Fundamental limits of shallow neural networks with small training sets