Escaping Saddle Points for Effective Generalization on Class-Imbalanced Data

Following this observation, we also find that optimization methods designed to escape from saddle points can be effectively used to improve generalization on minority classes.