Cost of a sparse autoencoder based on KL Divergence