Neural networks - the power of fuzzy thinking

The predictive qualities of neural networks are in ever increasing demand in high risk operations the world over. We take a look at the technology behind the fuzzy logic.

It might seem that the last thing anyone needs when things get complicated is fuzzy thinking. However, this is exactly the kind of condition in which neural networks, with their inherent fuzzy logic, start to shine.

We are going to see a lot more of these 'black box' plug-ins in the coming year or so. They are already turning up as fraud detection devices, churn spotters, pattern predictors, network traffic analysers, click stream trend detectors and in various other roles where people or businesses need to detect trends in patterns of data.

The term 'neural network' is something of a metaphor, playing on the way that neurons in the brain have multiple links with each other. Neural networks are often called 'self learning' programs.

Where conventional programs require someone to pre-analyse a problem and code the solution line by line, neural networks can sift meaningful patterns from seemingly undifferentiated streams of data.

Part of their usefulness is that they can produce approximate solutions from incomplete data sets, much like human 'intuition'.

Training neural networks
The idea behind neural computing goes back to the 1940s. But the whole concept got steamrollered by the success of conventional computing during the 1970s and 1980s.

New ideas in the field and advances in hardware, however, put neural computing back on the networking industry's agenda.

The industry also received some substantial UK government assistance. In 1993, the Department of Trade and Industry launched the Neural Computing Technology Transfer programme (NCTT), with £5.75m worth of funding. It was intended to speed up the adoption of neural computing in the UK.

There are several different kinds of neural networks, each of which is suited to a different task. The simplest type of neural computers arrange processing units called 'neurons' in layers. There is an input layer, an output layer and one or more middle strata called 'hidden layers'.

The NCTT has an excellent introductory paper on the subject available at www.brainstorm.co.uk/nctt.

Its account of the fundamentals goes something like this: each neuron within a neural network takes one or more inputs and produces an output. At each neuron, every input has an associated 'weight', which modifies the strength of each input connected to that neuron.

A neural network is 'trained' by presenting it with data sets and then manually adjusting the weights associated with each neuron to bring the output of the whole network closer to the desired result. The training phase ends when the network is deemed to provide a reasonably accurate set of answers to all the training problems.

The way in which the neurons are organised is called the network topology. The number of neurons in a neural network can range from 10 to thousands.

The most popular topology consists of three layers of neurons in which the output of each neuron is connected to all the neurons in the next layer (known as a feed forward network). This flows into the network through the input layer, passes through one or more intermediate 'hidden' layers and flows out of the network through the output layer.

Some networks allow backward as well as forward connections. Others allow connections between layers, but not between neurons in the same layer. Some networks even allow a neuron to be connected back to itself.

Martin Saunders, enterprise management product manager for northern Europe at Computer Associates (CA), claims that the company sees massive opportunities for predictive technology in the next few years.

Neugents' predictive potential
CA has been using what it calls 'neugents' (an abbreviation of neural agents) in its framework management product for the last two years.

"The problem we had, which neugents allowed us to address, was that our management tools in TNG Unicentre were good at saying what was happening at any instant in time, but they were not particularly predictive," said Saunders.

The tools could draw an IT manager's attention to a particular server running out of disk space. What they couldn't do was identify the pattern of events symptomatic of problems that would occur later on. With the advent of ebusiness and the additional focus on round-the-clock availability, it became more vital for framework management tools to gain this predictive capacity.

As Saunders notes, the performance matrix in a Windows 2000 server records some 1000 separate items. There is far too much going on there, far too many hidden correlations for a human brain to map.

However, this kind of data-driven matrix is meat and drink to a neugent, particularly when it has been trained to spot exceptions to standard performance patterns. When you load a neugent onto a server it just sits there monitoring the system.

After this 'reflective' period, the neugent will have observed that, of the 1000 items in the performance matrix, 100, say, did not change, so they can be ignored.

It will recognise that tens of other indicators all changed in the same way, and a small percentage of the indicators appeared critical as predictive indicators.

A typical management tool says: "You tell me what constitutes an unusual event, and I will alert you if I see it."

The problem with this approach is that you often don't know in advance what the determining event will be. An infinite number of things might go wrong and describing them all, in sufficient detail and up front, is not feasible.

What makes neugents useful is their ability to produce a statistical view of potential outcomes. They can flag a situation as 80 per cent likely to produce a catastrophic outcome for a server in an hour's time.

"After we produced our first neugents, our clients became interested in their predictive potential for other uses. Neugents are now being asked to look for profitable patterns in business data and in bank cheque fraud," said Saunders.

