Neural computing

Citation

Rudall, B.H. (1998), "Neural computing", Kybernetes, Vol. 27 No. 5. https://doi.org/10.1108/k.1998.06727eaa.006

Neural computing

Neural computing

Advances in neural computing

Neural nets and the subsequent developments in neural computing are very much the concern of those who are engaged in the study and development of cybernetics and systems. Cybernetics literature was full of references to developments in neural nets and networks many decades before the current interest. Most references, however, lacked examples of real-life applications, while today we see numerous publications that outline new and exciting uses in almost all areas of endeavour. In the UK, for example the government, through its research agencies and councils has encouraged developments and applications in neural computing. An analysis of the Engineering and Physical Sciences Research Council's (EPSRC) funded programme into neural computing research is currently being carried out by an expert panel of academics and industrialists. The aim of this exercise is to assess the status and balance of the current portfolio and to identify potential research opportunities in neural computing. The EPSRC programme coordinator for "Neural Computing ­ The Key Questions (NCTKQ)" is David Bounds. Tel: +44 (0) 121 359 3611.

Industrial potential in neural computing

Neural computing methods are now being applied to many important new industrial applications. A recent report gives details, for example, of the work in the UK, at Oxford University. The university has been working closely with Rolls-Royce and Oxford Instruments to develop novelty detection capabilities which will utilise the latest neural computing techniques. Novelty detection, the report says, is important in a wide range of pattern recognition problems in engineering and medicine, both of which are areas interested in identifying abnormal cases. Such cases are under represented in databases of available examples and this makes it impossible to train neural network classifiers. The solution developed by the Neural Network Research Group in Oxford University's Engineering Science Department is based on a two-step approach. Professor Tarassenko, who has led the research, explains that:

First, we learn a probabilistic description of normality using a statistical model. Then we identify abnormalities by testing for novelty against the description of normality.

It is claimed by the group that models used in this process can detect abnormalities without being trained on abnormal patterns. This is the example of an initiative that has been funded by the EPSRC's Neural Computing ­ The Key Questions (NCTKQ) programme, referred to in the previous section, and is one that has already shown great industrial potential. It has, for example, been successfully demonstrated at Rolls-Royce for over a year where the novelty-detection and data analysis software developed by the university has been in use. Currently three packages are being used to extract diagnostic information from jet engine vibration data.

At the Rolls-Royce Applied Science Laboratory, Dr Peter Cowley comments that:

We had discovered the potential usefulness of novelty detection before encountering Professor Tarassenko's work. Oxford's research has now added rigour to our own R&D efforts, which has brought significant improvements to the robustness we can achieve.

It is now expected that a major R&D programme at Rolls-Royce will take forward the University's novelty detection approach into mainstream aerospace commercial operations. The University is also working, we are told, with the Rolls-Royce industrial businesses on a separate novelty detection initiative on monitoring industrial gas turbines.

The other important application where the identification of abnormal cases is required is in the medicine area. The Medical Systems Division of Oxford Instruments has collaborated in the EPSRC Neural Computing project to extend previous long-term work with Professor Tarassenko's team. This has resulted in Oxford Medical's Questar Sleep Analysis System, which can monitor people suffering from sleep disorders by analysing electro-encephalograms (EEGs). This collaboration was focused on the analysis of the EEGs of people with epilepsy so that "inter-ictal" events could be identified. These occur between seizures, but are difficult to characterise. Oxford Medical's Business Development Director, Pauline Hobday says that:

Its results have encouraged us to develop plans to incorporate novelty detection into a variety of future products. Working with the University is an invaluable complement to our internal R&D activities. The partnership introduces us to broader realms at the forefront of advanced research, enabling the company to maintain its competitive edge in key medical technologies.

There is no doubt that these examples of the application of neural computing are but some of the many exciting new developments in the use of neural nets worldwide. The industrial potential is obviously widening for neural network applications and Oxford University's neural network novelty-detection approach is just one excellent foretaste of what is to come.

B.H. RudallNorbert Wiener Institute and University of Wales, UK