Exploring uncertainty in pseudo-label guided unsupervised domain adaptation. (December 2019)

• Record Type: Journal Article
• Authors: Liang, Jian; He, Ran; Sun, Zhenan; Tan, Tieniu
• Is Part Of: Pattern recognition. Volume 96(2019:Dec.)
• Publication Date: 2019-12
• Subjects: Unsupervised domain adaptation -- Pseudo labeling -- Feature transformation -- Progressive learning -- Transfer learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.106996
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We introduce the problem of addressing the uncertainties of target pseudo labels, which is important yet under-studied in the domain adaptation area. Specifically, we propose a novel approach that progressively includes more target samples into training and incorporates previously estimated class confidence scores to characterize both the within- and cross- domain relations. Especially, we provide a more accurate conditional distribution discrepancy than those of previous studies. To fully exploit the discriminative cross-domain structures, we compensate joint distribution adaptation by designing a new local triplet-wise instance-to-center margin for better separability. Experimental results demonstrate the superiority of our method over recent state-of-theart approaches. Particularly, on the challenging Office-Caltech dataset with VGG features show that our method advances the best reported average accuracies from 83.4% to 88.2% and 81.7% to 87.2%, respectively. Abstract: Due to the unavailability of labeled target data, most existing unsupervised domain adaptation (UDA) methods alternately classify the unlabeled target samples and discover a low-dimensional subspace by mitigating the cross-domain distribution discrepancy. During the pseudo-label guided subspace discovery step, however, the posterior probabilities (uncertainties) from the previous target label estimation step are totally ignored, which may promote the error accumulation and degrade the adaptation performance. To address this issue, we propose to progressively increase the number of target training samples and incorporate the uncertainties to accurately characterize both cross-domain distribution discrepancy and other intra-domain relations. Specifically, we exploit maximum mean discrepancy (MMD) and within-class variance minimization for these relations, yet, these terms merely focus on the global class structure while ignoring the local structure. Then, a triplet-wise instance-to-center margin is further maximized to push apart target instances and source class centers of different classes and bring closer them of the same class. Generally, an EM-style algorithm is developed by alternating between inferring uncertainties, progressively selecting certain training target samples, and seeking the optimal feature transformation to bridge two domains. Extensive experiments on three popular visual domain adaptation datasets demonstrate that our method significantly outperforms recent state-of-the-art approaches.