Shared wasserstein adversarial domain adaption

Abstract

In numerous real-world applications, obtaining labeled data for a specific deep learning task can be prohibitively expensive. We present an innovative framework for unsupervised training of deep neural networks, drawing inspiration from the adversarial learning paradigm. Our approach incorporates the cycle-consistency constraint to effectively constrain the generator. Furthermore, we capitalize on the reconstructed samples, treating them as "real" samples for the discriminator during classification. This idea stems from the success of Wasserstein GAN, which leverages its gradient property and promising generalization bound during network training. Simultaneously, we employ a shared latent-data space constraint to ensure compatibility between the source domain and its corresponding target domain. This constraint facilitates effective knowledge transfer from the source to the target domain, even in the absence of labeled data for the target domain. To enhance the performance of the target domain classifier, we introduce association chains that link the embeddings of labeled samples to those of unlabeled samples and vice versa. By encouraging correct association cycles that ultimately return to the same class from which the association began, and penalizing wrong associations leading to a different class, we ensure accurate predictions. Our proposed method, named Shared Wasserstein Adversarial Domain Learning (SWADL), combines these novel constraints. Through extensive evaluations on benchmark datasets such as MNIST, SVHN, and USPS, we demonstrate that SWADL consistently outperforms current mainstream methods. It achieves superior results in unsupervised domain adaptation tasks, addressing the challenge of limited labeled data in real-world scenarios. The code and models are available at https://github.com/Jayee-chen/Adversarial-Domain-Adaptation.git.