A new classifier based on Bayes' principle that assumes the clustering of attribute values while boosting the attribute differences is presented. The method considers the error produced by each example in the training set in turn and updates the connection weights associated to the probability
of each attribute of that example. In this process the probability density of identical attribute values flattens out and the differences get boosted up. Using four popular datasets from the UCI repository, some of the characteristic features of the network are illustrated. The network is found to have optimal generalization ability on all the datasets. For a given topology, the network converges to the same classification accuracy and the training time as compared to other networks is less. One of the examples indicates the possibility that the optimization of the network may be done in parallel.
