Manifold-tiling Localized Receptive Fields are Optimal in Similarity-preserving Neural Networks