Gradient Boosting builds a sequence of small models, each correcting the current ensemble's mistakes.