Learning Sparse Codes with a Mixture-of-Gaussians Prior