ABSTRACT
In this section single-layer neural network models are considered. Some of these models are simply single neurons, which, however, are used as the building blocks of larger networks. We discuss the perceptron which was developed in the late 1950s, and played a pivotal role in the history of neural networks. Nowadays, it is rarely used in real-life applications as more versatile and powerful models are available. Nevertheless, the perceptron remains an important model due to its simplicity and the influence it had in the development of the field. Today most neural networks consist of a large number of neurons, each largely resembling the perceptron. The adaline, also a single neuron model, was developed contemporaneously with the perceptron and is trained by the widely applied least mean square (LMS) algorithm. Both adaline and its extension known as madaline found many real applications, especially in signal processing. Notable is that the backpropagation algorithm is a generalization of LMS. A powerful technique, called learning vector quantization (LVQ) is also presented. This technique is used often in data compression and data classification applications. Another model discussed is the CMAC (cerebellar model articulation controller), which has many applications especially in robotics. All of these models are trained in a supervised manner: for each input, there is a target output, based on which an error signal is generated, based on which the weights are adapted. Also discussed are the instar and outstar models, single neurons which are closer to biology, and are primarily of theoretical interest.
