Neural Computation with Winner-Take-All as the Only Nonlinear Operation