Contemporary domain adaptation methods are very effective at aligning feature
distributions of source and target domains without any target supervision.
However, we show that these techniques perform poorly when even a few labeled
examples are available in the target. To address this s