Depth-Width Tradeoffs in Approximating Natural Functions with Neural Networks