On the complexity of neural network classifiers: a comparison between shallow and deep architectures