Knowledge transfer using context-sensitive multiple task learning.

摘要

Machine lifelong learning, or ML3, is concerned with building machines that continue to learn over time, while drawing on previously learned knowledge to aid in new learning. This thesis presents context-sensitive Multiple Task Learning, or csMTL, as a method of inductive transfer by encoding examples with contextual information. The research is inspired by problems encountered with standard Multiple Task Learning (MTL), which is shown to be an inappropriate method for ML3. The thesis presents relevant background of machine learning, artificial neural networks, and requirements for an ML 3 system. It also presents preliminary mathematical theories that explore possible reasons that csMTL produces better models than standard MTL. The focus of the thesis is upon empirical studies, which are conducted using five task domains, which include two synthetic domains and three real-world domains. The studies show that, using artificial neural networks, csMTL is able to produce more accurate models than MTL. Studies using inductive decision trees and the k-nearest neighbour algorithm suggest that the improvement is due to characteristics of the ANN model, and not to machine learning models in general.