Computers that act like the human brain, by learning and thinking in the same way, are being looked into by the Research Initiative in Pattern Recognition at the Royal Signals Research Establishment, Malvern, Worcestershire. The UKP3.6m project was launched three years ago, in 1986, to research into machine vision and neural networks. Carried out jointly […]
Computers that act like the human brain, by learning and thinking in the same way, are being looked into by the Research Initiative in Pattern Recognition at the Royal Signals Research Establishment, Malvern, Worcestershire. The UKP3.6m project was launched three years ago, in 1986, to research into machine vision and neural networks. Carried out jointly by eight companies and the Ministry of Defence it has recently won continued funding by the Department of Trade and Industry for the next two years running to December 1 1991. The project embodies co-located research where industries and people create one centre of resource, with British Telecom, British Aerospace, Pilkington, MEL/Philips, Thorn EMI, Plessey, STL and Smiths Industries each placing one representative research scientist at Malvern. The companies will get exclusive rights to intellectual property generated by the project for two years after the end of the project. This will provide a commercial incentive for the companies involved, even though the actual research is not directly for commercial use. The aim of the research is to teach computers to think without being formally programmed and thus to try to create a computer that can develop its own learning system that mimics the human reasoning process. Machine vision and research into neural network computing have been combined to form research into neural network learning systems. Machine vision covers such things as robots with stereoscopic vision for distinguishing objects. Neural networks are computers that use their own logic to reach solutions rather than being programmed with a set of absolute rules. In a neural network there is a whole array of simple computing elements, each connected to several others by links of varying and initially random strengths. When data is input, initially garbage comes out the other end but by adjusting the strengths of the links, by the use of algorithms, after several iteraTi-oUKP!at reinforce the correct links and weaken the rest, a recognised output can be obtained. The algorithms are formed by using the error obtained when data is processed and the known output is compared with the obtained output. A neural network learning system works by learning by examples, and the learning capabilities of the system can be evaluated by presenting it with unseen examples and noting how well it performs with them. A conventional computer with a CPU and a memory, which reads instructions in the form of a program, processes them and stores the data in memory, is very time-consuming for complex problems compared with what seems to be promised by neural networks. For example, Chase Manhattan Bank in New York has installed a neural network learning system for use with credit card fraud. It uses databases of financial information from various sources including Reuters. Each individual credit card holder develops a particular card-usage pattern and the computer learns the way in which the holder characteristically uses the card, taking into account small variations in usage. If the card is used abnormally, that means outside certain limits of the normal usage pattern, then the neural network, which has learned that user’s idiosyncratic spending patterns by experience, can sense the variation. It would be nearly impossible to program a conventional computer for each particular card.
Diagnosing back problems
Projects carried out at the Royal Signals Research Establishment include a medical diagnostic system. A pilot such system has already been produced, which can distinguish between the four categories of back pain when a patient’s symptoms are input. A doctor followed patients with back problems for two years, training a computer with their medical records. The system proved that neural networks could be trained to give accurate diagnosis and are sometimes better at it than a human doctor in particular areas of medicine. Another neural network at Malvern was taught to read with the Janet & John books. The computer was taught in much the same way as a child would learn, by sta
rting with a few words and gradually building up its vocabularly to approximately 10,000 words. A variety of hardware is used, including workstations and personal computers for smaller networks and Transputer parallel networked systems for larger problems. The next two years will be focussed mainly on further research into neural networks which are believed to be the next step in improving computer intelligence. – Elvadia Tolputt