Learning Efficient Coding of Natural Images with Maximum Manifold Capacity Representations