Neural networking is still a rarefied science to many, but try out the real thing and you’ll get a glimpse of what it is all about. On a sunny June day at Fitzwilliam College, Cambridge, 80 researchers assembled for a set of seminars under the auspices of the Neural Computing Applications Forum, and the energy generated by their enthusiasm and ideas sparked its own network: you could practically see the neurons flying around the room. Since the technology is based on the workings of the human brain, this is a useful example. In neural computing a set of data processors, each mimicking some of the data processing characteristics of biological neurons or nerve cells, are linked up. They are not programmed in the conventional manner but learn through data inputs to produce desired results. This cuts down on the time and manpower involved in complex programming as well as solving tasks that are traditionally difficult, inappropriate or impossible – or unpleasant. Some of the applications at Cambridge were: for spot-welding; to explore the characteristics of a bug that causes acne; estimate the lifespan of a building; and detect impurities in olive oil.

Adulterating oil

This last involves uncovering the work of the Mafia which is allegedly heavily into adulterating oil. Any application that involves pattern recognition, forecasting and complex data analysis will benefit from the neural approach. Some of the topics under discussion were: on-line algorithms for prediction and control application, neural control of an experimental batch distillation column and Kohonen networks in machine health monitoring. Much of this is incomprehensible to the layman, which is perhaps why the science is so little known. Enter the UK Department of Trade and Industry and a UKP5.75 neural computing technology transfer programme which aims to encourage take up over the next three years (CI No 2,204). The Department originally called a meeting in January 1990 to thrash out with academics and industrialist what neural computing could mean, and then the its technology adviser, the Information Technology Advisory Board, recommended the programme. Unsurprisingly, US and Japanese companies are ahead of Europe in developing the technology. Ray Browne, the Department’s programme co-ordinator, said the geeneral view is: We (the UK) are a year or two behind in implementing the technology, but on a par with understanding it… We may even be ahead in research. Others disagree, and say the Japanese are more like 10 years ahead. The Trade & Industry initiative has been welcomed by developers who are coming up with a wealth of applications which have not had the take-up they deserve. AEA Technology, the trading name of the UK Nuclear Energy Authority, has had a six-strong team in its applied neural computing centre since 1987. For three years it was part of an Esprit project named ANNIE, the Applications of Neural Networks in Industry in Europe. This looked at the use of neural networks in airline scheduling systems and the classifications of defects. It is now trying hard to market Countermatch, a bio-metric program for signature verification aimed at security applications and reduction of credit card fraud.

By Kate Potter

Running on a standard personal computer, the novel algorithm maps the test signature against a standard derived from an enrolment procedure. Temporal and spatial features are drawn from this mapping and used to classify the signature as authentic or forged. It is written in C so can be used on MS-DOS and Unix systems, and can be used as an on-line or off-line system – the protoype signature can be encoded onto a user’s card or held on a central database. But AEA is having trouble selling it. Andy Lewcock, business development manager, said he is finding customers do not believe bio-metric products work, so he now wants to team up with a systems house to make some sales. Incidentally, Lewcock is no accountant, but a physics graduate, and he believes that’s vital in making high tech businesses run well. Neural networks are not the an

swer to every computing problem, but they are excellent at clearing the way for a detailed analysis. When the Pentagon needed to study the differences between the sounds of a porpoise and a submarine, it could have spent nine months reaching its goal with conventional methods. Instead, it built a neural network and did the study in three days. And the technology can handle jobs that intimidate current systems. The UK Home Office is working with an unnamed software house to develop a neural system to match fingerprints taken by police across the UK. Most forces do this independently, but there is currently no nationwide collation of fingerprints. The British Textile Technology Group got into neural nets (as they are called among afficionados) four years ago, when it examined the feasibility of using them in fabric construction. It can now take a fabric with particular qualities and determine what is needed to produce it at the required price, or how to unite attributes like the desired strength, abrasion and length in one fabric. Software engineer Richard Cooke said the Group is about to enter a European project with fellow groups in France and Italy, but this will not be finalised for several months. The Group has also come up with a sub-UKP1,000 alternative to the spectrophotometer, an instrument that calculates and monitors colours by measuring the intensity of each colour or wavelength present in an optical spectrum. The machine costs around UKP20,000, and the Group’s low-cost solution uses cheap photo-sensors to make calculations from the voltage measured. Some of the other groups involved in neural nets are the computation department at the University of Manchester Institute of Science & Technology which is analysing the recurrence of breast cancer. Scientist Michael Turega said: We’re getting results, but we’re not about to publish them. London Underground hopes to keep its platforms clear with the aid of a surveillance system based on neural networks that detects overcrowding. Cynics feeling the heat might say the system should be focussed on air-conditioning instead. Neural computing may not please everyone, as when it is used to establish personal credit ratings. But it may speed up the process of making speech recognition a valid technology. Tony Robinson, a researcher at Cambridge University, believes it will be three years before a speech recognition card is available for the personal computer, something he is working on now: But there’s still masses to be done.

Speech recognition

However, Dragon Systems Inc plans to unveil a new speech recognition unit based on neural networks, to the Neural Computing Applications Forum’s next meeting in September. Whether neural nets will gain prominence through the success of an individual unit, or via the long term effects of the UK initiative remains to be seen. But is the Department setting the example of using neural computing in its own offices? Dr Robert Wiggins of the information and manufacturing technologies division, had to admit his department had not quite got around to it: Well, no. I don’t really know why. Perhaps it’s something to do with the cost.