Groups
Statistical learning theory explains why a model that fits training data can still predict well on unseen data by relating true risk to empirical risk plus a complexity term.
VC dimension measures how many distinct labelings a hypothesis class can realize on any set of points of a given size.