Cascading Negative Transfer in Networks of Machine Learning Systems

Abstract

Wide-spread use of transfer learning establishes inter-linkages between otherwise disparate parts of systems. These inter-linkages create systemic risks of cascading failure. This paper presents a formal framework for cascading negative transfer. A novel definition is provided for the efficiency of transfer learning networks. Catalysts and pre-conditions for cascading negative transfer are identified. Throughout, qualitative examples in unmanned aerial systems are used to clarify the presented theory. The conclusion is drawn that negative transfer can propagate among machine learning systems that are inter-linked by transfer learning, and that solution methods spanning algorithm design and systems engineering can address systemic risks from cascading negative transfer without eliminating transfer learning altogether.