Prototype-Based Multisource Domain Adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Unsupervised domain adaptation (UDA) addresses knowledge transfer from labeled source to unlabeled target domains.
• Multisource domain adaptation (MDA) extends UDA by leveraging multiple labeled source domains.
• Existing MDA methods often fail to optimally integrate information from diverse source domains.

Purpose of the Study:

• To propose a novel prototype-based method for multisource domain adaptation (MDA).
• To enable effective knowledge transfer from multiple labeled source domains to an unlabeled target domain.
• To improve the performance of domain adaptation by disentangling domain-invariant and domain-specific features.

Main Methods:

• A feature extraction network disentangles domain-invariant and domain-specific features.
• Inherent class prototypes and domain prototypes are estimated from these features.
• A prototype mapping is learned, and features are aligned twice at the feature level.

Main Results:

• The proposed method successfully transfers semantic category information using constructed class prototypes.
• The feature extraction network is progressively adjusted by aligning features with class prototypes.
• Experiments on public datasets demonstrate the effectiveness of the novel MDA approach.

Conclusions:

• The novel prototype-based MDA method effectively transfers semantic knowledge to unlabeled target domains.
• Dual feature-level alignment enhances domain-invariant features and prototype closeness.
• The inherent class prototypes serve as effective classifiers for the target domain.