By now I’m sure most of you with an interest in the computer industry will have heard about neural networks. The UK Department of Trade & Industry has been plugging them with a series of workshops as part of its three-year Neural Computing Technology Transfer Programme to encourage British firms to investigate how this technology can be applied to their businesses. So are you really going to apply neural networks to solve those business problems and improve efficiency? The Department’s approach has been to focus on case studies as the means to convince us of their industrial and not just academic worth.
Bizarre
The range of applications is broad and in some cases bizarre. At the Neural Computing Applications Forum two-day conference in Oxford last week for academics and industrialists, applications were being discussed that are currently under trial. These included an alarm monitoring system developed by a small firm, Neural Solutions Ltd, for hospital intensive care units to cut the number of false alarms about the severity of a patient’s condition that currently occur; an application for monitoring the health of machinery at the Blyth Power Station located north of Newcastle-upon-Tyne and a method to help solve the literary mystery as to whether William Shakespeare or John Fletcher wrote the play Two Noble Kinsmen or whether it was a collaboration between the two writers. A representative of the Department of Trade & Industry attending the conference raised the point that case studies in the UK are shown to be a particularly effective way of conveying the message, with 1,000 firms currently in the process of investigating the possible use of neural networks. However, case studies are all well and good, but managers have all got their own kinds of problem to solve with their own kinds of data. Case studies on the applications of the networks are just the icing on the cake. There is a whole string of work to be done before you get anywhere near using neural network yourself. You have to establish your training data, train the network, make sure you have a full understanding of the data you are using and then once you have run the network and got an answer you have to go back and check that this answer really makes sense. Case studies seem to add to the mystique of neural networks perhaps running the risk of people seeing them as magic computers that can think for themselves without emphasising the immense amount of statistical work that has to go on both before and after the network has provided some answers. Neural networks then are more of a step in a whole string of work that needs to be done to get useful answers.
By Abigail Waraker
From general discussion at the conference on the key factors in the application of the networks the issues that emerged focused firstly on how to convince people that neural networks really do work and can be a useful business tool and secondly the importance that the pre-processing carried out before and after the neural network is used is accurate. There was much greater focus at this conference than the one held six months ago on the importance of the pre-processing of data and also the idea that once the results have been obtained it is a good idea to go back and check them with alternative methods. No matter how good the software or neural network is, the prior knowledge held about the information could be wrong. An example given was a robot that was being trained to avoid bumping into objects. Different people took it in turns to guide it around the obstacles as training runs. But once the robot was trained and set to navigate the obstacles alone it consistently bumped into the obstacles even though it had never bumped anything in training. Looking into the problem, it was found that one of the trainers had always guided the robot to the left of an obstacle and the other consistently to the right, so the neural network made the logical conclusion and went straight ahead! This clearly shows the importance of fully understanding what information your test data holds. So the emph
asis now seems to be that it is important to explain to people exactly how the network works in a more intuitive manner so they feel they understand what is going on, rather than putting too much trust in an apparently magic solution demonstrated by case studies. The Department of Trade & Industry’s awareness programme also consists of workshops held at regular intervals throughout the country to help business managers understand exactly what is involved in using neural networks to gain a business advantage. Lionel Tarassenko of Oxford University raised the point that you need to be sure you have enough data, as insufficient data could lead to rather skewed results. Huge amounts are needed and as a rule of thumb this is as high as 10 times the amount as there are weights in the network for a single layer network.
Dangers
In addition, the time spent checking and rechecking results and even testing alternative measures all adds to the time and expense for a firm that is implementing these networks, for what might only amount to a small percentage increase in accuracy on the original result from the network. And then is the business manager going to want to put in the necessary resources if he can’t be sure he will get useful results in good time at the right price. Steve Roberts of Oxford University suggested that whether neural networks survive industrially will depend on whether those firms that have taken them up make returns on them in the next five years, which the where the Department sees its case studies fitting into the process of education. So it appears that if there is to be a future for neural networks then we need to get away from the idea that not a case of just getting some data, running the program and taking home your result. The key seems to be to understand what you are doing with the network and to be aware of the dangers of hidden data. Much as it creates great excitement when people talk about computers mimicking the processes of the human brain, this is not necessarily exactly what we want. Steve Roberts argued that neural networks are useless unless we go in search of honest networks that tell you if their results are less than good, otherwise we run the risk of making them too much like the failings of the human brain, in which if we make mistakes are inclined to cover it up and quickly make right the situation rather than openly admitting what we have done wrong.
