REVE: Regularizing Deep Learning with Variational Entropy Bound