A neural network is a crude but powerful simulation of the nervous system, a kind of “electronic biology” that attempts to mimic the way the brain and nerves work together.
The potential of neural networks is far-reaching.
Because they mimic natural processes, they are suited for applications that require human interpretation, such as radar and sonar target identification.
They are also suited for tasks related to machine vision and robotics, or speech, text and handwriting recognition. They also can solve highly complex patterning problems, such as integrated circuit layouts.
In banking, Chase Manhattan uses a “neural net” system to detect credit card fraud; a Security Pacific system analyzes commercial and automobile loan risks.
Medical researchers have designed neural networks into systems that detect abnormal heart sounds and interpret electrocardiograms.
In chemical detection, England’s Warwick University has developed an “electronic nose” to recognize smells. A company in Japan uses a similar system to test sushi for freshness.
In data analysis, Ford Motor Co. is developing a real-time, on-board auto diagnostic system, which simultaneously monitors all operating systems--engine, powertrain, suspension, electronic steering, anti-lock brakes, climate control and so on.
In recognition systems, International Imaging Systems of Milpitas, Calif., uses neural networks to recognize faces with a “casual glance” from a video camera.
In addition to data analysis and recognition uses, Japanese engineers are incorporating neural networks into a variety of fuzzy-logic applications. With fuzzy logic, you can have values other than the standard 0 and 1 of conventional computer logic systems, an electronic maybe to go with the yes or no.
By adding degrees to the traditional digital on/off controller, fuzzy-logic systems more nearly simulate the behavior of a human controller.
Japanese designers are incorporating fuzzy logic into a variety of control systems. A washing machine can be programmed to sample laundry water to determine how dirty the clothes are, then select the optimum detergent level and washing time.
In the past, neural networks existed only in computer software simulation form. On a conventional digital computer, neural programs do not run particularly fast. More recently, the industry has moved into hardware development. Digital and hybrid analog/digital neural network chips increase program speed about 1,000 times. Implementing neural networks in analog form on a chip introduces an enormous advantage in processing speed, up to 1 billion times faster than conventional digital processors.
This recent software-to-hardware shift has changed the way that the granddaddy of neural network companies operates. Until this year, Nestor Corp. of Providence, R.I., sold only neural network software. In late March, the company received a contract from the federal Defense Advanced Research Projects Agency, making Nestor responsible for the design of Intel’s new N1000 neural network chip. (Intel will reduce Nestor’s design to practical silicon wafers under a separate contract.) In early April, Nestor laid off one-third of its software-oriented staff, an action that suggests that the future of neural networks strongly favors chip hardware.
Federico Faggin, the father of the Intel chip that powered the personal computer revolution, is developing a new technology for producing neural networks. He suggests that “the evolution of neural networks could be similar to that observed after the introduction of the first microprocessor in 1971.”
Looking to the future of neural networks, Faggin sees “a fundamental change that is very dramatic. The architecture will be very different: It will be massively parallel. There will be no software.” Parallel architecture provides a number of electronic routes between a question and its answer. Such a design is inherently fault-tolerant: If one element does not work, others will. Processing continues.
As for sociological impact, neural networks will affect an area completely different from standard computers.
They will eliminate jobs higher on the corporate ladder. Whereas standard computers displaced blue-collar workers and assisted white-collar workers, neural networks will tend to displace white-collar workers--from middle management to top decision makers--by substituting their human-like decision-making capabilities for white-collar bodies.
Thus, neutral networks may help realize a long-frustrated goal of the computer age: improving white-collar productivity.