Neural Networks in Electrical Engineering
Howard J. Silver · 2006
Neural networks are a subject thought by some to be a branch of artificial intelligence (AI). Although neural network research dates back about sixty years, it is generally considered a newer problem solving technique than AI. Whereas AI is applied in the form of “if-then” rules, neural networks attempt to model the structure of the human brain and are based on selflearning. The structure is highly parallel, resulting in the ability to selforganize to represent information and rapidly solve problems in real time. However, even with today’s high-speed computers, artificial neural networks are limited by the fact that they can only replicate a small fraction of the brain’s total structure. Early work in neural networks was researched primarily by neuroscientists and psychologists, whose interests were in learning more about the brain. Soon after, mathematicians, physicists and computer scientists joined in. Engineers were attracted to the subject by recognizing its potential for solving problems for which other techniques, such as traditional computer programming, may not have been feasible or economical. Neural networks tend to do well at recognizing patterns within seemingly random data, as opposed to applications involving rules of logic. As a result, they have been applied to areas of business such as evaluation of loan applications and stock market pattern prediction, and to other areas such as handwriting and speech recognition and the detection of signals in a noisy environment. This paper will present a brief overview of a simple two-layer neural network structure and a supervised learning algorithm called Perceptron. In supervised learning the network is trained to map a given set of input patterns to known outputs. After the patterns have been “learned” the network can be tested for patterns with errors present. With the aid of MATLAB programs created by the author, Perceptron learning is first applied to a problem of alphabetic character recognition, which relates to the application of handwriting analysis. The algorithm is then used to train a network to distinguish among square, triangular and sinusoidal waveforms. The network is then tested by, superimposing a given level of “random” noise on each waveform, and determining whether the “noisy” waveforms can still be distinguished from each other. A third example