Neural Networks are constructs that present a stimulus of some sort to a collection of interconnected data paths deliberately modeled on the neurons in mammalian brains. Connections are arbitrarily altered searching for an arrangement of connections that yields optimum results. Typically, the interconnections are not actually connected or disconnected, instead, weighted connections are used, and the weightings are tinkered with. Several of algorithms are known for altering the weightings based on the history of past trials.
Theoretical analysis of neural networks started in the 1950s. Practical neural networks developed in the late 1980s.
Difficulties with Neural Networks revolve around their indeterminate nature. At best, they can be evaluated by their ability to arrive at acceptable answers to a set of test problems. There is no way to know if their answer to the next similar problem will be "correct" or even acceptable.
Neural networks are used in pattern recognition, and various classification tasks. Attempts have been made to apply them to a variety of jobs for which there is no known, practical, determinant algorithm. Neural Networks might be considered for any situation for which a guess is better than having no answer whatsoever. For example, neural network commercial products are available that will search the Internet for art based on samples of art that an individual preferences.
Return To Index Copyright 1994-2008 by Donald Kenney.