Exploiting Numerical Sparsity for Efficient Learning : Faster Eigenvector Computation and Regression