Partial-Hessian Strategies for Fast Learning of Nonlinear Embeddings