Unsupervised Domain Adaptation via Discriminative Manifold Embedding and Alignment

Abstract

Unsupervised domain adaptation is effective in leveraging the rich information from the source domain to the unsupervised target domain. Though deep learning and adversarial strategy make an important breakthrough in the adaptability of features, there are two issues to be further explored. First, the hard-assigned pseudo labels on the target domain are risky to the intrinsic data structure. Second, the batch-wise training manner in deep learning limits the description of the global structure. In this paper, a Riemannian manifold learning framework is proposed to achieve transferability and discriminability consistently. As to the first problem, this method establishes a probabilistic discriminant criterion on the target domain via soft labels. Further, this criterion is extended to a global approximation scheme for the second issue; such approximation is also memory-saving. The manifold metric alignment is exploited to be compatible with the embedding space. A theoretical error bound is derived to facilitate the alignment. Extensive experiments have been conducted to investigate the proposal and results of the comparison study manifest the superiority of consistent manifold learning framework.

Domain Adaptation via Maximizing Surrogate Mutual Information

“…Class-level methods align the conditional distribution based on pseudo-labels (Chen et al, 2020a;Luo et al, 2020a;Li et al, 2020b;Jiang et al, 2020;Liang et al, 2020;Venkat et al, 2020). Conditional alignment methods (Xie et al, 2018;Long et al, 2018) minimize the discrepancy between conditional distributions.…”

“…However, the conditional distributions from different categories tend to mix together, leading to performance drop. Contrastive learning based methods resolve this issue by discriminating features from different classes (Kang et al, 2019;Chen et al, 2020a;Luo et al, 2020a), but still face the problem of pseudo-label precision.…”

Exploring Category Attention for Open Set Domain Adaptation

“…For robustness, Saito et al [8] take into consideration the task-specific decision boundaries for better alignment of two domains. Luo et al [6] introduce a Riemannian manifold learning framework to make the representations to be both transferable and discriminative. Wang and Breckon [27] integrate supervised subspace learning with structured prediction to formulate an iterative domain adaptation framework.…”

“…As shown in Eq. 5and (6), the attention maps are determined by the features themselves. Thus, two categorysharing samples tend to have similar self-attention maps in the last layers of a network, due to the higher similarity between deeper features [44]- [46].…”

“…The OS (%) of all designed variations of our method. 2±0 6. 51.8±1.0 46.3±0.4 60.9±1.1 56.9±0.8 67.7±1.3 52.8±0.9 54.2±0.9 56.5±1.4 71.2±1.4 70.3±1.4 69.4±0.8 59.4 Ours w/o con 62.1±1.0 51.6±1.7 50.2±1.1 59.4±1.7 64.8±1.7 75.6±1.8 64.9±1.2 65.2±0.7 72.8±1.0 78.6±1.2 72.4±1.8 77.7±1.1 66.3 Ours w/o sharp 57.1±1.2 51.0±0.9 47.8±1.5 60.1±0.7 58.0±1.6 68.2±0.9 57.2±1.1 54.8±1.0 59.8±1.2 73.7±1.9 66.1±1.5 68.9±1.4 60.2 Ours w/o dis 61.7±1.0 54.2±0.8 56.3±1.3 70.6±0.6 68.4±1.4 77.3±0.8 64.2±2.2 65.3±1.6 73.6±1.5 83.3±0.9 76.0±0.9 80.4±1.7 69.3 Ours w/o S-CAM 55.3±1.4 54.3±0.9 54.3±1.8 60.6±1.9 68.4±1.6 73.4±1.2 65.6±1.2 64.9±1.5 67.3±0.9 80.8±1.2 75.9±0.9 78.5±1.2 66.6 Ours w/o C-CAM 60.5±1.5 55.9±0.9 56.9±1.9 71.6±0.7 72.3±1.2 81.0±1.6 63.5±1.2 65.7±1.0 70.3±1.5 79.8±1.5 76.4±1.0 81.0±0.8 69.6 Ours 63.0±0.9 57.4±1.0 59.7±1.4 72.3±1.7 72.7±1.8 81.2±1.3 67.2±1.1 68.4±1.9 74.7±1.0 84.9±1.2 77.1±0.8 82.8±1.1 71.8…”