BCN: an architecture for weightless RAM-based neural networks
Gareth Howells, M.C. Fairhurst, D.L. Bisset · 2002
This paper introduces a novel networking strategy for RAM-based neurons which significantly improves the training and recognition performance of such networks whilst maintaining the generalisation capabilities achieved in previous network configurations. The Boolean convergent network (BCN) is a RAM-based neural network where the inputs and output of the component neurons are taken from the values '0', '1' and the undefined value 'u'. The inputs to a neuron form an addressable set incorporating all memory locations which may be formed by treating any undefined value within the input as either a '0' or a '1'. The output of a neuron can be any defined value which occurs exclusively within the memory locations included within the addressable set. If the addressable set contains either no defined value or examples of both defined values, then the undefined value 'u' is output.>