Data Representation And Generalisation In An Application Of a Feed-Forward Neural Net

This paper examines the role of a feed-forward neural network in an application involving the segmentation of medical images. Particular emphasis is focussed on the the choice of input to the network in terms of both its representation and content. It is shown that pre-processing of the input information by a statistical classifier leads to significant improvement in the networks performance. An examination is also made of the networks ability to generalise. In particular the importance of the use of validation data during training is demonstrated.