capabilities of neural networks

Neural networks have the ability to map data clusters and data relationships from huge volumes of data. They can identify subtle relationships in complex data masses to reveal valuable trends, relationships and associations from which we can gain an improved understanding of the problems we face in disciplines as widely varied as medicine and commerce. Such knowledge can revolutionise the way we go about our business.

Neural networks can give predictions for results that we cannot hypothesise upon. For example, suppose we wanted to pre-select applicants for training in a particular skill. We would first produce a list of many people who had been successful in that skill. We would then consider what data relating to them might be relevant or loosely relevant in contributing to their success e.g. lifestyle, lineage, attitudes, academic ability, IQ, physical characteristics, and any other data that we could get our hands on and which might be a determining factor. Note that neural networks are tolerant of vague data, and data that cannot be quantified (one network describes colours as ‘bluish. Pinkish etc.). We would then train a neural network to associate the data with the levels of success achieved by the subjects. If trained successfully, the network would indicate which of the factors were the more significant, allowing the dataset to be pruned to eliminate low contributors. After refinement followed by extensive testing and validation the neural network would be able to predict successful candidates from a knowledge of the relevant data associated with them.

The neural network can then not only predict the likelihood of success for each candidate, it can also be used for ‘what/if’ analysis where inputs can be changed slightly to determine the effect on the prediction. This facility is used to good advantage in a number of neural networks used in the field of medicine. In summary:

• Well trained neural networks can make predictions even when we cannot.
• Neural networks can identify which data is most relevant to their success, so that low impact data fields can be dispensed with.
• Neural networks are tolerant of vague data and data that cannot be quantified easily.
• The ‘what/if facility’ in back propagation neural networks is a quality analysis tool.

Many neural network applications are already in use to help HR managers in their work. These applications include, pre-employment screening for particular work or posts and selection for specific work or promotion – see our example promotion prospects application.

If you wish to predict something, but cannot quantify precisely what, but have or can collect lots of data that might be useful in reaching your desired objective (even if some of that data is vague), then consider a neural network solution.