Size and Depth Separation in Approximating Natural Functions with Neural Networks