Deep Probabilistic Ensembles: Approximate Variational Inference through KL Regularization