The field of neural networking is today being explored rapidly. As Stergiou and Siganos (http://www.doc.ic.ac.uk/~nd/surprise_96/journal/vol4/cs11/report.html) point out, research had been nearly halted in the field for nearly three decades based on a misconception regarding the ability of a model regarding the mammalian system of vision and its ability to represent different shapes. Until this conceptual problem with the so-called ‘Perceptrons’ was accounted for, the limits of the field were through to be known. Today, on the contrary, the field of neural networking holds the potential to revolutionize the way computer software functions.

Foremost, neural networks are merely modeled after the nervous systems of mammals, and are not meant to mimic their specific function. Instead, the basic principle is borrowed from the biological understanding: in order to generate a desired outcome from a specific set of inputs, a complex intermediary network can use different methods of learning and thereby perform the required behavior. In this way, the network functions much like a nervous system, but these artificial nervous networks are most useful in solving different types of problems than conventional computer algorithms. These networks are instead able to perform primitive types of problem solving for the subjects in which they have been trained as ‘experts.’ Another advantage of this problem solving method is the parallel processes utilized by the network.

The architecture of these neural networks is usually either feedforward or a feedback mechanism. The former, also known as a top-down network, allows for one way flow of data. The more complex feedback networks, however, allow for the network to adapt to its own output, therefore making it approach an equilibrium, at which it remains until disturbed by further input. Neural networks, moreover, are often split into three distinct layers: input, hidden, and output units. Hidden units are essential in the way input information is transformed into output information. The ways in which networks are taught to respond to input are often separated into the classes of associative mapping and regulatory detection.

Neural networks, aside from the computing applications, may be useful in a number of other fields, such as data validation, customer research, and industrial process control. Neural networks can also be very useful in medical settings, where they are used to diagnose patients in a way which is not entirely algorithm dependent. They have even been suggested for use in fields as diverse as marketing and credit evaluation.