Problems which are well suited to neural networks are those in which there are large sets of examples, the task involves associating objects in one set with objects in another set, the problem is not easily solved using another approach, or the problem requires recognition of patterns from incomplete, noisy, or partially incorrect input.
In addition, neural networks are appropriate for applications requiring fault tolerance. If a connection or processing element were to fail, the network would experience a graceful degradation in accuracy rather than a complete failure.
Neural networks are currently being used in several interesting application areas. These areas are summarized in the following paragraphs:
Adaptive Noise Cancelling. This technology cleans up echoes on telephone lines and reduces transmission errors on modems. The adaptive filtering process was developed in the 1950s by Bernard Widrow, a pioneer in the field.
Airline Seat Management. This system by BehavHeuristics advises airlines on seat yield management to increase bookings and profits.
Bomb Sniffer. SAIC developed a bomb detector for the Federal Aviation Administration (FAA) using thermal neuron analysis and a neural network to interpret the results.
Character Recognition. The Electric Power Research Institute (EPRI), the U.S. Postal Service, the Internal Revenue Service (IRS), banks, and credit card companies are advancing the state-of-the-art in hand-lettered text recognition. Constrained character recognition is currently employed while it is anticipated that unconstrained recognition will be available within the next five years.
Diagnosis and Control. One application uses a neural network to determine when a machine needs to be rebuilt and which sensors are most useful in determining the machine's health.
Financial Analysis. AVCO Financial Services uses a neural network system for credit risk analysis. Peat Marwick uses a neural network to predict bank failures.
Process Monitoring. GTE Laboratories is using a neural network in its fluorescent bulb manufacturing plant to identify parameters which most affect production. Heat, pressure, and chemical inputs are compared to yield and performance. Other process monitoring and control applications are being explored by the National Science Foundation (NSF).
Robotic Arm Control. Neural networks are used to increase both the accuracy and performance of computing the inverse kinematics for a robotic arm.
Speech Recognition. Intel has had an application in place since 1983 which recognizes a 100-word vocabulary for a single user with better than 99% accuracy.
Vision. An automated inspection and a visual navigation system provide two examples of vision-based neural network applications.
A neural network trains by accepting information and conclusions. It reviews the material repeatedly, adjusting from its mistakes until it has finally adapted itself to correctly perform the task.
The two key features of neural networks are that they are trainable and that they are naturally highly parallel. Neural networks are not so much programmed as they are trained. Their parallel nature facilitates high speed computations and also provides a measure of fault tolerance.
Not every neural network built adapts correctly to the problem, but there is a growing body of literature and experience on how and when to apply neural networks.
Neural networks do not replace, but rather complement, other methods of computing. While expert systems are more appropriate for inferencing and explaining decisions based on domain knowledge rules, neural networks are stronger at recognizing patterns.
The following references are provided as suggested reading for additional information on Parts I and II of this series:
Ahmad, Z., "Improving the Solution of the Inverse Kinematic Problem in Robotics Using Neural Networks," Journal of Neural Network Computing, Vol. 1, No. 4, Spring 1990.
Anderson, J.A. and Rosenfeld, E., Eds. Neurocomputing: Foundations of Research. MIT Press, Boston, MA, 1988, p. 125.
Antognetti, P. and Milutinovic, V., Eds. Neural Networks: Concepts, Applications, and Implementations. Volumes I-IV. Prentice Hall, Englewood Cliffs, NJ, 1991.
Caudill, M. and Butler, C. Naturally Intelligent Systems. MIT Press, Cambridge, MA, 1990.
Coleman, K.G., et. al., "Neural Networks for Bankruptcy Prediction," AI Review, Summer 1991, pp. 48-50.
DARPA Neural Network Study. MIT Press, Lexington, MA, 1988.
Dietz, W.E., Kiech, E.L., and Ali, M., "Jet and Rocket Engine Fault Diagnosis in Real Time," Journal of Neural Network Computing, Vol. 1, No. 1, Summer 1989.
Doherty, R., "FAA Adds 40 Sniffers," Electronic Engineering Times, September 4, 1989, p. 16.
