A Conceptual Foundation for Generalisation in Natural and Artificial Learning Agents

Generalisation is ubiquitous in both natural and artificial learning agents. Since the early 20th Century, numerous models have been developed that attempt to explain how generalisation arises. While some of these models propose specialised mechanisms applicable only to specific tasks or learning agents, others use common principles which may be more broadly applicable. It remains unclear to what extent the more specialised models are necessary to explain generalisation, and to what extent a shared set of explanatory principles may suffice. To demonstrate the potential of simple principles to explain generalisation, we present a minimal conceptual framework which describes how various forms of generalisation may arise in any learning agent and for any task. Based on the idea that agent behaviours are determined by a set of agent-internal parameters, we illustrate how a wide range of classical generalisation phenomena, including positive, negative, and asymmetric transfer, can arise from the same general mechanisms. We highlight several related generalisation phenomena typically overlooked in the classical literature, and use the presented framework to clarify numerous challenges for the empirical study of generalisation. We emphasise that specialised theories of generalisation may still be required to explain specific empirical observations. The presented framework provides a minimal foundation from which empirical observations requiring more specialised explanations can be identified.