SPARLING: Learning Latent Representations with Extremely Sparse Activations