A Node Splitting Neural Network with Integer

A node splitting network that classiies binary patterns by Hamming distance has been presented in 5]. There it has been shown that the node splitting training algorithm can produce a network with much fewer connections than Lippmann's Hamming net 9]. Now this network (NOSDIC) has been further developed to process integer values. For a pattern recognition problem in mobile radio communications traces of quantized signal levels have been used as input to the NOSDIC network. In this application the new version taking integer input vectors has resulted in a much smaller and more eecient network. The Node Splitting Distance Classifying Network (NOSDIC) originally designed for binary patterns 5] has been extended to take integer input values. Basically, the network consists of two layers of neurons whose number grows during training. In the rst layer, neurons take their input from selected pattern components and partially compute the distances between the test pattern and all trained (exemplar) patterns. The outputs of the rst layer are combined by summing nodes in the second layer to give the total distances. Classiication is achieved by nding the minimum distance, which may be implemented by decrementing devices 3] or a recurrent network 9]. The neuron model applied here is a summing node without threshold and saturation. Yet the rst layer nodes do not have a linear transfer function, because each input connection computes an absolute diierence, rather than a dot product as in classical neural network models. The current network has been innuenced by weightless neural systems, some of which react to pattern overlap (WISARD 2]) or classify binary patterns by Hamming distance, see 7] and in particular the Generalizing Random Access Memory, GRAM 1]. The computation of vector distances is an alternative to the notion of vector similarity based on dot products as in the McCulloch-Pitts neuron model 13] and has been used for a neuron model before 10]. The neuron function used here is: d n n = X i je i ? p i j n (1) where p i are input pattern components, e i refers to the exemplar pattern, and d n n relates to the n th order vector distance d n by the n th power. Because of the linear sum in (1) the computation of the distances between patterns can be distributed over many nodes. Taking the n th root may be omitted for comparison of distances, because …