Page 171 -

5.3 The Perceptron Concept 159

5.3 The Perceptron Concept

The network unit depicted in Figure 5.7, with a threshold function h(x) as
activation function, was studied by Rosenblatt (1962) who, inspired by the
similarity with the physiological structure of the neurons, called it perceptron. A
similar device was studied by Widrow and Hoff (1960) under the name of adaline
(ADAptive LINear Element). We will consider, in later sections, cascades of such
units that bear some resemblance to networks of physiological neurons, the inputs
playing the role of the synapses and the outputs playing the role of the axons.
Within this analogy, positive weights are interpreted as reinforcing connections
and negative weights as inhibiting connections. It was this analogy that earned the
perceptron and its other relatives the name of arttficial neural networks or simply
neural networks or neural nets for short. However, the reader must not carry the
analogy too far, as it is quite coarse, and one should consider the engineering
terminology neural networks only as a convenient way of referring to artificial
connectionist networks with learning properties.
The perceptron output is given by:

It is of course also possible to use transformed inputs, as already mentioned in
the previous section.

Figure 5.11. Classification of a feature vector x using the perceptron rule, based on
the distance 1-1, to the discriminant d(x).

Consider a two-class situation, w, with target value +I and with target value
-1. Then, from (5-13) and the definition (5-10a) of the hard-limiting function, we