Domain Adaptive Ensemble Learning

The problem of generalizing deep neural networks from multiple source domains to a target one is studied under two settings: When unlabeled data available, it multi-source unsupervised domain adaptation (UDA) problem, otherwise generalization (DG) problem. We propose unified framework termed adaptive ensemble learning (DAEL) address both problems. A DAEL model composed CNN feature extractor shared across and classifier heads each trained specialize in particular domain. Each such an expert its own non-expert others. aims learn these experts collaboratively so that when forming ensemble, they can leverage complementary information other be more effective for unseen To this end, used turn as pseudo-target-domain with providing supervisory signal the non-experts learned sources. For UDA setting where real does not exist, uses pseudo-label supervise learning. Extensive experiments on three datasets DG show improves state art problems, often by significant margins. code released at \url{https://github.com/KaiyangZhou/Dassl.pytorch}.