When correspondences between domains are not available, we resort to purely discriminative information, assuming that there is few labeled data available in the target domain.
In contrast to other approaches in Domain Adaptation and Transfer Learning, we assume that there exists a latent space Z where a shared decision bounday exists. The form of the mapping from the original feature space to Z is key to our approach, and we show that the proposed solution can compensate for the discrepancies between domains, outperforming the baselines.
For more information please check the publications below.