Neural Network Approximation: Three Hidden Layers Are Enough