Feedforward connectionist networks are primarily used for pattern classification. The reason is that such networks are good generalization devices, offering a previously unavailable possible explanation for the human phenomenon of pattern recognition. It is natural to think that `mysterious' training procedures, such as backpropagation, are responsible for the generalization abilities of such networks. This is, however, not true. This tutorial will demystify the training procedure explaining at a conceptual level what it does. We will see that the generalization abilities -- the real discovery of connectionist networks -- depend on the network architecture and not the training procedure. Nonetheless, the backpropagation training procedure is often criticized for not affording a plausible brain implementation. To counter this, we will show the feasibility of a training procedure that is based on the existence of a global reward signal combined with synaptic noise, both of which are biologically plausible assumptions.