Learning Sparsity of Representations with Discrete Latent Variables