Adaptive Transfer Kernel Learning for Transfer Gaussian Process Regression.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Transfer regression addresses challenges in cross-domain knowledge transfer, crucial for applications like engineering design and localization.
• Effective domain relatedness modeling is key to adaptive knowledge transfer, enabling better performance across different domains.

Purpose of the Study:

• To formally define and explore transfer kernels for explicitly modeling domain relatedness.
• To propose advanced transfer kernel forms (Trkαβ and Trkω) addressing limitations of basic forms in complex data.
• To develop Gaussian process models (TrGPαβ and TrGPω) incorporating these novel transfer kernels.

Main Methods:

• Formal definition and introduction of three basic general forms of transfer kernels.
• Proposal of two advanced transfer kernel forms, Trkαβ (multiple kernel learning) and Trkω (neural networks).
• Development of Gaussian process models (TrGPαβ, TrGPω) with guaranteed positive semi-definiteness and interpretable domain relatedness.

Main Results:

• Demonstrated effectiveness of Trkαβ and Trkω in modeling domain relatedness.
• Empirical studies confirmed the superior performance of TrGPαβ and TrGPω in transfer adaptiveness.
• The proposed methods successfully handle complex real-world data, improving transfer regression.

Conclusions:

• Transfer kernels provide an effective mechanism for explicit domain relatedness modeling in transfer learning.
• The advanced forms Trkαβ and Trkω, integrated into TrGPαβ and TrGPω, significantly enhance transfer adaptiveness.
• This work offers a robust framework for improving cross-domain knowledge transfer in various applications.