Abstract:
In this paper, we present a novel unsupervised domain adaptation framework, Multi-Partition Feature Alignment
Network, that learns a deep neural model for the target domain
without the need for any supervision. Recent leading approaches
for unsupervised domain adaptation are based on adversarial
alignment. Aligning the global distribution of the domain representations via adversarial training does not guarantee the classwise distribution alignment. The proposed approach is built
on adversarial learning with the focus on carefully aligning
class-wise domain representations. Our algorithm utilizes the
pseudo-labels (the predicted labels) of the target features to
stimulate class-wise alignment. As the pseudo-labels of individual
target features can be erroneous, instead of iteratively aligning
individual target samples, the proposed framework introduces
a generic class-specific multi-partition alignment procedure that
enables superior class-discriminative alignment of domain representations. The competitive performance of the proposed framework against state-of-the-art approaches over a wide variety of
visual recognition tasks, namely, the digits classification task
and the object recognition task, validates its effectiveness for
unsupervised domain adaptation.
