Groups
Generalization bounds explain why deep neural networks can perform well on unseen data despite having many parameters.
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.