The Connection Between Approximation, Depth Separation and Learnability in Neural Networks