The training error decreases with increasing neuron count and plateaus beyond 28 neurons per hidden layer. For the two-hidden-layer network, error stabilization is ...