Manchester-based Recognition Research Ltd, a company specialising in neural network computing, has been set up by Paul Gregory and David Bounds, two experts in the area of pattern recognition and neural computing that now believe the technology has advanced sufficiently over the past two years and that there is an emerging market for neural networks for commercial applications. In recent years information, often large quantities, has come to be regarded as a vital strategic resource for companies, but only if it can be understood and acted upon. Humans are very good at extracting patterns from apparent chaos, tasks such as speech recognition, identifying a face in a crowd or spotting a new trend in the market.

Apparent chaos

These are very difficult, if not often impossible for conventional computing techniques. Neural computing offers techniques that enable computers to be trained to recognise patterns and to extract some order from apparent chaos. Whereas a conventional computer uses instructions and data with between one and 20 processors, neurocomputing mimics networks of biological neurons, processing information either in digital or analogue form with up to 1m processors. Both use a transistor rather than a neuron. General purpose neurocomputers can be further subdivided into commercial co-processors that can plug into the back of a personal computer, or interface to a Sun workstation or DEC VAX, and parallel processor arrays of cellular arrays composed of a large number of processing units such as simple RISCs with possibly a small amount of memory. Neural networks are the programs for neurocomputers. Each ‘program’ is made up of interconnected processing elements. Each element is associated with a state, either on or off, and each connection is associated with a weight. A neural network that learns patterns does so by adjusting the connection weights and through these adjustments the network exhibits properties of generalisation and classification (CI No 1,337). For example, an an expert system, a decision learning system can adapt automatically to changing risk environments, eliminating costly updating of static expert systems. As a predictive modelling tool it can also learn to recognise complex patterns involving many interdependant variables, which are too difficult to process using conventional statistics. Applications group into commercial such as loan application scoring, airline fare optimisation, gas and petroleum exploration, and industrial, including manufacturing product inspection, process control and fault detection. Military applications include radar and sonar target detection, image object detection and recognition and applications in electronic warfare. The UK has been slow to implement this revolutionary technology but the US is already using neural computing on a commercial basis. Many major US financial institutions such as the Citibank, Chase Manhattan and Chemical banks use neural networks successfully for credit card applications. Typical credit card applications include credit assessment for initial authorisation, setting limits and authorising actions as well as detecting fraudulent transactions and predicting charge-offs. Providence, Rhode Island-based Nestor Inc has developed a mortgage underwriting mimic trained on the decisions of good underwriter and a system to predict loan performance based on the actual outcomes of specific loans. So far bad debts have been reduced by 7%. –

By Elvadia Tolputt

Other risk-assessment problems include insurance. Many Swiss banks held a conference last year to understand the concept and application of trained networks, but it has not been revealed how far they have taken it. Paul Gregory believes that the same type of credit card fraud-control system that is used in the US may be installed in the UK sometime this year. Five years ago when neural computing was a speculative subject large British companies, such as British Aerospace and British Telecom, did invest large sums of money into research. At present the majority of the companies involved

in neural network research and applications, including Thomson-CSF, Philips, Siemens, NEC Corp, British Aerospace and British Telecom, are being very secretive about future commercial implementation. In spite of this, a large number of corporates spoke about their research at the Third European Seminar on Neural Computing last week. Gregory and Bounds have set up the company along the lines of the large number of smaller US companies that start up with a good idea and work like mad to develop it into a marketable product; they normally either go bust or make a million. The big companies do have a lot going for them with large resources and research facilities and there are limitations on smaller companies but as they are often in a niche with very specialised knowledge, Bounds believes that smaller companies often end up with the best technology. Believing that training is the first step in applying neural computing to any application, six months ago Recognition Research started to introduce the concept of the technology by training courses. These range from introductory to advanced and include hands on experience using NeuralWorks Professional II, a neural computing simulation package. Delegates have come from many major organisations including Racal, ICL and Royal Insurance.

NeuralWare

In January 1990 the company was appointed distributor for the UK and France for Pittsburgh, US-based NeuralWare products which are being used to develop and implement a number of real-world applications, including national defence, space exploration and medicine with over 3,000 installations worldwide. Recognition provides consultancy for companies who want to implement these products within companies, either to advise on opportunities and to provide a full or part solution. Last week, the company launched an accelerator card, the AL860, which will reduce the ammount of time for these applications on 80386 and 80486 AT-alikes from hours to minutes making this technology much more feasible. The card comes with 4Mb, 8Mb or 16Mb of cached DRAM, four way ported to Intel’s 80860 RISC processor, and plug in daughter cards allow for memory expansion upto 64Mb. Combining integer processing at 33 MIPS with single precision floating point computation at 66 MFLOPS, it will bring supercomputer performance to the desktop, Paul Gregory believes, and will lead to a whole range of data analysis tools that could only be implemented previously on large mainframes. A neural network application, written in C, ran 30 times faster on the AL860 than on a standard 80386 personal computer. AL860 can also be used for compute-intensive applications in simulation and finite element analysis, image, speech and signal processing and financial analysis. Recognition Research is set up along the lines of new wave US companies, employees work from their own homes using phone, facsimile and computer links. It seems a crude method but it has worked well for many companies in the US. Most of the specialists are employed on a contractual basis because the company sees this as the most efficient way to apply the talents of these highly skilled people.