ON PATTERN-CLASSIFICATION WITH SAMMONS NONLINEAR MAPPING - AN EXPERIMENTAL-STUDY

Sammon's mapping is conventionally used for exploratory data projectio n, and as such is usually inapplicable for classification. In this pap er we apply a neural network (NN) implementation of Sammon's mapping t o classification by extracting an arbitrary number of projections. The projection map and classification accuracy of the mapping are compare d with those of the auto-associative NN (AANN), multilayer perceptron (MLP) and principal component (PC) feature extractor for chromosome da ta. We demonstrate that chromosome classification based on Sammon's (u nsupervised) mapping is superior to the classification based on the AA NN and PC feature extractor and highly comparable with that based on t he (supervised) MLP. (C) 1998 Pattern Recognition Society. Published b y Elsevier Science Ltd. All rights reserved.