On the Nyström Method for Approximating a Gram Matrix for Improved Kernel-Based Learning