Locality Preserving Joint Transfer for Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Domain adaptation (DA) seeks to transfer knowledge from a source domain with ample labels to a target domain with scarce labels.
• Existing DA methods often focus on either feature-level or sample-level adaptation, while recent work suggests the importance of both.
• Optimizing feature and sample adaptation concurrently can lead to synergistic improvements in knowledge transfer.

Purpose of the Study:

• To propose a novel domain adaptation approach that jointly exploits feature adaptation and sample adaptation.
• To enhance knowledge transfer by preserving the manifold structure of data through local consistency.
• To enable adaptation across different feature types (e.g., handcrafted and deep features) in heterogeneous domain adaptation.

Main Methods:

• Jointly optimizing feature adaptation via distribution matching and sample adaptation via landmark selection.
• Incorporating local consistency during knowledge transfer to preserve sample manifold structures.
• Utilizing label propagation for predicting new instance categories.
• Learning domain-specific projections for both homogeneous and heterogeneous domain adaptation.

Main Results:

• The proposed method significantly outperforms conventional and end-to-end deep learning approaches on five benchmark datasets.
• Demonstrated effectiveness in both homogeneous and heterogeneous domain adaptation scenarios.
• Showcased the ability to improve deep feature accuracy using handcrafted features via heterogeneous adaptation.

Conclusions:

• The joint optimization of feature and sample adaptation, coupled with manifold structure preservation, is a powerful strategy for domain adaptation.
• The proposed approach offers a flexible and effective solution for various domain adaptation tasks, including cross-feature type adaptation.
• This work highlights the potential of combining different adaptation levels and feature representations for robust knowledge transfer.