Just as a child can learn to identify a bus as being both a green two decker vehicle full of people and an empty red single decker vehicle, so can neural computing. So how can a computer think like a human brain? Well, it doesn’t think as such, but like a child, these computers learn by example. Mimicking the biological processes of the brain, a neural network can generalise from experience, and unlike conventional computing you don’t have to feed it exact and complete data. It is capable of coping with previously unspecified situations. Instead of working to precise algorithms and equations, neurals can be trained. A neural network is a series of neurons, or elementary processing units, connected together in layers. Each neuron takes input either from the external world or from the output of other neurons. Output signals from each neuron propagate their effect across the whole network to the last layer where the results are output to the real world.
The relations between each neuron are set during the training of the network, in which numerous examples are presented to the network, and each neuron weights the recurring characteristics of a specific aspect to establish a pattern for future recognition, in the same way that the child learns to recognise different types of buses. So, what role does this intriguing and relatively unexplored realm of computing have to offer the practical business user? At present most neural technology applications are concentrated in the military and defence sector or are being run as research projects by educational establishments. Although this trend is changing and the highest new take-up of the technology is in the industrial and manufacturing industries. According to recent research by the UK Department of Trade & Industry, almost 70% of the UK’s leading blue-chip companies are either investigating the potential of neural computing or are actually developing neural applications. Amongst them are Powergen Plc, Boots Plc and British Gas Plc; although as ever Japan and the US are some years ahead in development. The Department is now one year into its Learning Solutions Awareness Campaign and the other day hosted a series of neural computing workshops for managers and software engineers to explain what the technology is and to demonstrate how it works. Obviously the clear advantage of neural computing lies in its ability to cope with incomplete information, often a frustrating aspect of conventional systems. The majority of potential applications lie in pattern recognition and complex data analysis and prediction. Neural Technologies Ltd of Petersfield, Hampshire was demonstrating PictureNet, a picture recognition system, to show how neural computing works. A pre-processor converts a simple hand-drawn picture into black only and white only pixels and then forms the whole object into a 16 by 16 grid of black and white pixels; 256 input neurons feed into 80 middle layer neurons and onto 25 output neurons, which can identify up to 25 objects. Running on an 80486 processor, it can recognise an object every 0.03 seconds. Post processing involves selecting the highest value output as the most likely name of the drawing.
By Abigail Waraker
Redifon Technology Ltd of Leatherhead, Surrey was demonstrating its Speech Detection System. The company has combined neural computing and signal processing techniques with audio pre-processing to create a system that can detect speech in noisy environments. The demodulated audio output from high frequency radio receivers is often plagued by unwanted noise, making listening for incoming calls very tiring. A range of pre-processing algorithms were developed to identify significant voice features and a neural network is then used to combine the algorithms for improved speech detection. A squelch device can then be used to cut out undesired background noise so only the speech can be heard. This device would have an obvious use for the Air Force, for example, or for Morse or music detection. Advanced Recognition Ltd of Windsor, Berkshire has
applied neural network technology to cheque processing. Instead of manual keying of each amount, which is labour-intensive and time-consuming, the Courtesy Amount Recognition device is trained to recognise the figures written on a cheque and automatically encode the amount so it can be debited from the payer’s account. Only those amounts it cannot read then have to be keyed by hand. At this stage a significant percentage of cheques are not recognised and have to be keyed manually, but nevertheless keying time is still saved. Technology that can recognise the written amount and then compare that with the amount in figures is still about 10 years away. Boots of Nottingham is using a neural computing system for strategic forecasting to aid future investment decisions. The company found distinct advantage in the processing speed and ease of use. Unlike regression packages that produce statistics that the user subsequently has to interpret, the neural network can produce results immediately, saving considerable analysis and programming time.
The company is using the system to determine the relationships in rents, locations and performance in new areas. British Gas is using neural technology to forecast gas demand to improve efficient running of the national gas pipeline, again to save money. There is potential here for companies that need to predict demand for products with a limited shelf life, an advantage being that the systems are faster because they consist of interconnected processing units operating simultaneously on the same problem, which means that they have the potential to operate at considerable speeds. The Driving & Vehicle Licensing Agency in Swansea is monitoring budgetary and manpower planning based on the large number of transactions at peak times, and has managed to reach an acceptable turnaround standard required by the government’s Citizen’s Charter initiative to improve public services. Those attending the conference did not appear especially convinced by the theory. However, the demonstrations brought the theory to life and the practical uses were impressive. The government wants to demonstrate areas where the potential user may develop such a network. In addition to its own efforts it is promoting awareness of neural computing groups and clubs. The British Neural Network Society at the University of Portsmouth runs conferences and workshops as does the Neural Computing Applications Forum in Malvern, Worcestershire and the International Neural Network Society in Liverpool. The Department of Trade & Industry are also offering some financial assistance to firms wanting to develop neural systems under its Enterprise Initiative, depending on a company’s eligibility.