If you are baking a cake and you turn the oven up to twice the temperature the recipe recommends, it doesn’t take half as long to bake. So where does this tip fit in with neural computing? Neural networks look for complex patterns in large amounts of data. They carry out non-linear modelling. The time-to-temperature relationship of baking a cake is a non-linear relationship. The temperature increase does not respond directly to the time to cook. Neural networks are commonly described as computers that mimick the processes of the human brain. But as neural computing techniques become more prevalent in mainstream computing, less mystical definitions of how they work are more informative. Essentially they are techniques for non-linear data analysis and visualisation consisting of simple processing elements organised to enable the synthesis of numerical models.
Cacophony
Neural computing has become more accepted since the late 1980s. People have started looking to buy neural-based tools and services. You don’t need a PhD in rocket science. People are using them because they get good results in their business, says Alan Hall, business manager at Scientific Computers Ltd. The Crawley, West Sussex-based company provides both a consultancy service for people that want to use neural networks, and supplies software. Redifon Technology Ltd of Leatherhead, Surrey has used Scientific Computers’ NeuralWorks Professional II/Plus package to develop a system for the emergency services that cuts out the background radio interference so that the voices can be heard clearly. Operators can now listen to the radio transmissions for longer. Previously the background cacophony was so exhausting that the operators could only listen for two-hour periods. Now there is silence except when the network detects a human voice. The network was trained with samples of transmissions where human speech was and was not present. Neural networks are trained using historical data. Data the company already understands. So when the network produces an error, that error is used to adjust the model and change the coefficents to get the right answer. In a feed forward or multilayer perceptron type of network, this process is called back propagation. A lot of effort in building a successful neural network is in data analysis. Making sure you haven’t got awful examples in the first place is important as these will result in a bad output from the network. So if for example raw data on people’s salaries is presented to the network, in a scale from earnings of ú8,000 to ú250,000 per year, the majority of people will be bunched together at the lower end. If the network is calculating on a scale of 0 to 1 most data will be around 0.05 to 0.06. This is a very narrow segment of the scale which make it more difficult for the network to make accurate analysis.
By Abigail Waraker
So pre-processing needs to be carried out. In this case a logarithm transform of the data makes the task easier for the network. The criticism often levelled at neural networks is that people cannot understand how the results they throw out are reached. This is particularly common with the feed forward type of network: the type the UK Department of Trade & Industry estimates is used in 95% of neural-based applications. A feed forward network is generally made up of three layers. An input and output layer and a middle or hidden layer. Data is fed in to the input layer and the network assesses weights given to the importance of various factors in that data, then produces the result at the output layer. Scientific Computers is using another type of network, the Kohonen network, to offer insight into how a network derives its answers. This is done by inputting data that a feed forward network would offer as an output. So in the case of a network to assess credit risk, the anticipated likelihood of someone being a good, bad or medium risk would used be as an input. This likelihood is assessed using credit scoring methods curently used by companies lending money. Th
e Kohonen network is forcibly skewed with input data to see what common features the people have, say in the poor credit sector, that have not previously been noted. The aim of this is not to prejudice the results against certain people before the network has even had a chance to run, but to spot new factors linking people that have similar credit histories, so marketing can be efficiently targetted. When the Kohonen network organises itself, it extracts prototypes, with people that are of similar risk being clustered together on the visual representation created by NeuralWorks Professional II/Plus. There are separate maps for those who are good, medium and bad risks. The user can then look at the clusters and discover new features that link people within one of the three categories. In the Scientific Computers demonstration, good credit risk maps look like good payment maps, so people that pay their bills on time can be seen as good credit risks. The correlation is straightforward. But it is less easy to spot intermediate credit risk people. The Kohonen model can help you determine why these people are in the medium risk category, by profiling one kind of intermediate credit risk person and then another and comparing the input data on them to spot similarities. In early February, Scientific Computers will release Predict, a ú4,500 do-it-yourself neural network package that creates the neural network for you. Scientific Computers says it is a complementary product to NeuralWorks Professional II/Plus.
Neural novice
Predict takes the user through the stages necessary to create his own network. That is, it automatically carries out data analysis, then scales and transforms the data. It performs automatic training and can select which data it considers useful to train the network. Once it has built a prototype model, genetic algorithms test it to discover how good it is. And finally it builds a feed forward network. The product will help the neural novice take care of the stages necessary to build and deploy a neural network. After the network has come up with its answer the more advanced user can go through and find which fields of data Predict has used and which mathematical transforms it has carried out on that data. It also shows which data has not been us ed and at what stage a particular category of data was thrown out. Predict can sit within the Excel spreadsheet as a pull-down menu or run from a script file. The company also wants to apply it to the Applix spreadsheet to win the Unix market. Whether it’s a piece of cake to use remains to be seen.
