Learning Representations by Maximizing Mutual Information in Variational Autoencoder