Adaptive High Order Neural Trees for Pattern Recognition

In this paper, a new classifier, called adaptive high order neural tree (AHNT), is proposed for pattern recognition applications. It is a hierarchical multi-level neural network, in which the nodes are organized into a tree topology. It successively partitions the training set into subsets, assigning each subset to a different child node. Each node can be a first-order or a high order perceptron (HOP) according to the complexity of the local training set. First order perceptrons split the training set by hyperplanes, while n-order perceptrons use n-dimensional surfaces. An adaptive procedure decides the best order of the HOP to be applied at a given node of the tree. The AHNT is grown automatically during the learning phase: its hybrid structure guarantees a reduction of the number of internal nodes with respect to classical neural trees and reaches a greater generalization capability. Moreover, it overcomes the classical problems of feed-forward neural networks (e.g., multilayer perceptrons) since both types of perceptrons does not require any a-priori information about the number of neurons, hidden layers, or neuron connections. Tests on patterns with different distributions and comparisons with classical neural treebased classifiers have been performed to demonstrate the validity of the proposed method.