Blending-target Domain Adaptation by Adversarial Meta-Adaptation Networks

(Unsupervised) Domain Adaptation (DA) seeks for classifying target instanceswhen solely provided with source labeled and target unlabeled examples fortraining. Learning domain-invariant features helps to achieve this goal,whereas it underpins unlabeled samples drawn from a single or multiple explicittarget domains (Multi-target DA). In this paper, we consider a more realistictransfer scenario: our target domain is comprised of multiple sub-targetsimplicitly blended with each other, so that learners could not identify whichsub-target each unlabeled sample belongs to. This Blending-target DomainAdaptation (BTDA) scenario commonly appears in practice and threatens thevalidities of most existing DA algorithms, due to the presence of domain gapsand categorical misalignments among these hidden sub-targets. To reap the transfer performance gains in this new scenario, we proposeAdversarial Meta-Adaptation Network (AMEAN). AMEAN entails two adversarialtransfer learning processes. The first is a conventional adversarial transferto bridge our source and mixed target domains. To circumvent the intra-targetcategory misalignment, the second process presents as learning to adapt'': Itdeploys an unsupervised meta-learner receiving target data and their ongoingfeature-learning feedbacks, to discover target clusters as ourmeta-sub-target'' domains. These meta-sub-targets auto-design ourmeta-sub-target DA loss, which empirically eliminates the implicit categorymismatching in our mixed target. We evaluate AMEAN and a variety of DAalgorithms in three benchmarks under the BTDA setup. Empirical results showthat BTDA is a quite challenging transfer setup for most existing DAalgorithms, yet AMEAN significantly outperforms these state-of-the-artbaselines and effectively restrains the negative transfer effects in BTDA.