Vicinal and categorical domain adaptation. (July 2021)

• Record Type: Journal Article
• Authors: Tang, Hui; Jia, Kui
• Is Part Of: Pattern recognition. Volume 115(2021)
• Publication Date: 2021-07
• Subjects: Unsupervised domain adaptation -- Categorical domain adaptation -- Vicinal domain adaptation -- Cross-domain weighting -- Domain augmentation; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.107907
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose novel adversarial losses at multiple levels on both the source and target domains to promote categorical domain adaptation. Based on a joint domain-category classifier, the category-level adversarial loss improves over the domain-level one by a heterogenous, cross-domain weighting design. We propose to use vicinal domains to augment the alignment of original domains. We present novel adversarial losses for vicinal domain adaptation based on the above designs, giving rise to Vicinal and Categorical Domain Adaptation. To recover the intrinsic target discrimination damaged by adversarial feature alignment, we propose Target Discriminative Structure Recovery, which fine-tunes the trained model by semantically anchored spherical k-means. We analyze the working mechanisms of our key designs in principle. Particularly, we explain our cross-domain weighting scheme by connecting it with information theory and optimization equilibrium. We achieve the new state of the art on several commonly used benchmark datasets. Abstract: Unsupervised domain adaptation aims to learn a task classifier that performs well on the unlabeled target domain, by utilizing the labeled source domain. Inspiring results have been acquired by learning domain-invariant deep features via domain-adversarial training. However, its parallel design of task and domain classifiers limits the ability to achieve a finer category-level domain alignment. To promote categorical domain adaptation (CatDA), based on a joint category-domain classifier, we propose novel losses of adversarial training at both domain and category levels. Since the joint classifier can be regarded as a concatenation of individual task classifiers respectively for the two domains, our design principle is to enforce consistency of category predictions between the two task classifiers. Moreover, we propose a concept of vicinal domains whose instances are produced by a convex combination of pairs of instances respectively from the two domains. Intuitively, alignment of the possibly infinite number of vicinal domains enhances that of original domains. We propose novel adversarial losses for vicinal domain adaptation (VicDA) based on CatDA, leading to Vicinal and Categorical Domain Adaptation (ViCatDA) . We also propose Target Discriminative Structure Recovery (TDSR) to recover the intrinsic target discrimination damaged by adversarial feature alignment. We also analyze the principles underlying the ability of our key designs to align the joint distributions. Extensive experiments on several benchmark datasets demonstrate that we achieve the new state of the art.