A Phenomenon in machine learning where a machine learning model generalizes to a test set only long after it achieved perfect loss on the training set.