Learning Discrete Representations via Information Maximizing Self-Augmented Training