8 Supervised Learning
     One of the main issues covered in this chapter is generalization in neural networks. After some preliminary remarks, perceptrons and their learining rule are finely presented and illustrated. The strength of such an approach is observed to derive from the fact that now, unlike the linear case of the network in Chapter 6, the activation function is non-linear, allowing better generalization properties. The backpropagation algorithm is nicely introduced and illustrated in terms of an example involving the learning of family trees. The next section covers recurrent networks, which are described and illustrated in terms of the XOR problem. The chapter concludes by discussing some special advantages of the sigmoid function.