Transfer Learning with Kernel Methods

Area of Science:

• Machine Learning
• Computational Science
• Artificial Intelligence

Background:

• Kernel methods offer a computationally efficient approach for various machine learning tasks.
• Scalable transfer learning for kernel methods across diverse tasks with differing label dimensions remains a challenge.

Purpose of the Study:

• To propose a novel transfer learning framework for kernel methods.
• To enable scalable adaptation of models across general source and target tasks.
• To analyze the performance characteristics of kernel-based transfer learning.

Main Methods:

• Developed a framework to project and translate source models to target tasks.
• Applied the framework to image classification and virtual drug screening.
• Investigated performance scaling laws with varying numbers of target examples.

Main Results:

• Demonstrated the framework's effectiveness in image classification and virtual drug screening.
• Identified simple scaling laws governing transfer-learned kernel performance.
• Derived exact scaling laws in a simplified linear setting.

Conclusions:

• The proposed framework facilitates effective and scalable kernel-based transfer learning.
• Performance scaling laws provide insights into model adaptation efficiency.
• This work advances the application of kernel methods in complex domains.