One client uses CA's neugent technology to spot 'cheque kiting', for example. This is where a fraudster opens five bank accounts, deposits small sums in four and a larger sum in one and then moves the money rapidly between all five accounts to simulate a profitable business.

This false record is then used to open up lines of credit at all five banks, which the fraudster then cashes in before boarding a plane to distant parts.

To human eyes, the transactions involved in cheque kiting are very hard to spot, because they are buried in inter-bank transactions and masked by large numbers of legitimate transactions.

However, a neural network can identify and highlight the spurious nature of the pattern associated with cheque kiting relatively easily.

Malcolm Britto, a consultant with Nortel Networks, points out that Nortel uses neural networks as one of a number of tools to add intelligence to the network and to examine patterns of content traffic movements from internet service providers and users.

One of Nortel's newer technologies, stemming from its acquisition of Alteon, is routing based on ISA level four or level seven criteria. "Intelligent traffic management is carried out by a number of vendors, but we are the first to do content switching," said Britto.

Trying to allocate traffic flows equitably, based on IP address information alone, with no notion of content, is near impossible. Opening up the packets so the network can make decisions based on content-level data, using layers four to seven of the OSI stack, makes a lot of sense.

Content switching algorithms can look at the patterns of data and sift traffic to see which requests, for example, are really requests for local information that do not need to be fulfilled by being routed via internet servers all over the world.

This, combined with intelligent cacheing policies, can speed up response times massively, providing users with a better service. One UK company that has specialised in applying neural networks to a large range of problems is Hampshire-based Neural Technology.

Dr George Bolt, the company's director of product innovation, has written and lectured extensively on the subject. He likes to emphasise that neural agents and networks should not be thought of as mysterious 'black boxes', but rather as statistical devices no different to any other set of mathematical equations.

"A neural network is just a complex, repeated, mathematical equation that works on numbers," he explained. Neural technology can be loaded as software onto standard PCs, and can be integrated as a component inside other, conventional programs.

In fact, Bolt and others in the neural network game recognise that it is in their interests to downplay and dispel the idea that the technology is a 'black box'.

Incorporating the technology
One of the factors that has slowed down the uptake of neural network technology is corporates' unhappiness at not being able to see the workings behind the outputs from neural networks.

The reasoning here is that if you don't know what has led a particular device to generate a specific conclusion, it is hard to know what value to put on that conclusion, particularly with self-training neural networks. In other words, it can be far from obvious why a neural network is issuing this or that 'oracular' prediction or statement.

Of course, if the network's output proves right more often than not, you can simply disregard the opacity surrounding its workings and act on the outputs.

However, there is something about this that smacks of flying blind, and it tends to leave clients uneasy. For this reason, Neural Technology likes to insist that it can interrogate its agents and uncover the sequences behind their decision processes.

One of Neural Technologies' more successful products is its Decider risk management and weighted credit score-card application.

Gordon McFadyen, fraud prevention manager in the Bank of Scotland's risk management department, was one of the earliest users of the system, and he explains how the bank broadened the use of the system to include fraud detection.

Processing capacity
"When Neural Technologies came to us, it was initially demonstrating Decider's capabilities as a credit score-card system. However, at the time we believed we had a problem with application fraud; people setting out to obtain credit with no intention of making any repayments. We asked the company if it could build a neural network that could detect these kinds of applications with a reasonable level of confidence," he said.

Neural Technologies came up with a version of its neural network technology that was able to identify the patterns associated with this kind of fraud, and the Bank of Scotland became a neural networks user.

Since then, according to McFadyen, the bank has deployed several such systems in various departments and areas. "There is no one tool in banking that is going to be the silver bullet for all problems, but used in conjunction with other processes and procedures, neural networks have proved their worth," he said.

One of the problems with putting a precise value on the savings generated by the Decider product in his department, McFayden points out, has to do with its role as long stop or sweeper (to mix sporting metaphors).

The bank uses the product as a last line of defence to catch fraudulent applications that have not been picked up by all the other processes and procedures the department has in place.

Therefore, the system only gets credited for what it actually finds, not for the fraud it could have detected had the other systems not been in place.

Bolt points out that what neural networks bring to the party over and above ordinary statistical analysis tools is the capacity to process huge volumes of data.

Most statistical approaches fail (meaning they take impractical amounts of time) when it comes to performing accurate modelling where there are a high number of complex, non-linear interactions among input variables.

This leads to short cuts being taken, Bolt explains. A database with 30 variables will end up being cut down to, say, just 10. By contrast, neural techniques use all the available data to produce their predictive models.