Robust Ensembling Network for Unsupervised Domain Adaptation

Recently, in order to address the unsupervised domain adaptation (UDA) problem, extensive studies have been proposed achieve transferrable models. Among them, most prevalent method is adversarial adaptation, which can shorten distance between source and target domain. Although learning very effective, it still leads instability of network drawbacks confusing category information. In this paper, we propose a Robust Ensembling Network (REN) for UDA, applies robust time ensembling teacher learn global information transfer. Specifically, REN mainly includes student network, performs standard training updates weights network. addition, also dual-network conditional loss improve ability discriminator. Finally, purpose improving basic utilize consistency constraint balance error Extensive experimental results on several UDA datasets demonstrated effectiveness our model by comparing with other state-of-the-art algorithms.