Information update on neural tree networks

A method for information update on a supervised neural structure is presented. Neural trees are hybrid concepts between decision trees and neural networks. The method, applied to neural trees, combines the advantages of classical and neural classifiers, allowing both the update of the system without destroying previous information, and the use of all available features, inferring by itself which are the most important ones and the relation between them. The algorithm is named IUANT-information update algorithm for neural trees. It is robust to noise and also supplies good performance in comparison with the standard approach to retraining the tree. Moreover, it allows a large gain of time in the training phase. An application of the method is presented on a large (>3000) database of images