A Systems Theoretic Perspective on Transfer Learning

Tyler Cody, Stephen C. Adams, Peter A. Beling

Abstract: The machine learning formulation of transfer learning is incomplete from a systems theoretic perspective. It focuses on algorithm parameters, features, and samples, and neglects the perspective offered by considering system structure and system dynamics. Furthermore, while the machine learning formulation serves to classify methods and literature, the systems theoretic formulation presented herein serves to provide a framework for the top-down design of transfer learning systems, including a novel definition of transfer learning and identification of key design parameters. We dichotomize the transfer learning problem into a question of transferring system structure and dynamics. We formulate our framework in the context of input-output systems. We use an actuator system as a case study throughout the paper to ground the discussion in a real-world transfer problem.