The Learning Dynamcis of a Universal Approximator