A comparison of multilayer perceptron neural network and Bayes piecewise classifier for chromosome classification

The performance of a multilayer perceptron (MLP) neural network (NN) as a classifier of human chromosome was compared to that of a Bayes piecewise classifier. Both classifiers were trained to classify 5 types of chromosomes according to density profile features. The MLP NN classifier outperformed the Bayes piecewise classifier for all the combinations of features and for all the sizes of training sets. The MLP classifier was found to be almost unsusceptible to the ratio of the number of training vectors to the number of features, where the piecewise classifier was highly depended on this ratio. The piecewise classifier required higher number of training vectors whenever there was an increase in the number of features used. Therefore, the Bayes piecewise classifier is limited to large data sets. However, the MLP classifier performed well even for small data sets. As far as our chromosome data is considered, the MLP NN classifier ability to generalize from the training set to test vectors is evidently stronger than that of the Bayes piecewise classifier