On Ultrametricity in Feedforward Neural Network Error Surfaces