developing a neural network

The first requirement is lots of data relevant to the issue that the neural network is to address. Generally, more data means a higher quality result. It is normally necessary to use data that already exists, but quite often a network built using existing data can be improved by collecting and using new data that is added over time. In the case of a back propagation, forward projecting network that will make predictions based on input data, a network is designed, built and presented with input and output data that has been modelled into a form that the neural network can interpret.

For example, data can be presented to the network as numeric, symbolic, Boolean etc, depending upon how the network is to interpret it. A percentage of the input data is held back from training, to be used to test the network when training is complete.

The network is trained through many iterations to associate the input data with the output. This process is likely to include significant redesign of the network as the process is refined and as the network ‘learns’ the associations. The trained network is tested using the data held back from training.

The network is trialled to evaluate its performance, and given a user interface.

The result is an application that can predict the outcome from new and previously unseen datasets.

See the example.