2019/07/07 by Chen Ziliang, Chen, Ziliang, Jingyu Zhuang +5
Computer Science · Medicine · #COVID-19 diagnosis using AI #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML)
paper · pdf · doi:10.48550/arxiv.1907.03389
openalex publication_date 2019/07/07 · openalex created_date 2021/08/30 · openalex updated_date 2026/07/28
(Unsupervised) Domain Adaptation (DA) seeks for classifying target instances\nwhen solely provided with source labeled and target unlabeled examples for\ntraining. Learning domain-invariant features helps to achieve this goal,\nwhereas it underpins unlabeled samples drawn from a single or multiple explicit\ntarget domains (Multi-target DA). In this paper, we consider a more realistic\ntransfer scenario: our target domain is comprised of multiple sub-targets\nimplicitly blended with each other, so that learners could not identify which\nsub-target each unlabeled sample belongs to. This Blending-target Domain\nAdaptation (BTDA) scenario commonly appears in practice and threatens the\nvalidities of most existing DA algorithms, due to the presence of domain gaps\nand categorical misalignments among these hidden sub-targets.\n To reap the transfer performance gains in this new scenario, we propose\nAdversarial Meta-Adaptation Network (AMEAN). AMEAN entails two adversarial\ntransfer learning processes. The first is a conventional adversarial transfer\nto bridge our source and mixed target domains. To circumvent the intra-target\ncategory misalignment, the second process presents as ``learning to adapt'': It\ndeploys an unsupervised meta-learner receiving target data and their ongoing\nfeature-learning feedbacks, to discover target clusters as our\n``meta-sub-target'' domains. These meta-sub-targets auto-design our\nmeta-sub-target DA loss, which empirically eliminates the implicit category\nmismatching in our mixed target. We evaluate AMEAN and a variety of DA\nalgorithms in three benchmarks under the BTDA setup. Empirical results show\nthat BTDA is a quite challenging transfer setup for most existing DA\nalgorithms, yet AMEAN significantly outperforms these state-of-the-art\nbaselines and effectively restrains the negative transfer effects in BTDA.\n