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 covariance matrix, is experimentally investigated. When PCs are employed, fewer experiments are needed and the network configuration can be set precisely without trial-and-error experimentation. Tested on five real-world databases, it is shown that very few PCs are required to achieve a shorter training period, lower mapping error and higher classification accuracy, compared with those based on random initialisation.