A Systems Theory of Transfer Learning with Application

Machine learning is an emerging technology with few principled engineering frameworks to guide its application. In particular, theoretical frameworks for understanding the interrelationships between systems and their learning processes are underdeveloped. The presented research addresses this gap by using Mesarovician abstract systems theory as a mathematical superstructure for learning theory, using the synthesized theory to characterize transfer learning systems, and operationalizing the resulting findings towards an empirical methodology for system design and operation. In particular, transfer distance, the abstract distance knowledge must traverse to be transferred from one system to another, is used as a metric for generalization difficulty, and thereby as a mechanism for relating the generalization of component learning systems to overall system design and operation. In sum, the presented research develops a systems theoretic framework for transfer learning and shows how it can be used to develop and organize best practices and tradecraft in systems engineering for artificial intelligence.