Blending-Target Domain Adaptation by Adversarial Meta-Adaptation Networks

Abstract

Figure 1. The comparison of MTDA and BTDA (color orange and blue denote source and target) setups. In MTDA (a), target domains are explicitly separated and we are informed by which target an unlabeled sample originates from. In BTDA (b), sub-target IDs are encrypted. If we treat them as a combined single target, transfer learning will lead to the adaptation on a multi-target mixture instead of each hidden target (gray distribution curves in (a), (b)). It implies category-shifted adaptation and negative transfer in practice. Best viewed in color. Abstract(Unsupervised) Domain Adaptation (DA) seeks for classifying target instances when solely provided with source labeled and target unlabeled examples for training. Learning domain-invariant features helps to achieve this goal, whereas it underpins unlabeled samples drawn from a single or multiple explicit target domains (Multi-target DA). In this paper, we consider a more realistic transfer scenario: our target domain is comprised of multiple sub-targets implicitly blended with each other, so that learners could not identify which sub-target each unlabeled sample belongs to. This Blending-target Domain Adaptation (BTDA) scenario commonly appears in practice and threatens the validities of most existing DA algorithms, due to the presence of domain gaps and categorical misalignments among these hidden sub-targets.To reap the transfer performance gains in this new scenario, we propose Adversarial Meta-Adaptation Network (AMEAN). AMEAN entails two adversarial transfer learning processes. The first is a conventional adversarial transfer to bridge our source and mixed target domains. To circumvent the intra-target category misalignment, the second process presents as "learning to adapt": It deploys an unsupervised meta-learner receiving target data and their ongoing feature-learning feedbacks, to discover target clus-ters as our "meta-sub-target" domains. These meta-subtargets auto-design our meta-sub-target DA loss, which empirically eliminates the implicit category mismatching in our mixed target. We evaluate AMEAN and a variety of DA algorithms in three benchmarks under the BTDA setup. Empirical results show that BTDA is a quite challenging transfer setup for most existing DA algorithms, yet AMEAN significantly outperforms these state-of-the-art baselines and effectively restrains the negative transfer effects in BTDA.

Additive Adversarial Learning for Unbiased Authentication

“…The C-MNIST data set is originally built by Lu et al Table 2 summarizes the performance comparisons on C-MNIST. The results clearly show that our methods significantly outperform the direct learning method, which proves the effectiveness of our 49.90 50.09 10.89 MCD-SWD [24] 50.12 49.89 10.69 ML-VAE [3] 77.26 28.06 18.73 DADA [48] 83.90 22.29 15.83 BTDA [8] 85.14 20.82 26.43 JiGen [5] 82. 44 [3] 75.29 36.07 7.97 DADA [48] 83.37 28.06 11.36 BTDA [8] 78.…”

“…Recently, several domain-agnostic learning approaches [48,8,36,49] emerge to typically handle the domain-adaptation problem where the target domain may contain several sub-domains without domain labels [48]. DADA [48] and OCDA [36] propose novel mechanisms and achieve effective domain-adaptation performances, but do not discover latent domains in the target domain and exploit the information.…”

“…DADA [48] and OCDA [36] propose novel mechanisms and achieve effective domain-adaptation performances, but do not discover latent domains in the target domain and exploit the information. By contrast, BTDA [8] clusters raw and deep features to discover latent domains. However, its latent-domain discovery is only feature-based and does not exploit the heterogeneous predictive relationships of p(y | x).…”

Foreground-Aware Stylization and Consensus Pseudo-Labeling for Domain Adaptation of First-Person Hand Segmentation

“…Although single-target domain adaptation methods are sufficient when a well-defined target domain exists, target domains often consist of diverse distributions in the wild [11]. We therefore further validate our method in a multi-target domain adaptation setting [2], [11] that has not been explored much. We aim to simultaneously adapt to multiple target domains by one-time adaptation training.…”

“…When we cannot access the target data due to privacy issues [47], [53], [60], generalizing to an unseen domain is also necessary. We also experimented on a multitarget domain adaptation setting [2], [11], which aims to simultaneously adapt to multiple target domains, and a multiauxiliary domain generalization setting [62], which aims to generalize to an unseen test domain utilizing knowledge from auxiliary domains.…”