Efficient Representation of Low-Dimensional Manifolds using Deep Networks