Generalization — doing well on unseen data
Performing well on new data from the same source as the training data, not only on the training examples themselves.
Measured on a held-out validation or test set (section 4.4). A model that memorizes its training examples fits them perfectly and generalizes badly.