Unsupervised Domain Adaptation by Backpropagation

The paper tackles domain adaptation for deep architectures, motivated by the fact that top-performing models require massive labeled data that may be missing for a target task, while labeled data of similar nature from a different domain (such as synthetic images) is available. The proposed approach trains on a large amount of labeled data from the source domain and a large amount of unlabeled data from the target domain, with no labeled target-domain data needed. It works by promoting the emergence of deep features that are both discriminative for the main source task and invariant with respect to the shift between domains.

This adaptation behavior can be achieved in almost any feed-forward model by augmenting it with a few standard layers and a simple new gradient reversal layer, so the whole augmented architecture can be trained with standard backpropagation using common deep-learning packages. In a series of image classification experiments, the method achieves strong adaptation despite big domain shifts and outperforms the previous state of the art on the Office datasets, providing a simple and broadly applicable route to learning domain-invariant representations.