Uncertainty Autoencoders: Learning Compressed Representations via Variational Information Maximization