Joint Adversarial Domain Adaptation

Abstract

Domain adaptation aims to transfer the enriched label knowledge from large amounts of source data to unlabeled target data. It has raised significant interest in multimedia analysis. Existing researches mainly focus on learning domain-wise transferable representations via statistical moment matching or adversarial adaptation techniques, while ignoring the class-wise mismatch across domains, resulting in inaccurate distribution alignment. To address this issue, we propose a Joint Adversarial Domain Adaptation (JADA) approach to simultaneously align domain-wise and classwise distributions across source and target in a unified adversarial learning process. Specifically, JADA attempts to solve two complementary minimax problems jointly. The feature generator aims to not only fool the well-trained domain discriminator to learn domain-invariant features, but also minimize the disagreement between two distinct task-specific classifiers' predictions to synthesize target features near the support of source class-wisely. As a result, the learned transferable features will be equipped with more discriminative structures, and effectively avoid mode collapse. Additionally, JADA enables an efficient end-to-end training manner via a simple back-propagation scheme. Extensive experiments on several real-world cross-domain benchmarks, including VisDA-2017, ImageCLEF, Office-31 and digits, verify that JADA can gain remarkable improvements over other state-of-the-art deep domain adaptation approaches. CCS CONCEPTS• Computing methodologies → Transfer learning; Image representations; • Computer systems organization → Neural networks.

Cross-domain Cross-modal Food Transfer

“…Domain adaptation is not a new problem and has been intensively studied in [12,20,21,25,26,31,32]. The goal is to transfer the knowledge learnt in a source domain to annotate data in a target domain.…”

Imbalanced Source-free Domain Adaptation

“…Metric-based methods [22,23,31,33] propose statistics to quantify distribution divergence. Adversarial methods [9,20,26,40,52] train a pair of networks in an adversarial fashion to realize implicit alignment. Reconstruction methods reconstruct data in the new domain by an encoder-decoder or GAN discriminator [3,11].…”

Domain Adversarial Tangent Learning Towards Interpretable Domain Adaptation

“…The Conditional Domain Adaptation Network (CDAN) [6] utilizes a multi-linear map g(X) ⊗ d(X) to model uncertainty and class affiliation into L d . The Joint Adversarial Adaptation Network [1] introduces two non-equal classifiers g 1 (X t ), g 2 (X t ) predicting target labels. The network is learned to maximize the difference between both predictions using L 1 -norm, and by GRL, the feature extractor learns a classifier-level independent representation.…”

“…Deep Domain Adaptation (DDA) is a technique to learn a network capable of adapting from a training or source domain to an evaluation or target domain, assuming the domain data distributions are related but inevitably different. Joint Adversarial Domain Adaptation (JADA) [1] is the current state of the art in DDA and learns the following multi-task schema: a classifier, a domain discriminator, and a local adaptation divergence on top of a feature extractor network. The classifier is learned on labeled source data and should generalize well to the target domain.…”