Ganin Yaroslav ; Lempitsky Victor

Abstract

Top-performing deep architectures are trained on massive amounts of labeleddata. In the absence of labeled data for a certain task, domain adaptationoften provides an attractive option given that labeled data of similar naturebut from a different domain (e.g. synthetic images) are available. Here, wepropose a new approach to domain adaptation in deep architectures that can betrained on large amount of labeled data from the source domain and large amountof unlabeled data from the target domain (no labeled target-domain data isnecessary). As the training progresses, the approach promotes the emergence of "deep"features that are (i) discriminative for the main learning task on the sourcedomain and (ii) invariant with respect to the shift between the domains. Weshow that this adaptation behaviour can be achieved in almost any feed-forwardmodel by augmenting it with few standard layers and a simple new gradientreversal layer. The resulting augmented architecture can be trained usingstandard backpropagation. Overall, the approach can be implemented with little effort using any of thedeep-learning packages. The method performs very well in a series of imageclassification experiments, achieving adaptation effect in the presence of bigdomain shifts and outperforming previous state-of-the-art on Office datasets.