EDventure Holdings Newsletter, Release 1.0, New York, July 9, 1989, pp. 8-9.
Fukushima, K., Miyake, S., and Takayuki, I., "Neocognitron: A Neural Network Model for a Mechanism of Visual Pattern Recognition," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 13, No. 5, September/October 1983, pp. 826-34.
Glover, D.E., "Optical Processing and Neurocomputing in an Automated Inspection System," Journal of Neural Network Computing, Vol. 1, No. 2, Fall 1989.
Green, L., "Neural Networks Still Waiting on the Brink," Information Week, Issue 186, September 12, 1988, p. 52.
Grossberg, S., Ed. Neural Networks and Natural Intelligence. MIT Press, Cambridge, MA, 1988.
Hatsopoulos, G.G. and Warren, W.H. Jr., "Visual Navigation with a Neural Network," Neural Networks, Vol. 4, No. 3, 1991, pp. 303-317.
Hebb, D.O. The Organization of Behavior. Wiley, New York, 1949.
Hecht-Nielsen, Robert, "Neurocomputing: Picking the Human Brain," IEEE Spectrum, Vol. 25, No. 3, March 1988, pp. 36-41.
Isik, C. and Uhrig, R.E., "Neural Networks and Power Utility Applications," EPRI Knowledge Based Technology Applications Center Seminar, September 1991.
Johnson, R.C., "Neural Nose to Sniff Out Explosives at JFK Airport," Electronic Engineering Times, Issue 536, May 1, 1989, pp. 1, 86.
Kim, E.J. and Lee Y., "Handwritten Hangul Recognition Using a Modified Neocognitron," Neural Networks, Vol. 4, No. 6, 1991, pp. 743-750.
Levin, E., Gewirtzman, R., and Inbar, G.E., "Neural Network Architecture for Adaptive System Modeling and Control," Neural Networks, Vol. 4, No. 2, 1991, pp. 185-191.
Lippmann, Richard P., "An Introduction to Computing with Neural Nets," IEEE ASSP Magazine, April 1987, pp. 4-22.
Minsky, M. and Papert, S. Perceptrons: An Introduction to Computational Geometry. MIT Press, Cambridge, MA, 1969.
Nelson, M. McCord and Illingwirth, W.T. A Practical Guide to Neural Nets. Addison-Wesley, Reading, MA, 1991.
Pollack, A., "More Human than Ever, Computer is Learning to Learn," New York Times, September 15, 1987, Section C, p. 2.
Reed, F., "USPS Investigates Neural Nets," Federal Computer Week, Vol. 3, No. 4 , January 23, 1989, p. 10.
Reed, F., "Neural Networks Fall Short of Miraculous But Still Work," Federal Computer Week, Vol. 3, No. 4 , January 23, 1989, p. 28.
Rumelhart, D.E. and McCleland, J.L., Eds. Parallel Distributed Processing, Volumes I-II. MIT Press, Cambridge, MA, 1986.
Schwartz, T.J., "IJCNN `89," IEEE Expert, Vol. 4, No. 3, Fall 1989, pp. 77-78.
Scheff, K. and Szu, H., "Gram-Schmidt Orthogonalization Neural Networks for Optical Character Recognition," Journal of Neural Network Computing, Vol. 1, No. 3, Winter 1990.
Schalkoff, R.J. Pattern Recognition: Statistical, Structural, and Neural Approaches. John Wiley & Sons, New York, NY, 1992.
Tom, M.D. and Tenorio, M.F., "Short Utterance Recognition using a Network with Minimum Training," Neural Networks, Vol. 4, No. 6, 1991, pp. 711-722.
Vemuri, V. Artificial Neural Networks: Theoretical Concepts. Computer Society Press of the IEEE, 1988.
VerDuin, W.H., "Neural Networks for Diagnosis and Control," Journal of Neural Network Computing, Vol. 1, No. 3, Winter 1990.
Werbos, P.J., "Neural Networks for Control and System Identification," Heuristics, Vol. 3, No. 1, Spring 1990, pp. 18-27.
