected by different devices or under varying conditions and thus they often are not part of a same domain, yielding poor results. This is due to the domain shift between data distributions and has been the goal of a research area known as unsupervised domain adaptation. Many prior works have been designed to transfer knowledge between two domains: one source to one target. Since data may be taken from different sources and with different distributions, multi-source domain adaptation has received increasing attention. This paper presents the Multi-Source DomaIn Alignment Layers (MS-DIAL), which reduce the domain shift between multiple sources and a given target by embedding domain alignment layers in any given network. Except for the embedded layers, all the other network parameters are shared among all domains, saving processing time and memory usage. Experiments were performed on digit and object recognition tasks with five public datasets widely used to evaluate domain adaptation methods. Results show that the proposed method is promising and outperforms state-of-the-art approaches.