ABSTRACT
At the end of the last chapter we set out a programme that aimed to train all the weights in multilayer nets with no a priori knowledge of the training set and no hand crafting of the weights required. It turns out that the perceptron rule is not suitable for generalization in this way so that we have to resort to other techniques. An alternative approach, available in a supervised context, is based on defining a measure of the difference between the actual network output and target vector. This difference is then treated as an error to be minimized by adjusting the weights. Thus, the object is to find the minimum of the sum of errors over the training set where this error sum is considered to be a function of (depends on) the weights of the network. This paradigm is a powerful one and, after introducing some basic principles and their application to single-layer nets in this chapter, its generalization to multilayer nets is discussed in Chapter 6.
