Unsupervised Domain Adaptation via Discriminative Manifold Propagation

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Unsupervised domain adaptation (UDA) leverages labeled source data for unlabeled target domains.
• Deep learning and adversarial methods show promise but face challenges with pseudo-labeling and batch training limitations.
• Existing methods struggle with preserving target domain data structure and global feature representation.

Purpose of the Study:

• To propose a novel Riemannian manifold learning framework for UDA.
• To enhance feature transferability and discriminability simultaneously.
• To address limitations of hard pseudo-labels and batch-wise training in deep UDA.

Main Methods:

• A probabilistic discriminant criterion using soft labels on the target domain.
• A global approximation scheme based on pre-built prototypes for characterizing global structure.
• Manifold metric alignment to ensure compatibility with the embedding space.
• Derivation of theoretical error bounds for alignment metrics.

Main Results:

• The proposed framework effectively handles both vanilla and partial unsupervised domain adaptation settings.
• Experimental results demonstrate the superiority of the discriminative manifold learning framework.
• The method improves feature transferability and preserves intrinsic data structure.

Conclusions:

• The Riemannian manifold learning framework offers a robust solution for UDA challenges.
• Soft labels and global approximation enhance the adaptability and discriminability of features.
• The framework provides a promising direction for future research in domain adaptation.