On the initialisation of Sammon's nonlinear mapping

The initialisation of a neural network implementation of Sammon's mapping, either randomly or based on the principal components (PCs) of the sample co variance matrix, is experimentally investigated. When PCs are employed, few er experiments are needed and the network configuration can be set precisel y without trial-and-error experimentation. Tested on five real-world databa ses, it is shown chat very few PCs are required to achieve a shorter traini ng period, lower mapping error and higher classification accuracy, compared with those based on random initialisation.