Learning is defined operationally as improvement, with respect to a stated class of tasks and a stated performance measure, that accrues with experience — the canonical three-part formulation (task, performance measure, experience) that grounds machine learning in the older philosophical tradition of inductive inference. Because the definition is purely behavioral it is necessary but not sufficient: systems that satisfy it need not be said to learn in any meaningful sense. The definition organizes the field into supervised learning (learning an input-to-output map, subdivided into classification for categorical outputs and regression for continuous ones), unsupervised learning (discovering structure in data, as in clustering and association rule mining), and reinforcement learning (learning to control a system under a cost criterion), each carrying its own performance measures and its own methodological difficulties.
Learning is defined operationally as improvement, with respect to a stated class of tasks and a stated performance measure, that accrues with experience — the canonical three-part formulation (task, …