Guided Discrimination and Correlation Subspace Learning for Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain adaptation is crucial for transfer learning, aiming to apply knowledge from source to target domains.
• Existing methods often overlook feature discriminability, correlation, and the risk of negative transfer.

Purpose of the Study:

• To propose a novel domain adaptation method, Guided Discrimination and Correlation Subspace Learning (GDCSL), for cross-domain image classification.
• To address limitations in existing methods by focusing on domain-invariant, category-discriminative, and correlated features, while avoiding negative transfer.

Main Methods:

• GDCSL minimizes intraclass scatter and maximizes interclass distance for discriminative features.
• A novel correlation term extracts maximally correlated features between domains.
• Sample reweighting and a label selection scheme mitigate negative transfer and improve pseudo-label accuracy in the semi-supervised extension (Semi-GDCSL).

Main Results:

• GDCSL effectively learns domain-invariant, category-discriminative, and correlated features.
• The method demonstrates superior performance compared to state-of-the-art domain adaptation techniques on benchmark datasets.
• The semi-supervised extension (Semi-GDCSL) further enhances performance through improved pseudo-labeling.

Conclusions:

• GDCSL offers a robust framework for cross-domain image classification by addressing key overlooked factors in domain adaptation.
• The proposed method significantly improves transfer learning efficiency and accuracy.
• GDCSL and its semi-supervised variant represent advancements in domain adaptation research.