A Competitive Neural Network based on dipoles

Abstract

In a competitive neural network, a process unit (node) in the competitive layer is completely described by the vector of weight from the input node to it. Each such weight vector becomes the centroid of a cluster of inputs since the principal function of a competitive learning network is discovers cluster of overlapping input. In this paper we propose a competitive neural network where each process unit has a couple of weight vectors (dipoles) that becomes a line segment as representation of a cluster. A weight update is formulated such that the dipole associated with each process unit is as near as possible to all the input samples for which the node is the winner of the competition. This network allows the formation of groups or categories by means of unsupervised learning, where each class or category is identified by a line segment instead of a centroid. The line segment leads to a better representation of a group or class that a centroid that gives us only the position of the cluster. The network has been applied to the formation of groups or categories using the data IRIS, where the unsupervised learning algorithms reach between 12 and 17 incorrect classifications. However, while many partitional clustering algorithms and competitive neural networks are only suitable for detecting hyperspherical-shaped clusters, the proposed network gets only 5 incorrect classifications and is also suitable for detecting hyperspherical-shaped clusters.