Evaluation of training and mapping Sigma-pi networks to a massively parallel processor

This paper presents a methodology for training and mapping Sigma-pi networks on to a massively parallel processing (MPP) system. The implementation uses a Sigma-pi neuron model that can be viewed as an associative element which enables one to easily map the model to a MPP associative string processor (ASP) structure. The novelty of this paper is that it utilises the associative nature of the Sigma-pi neuron model and their bounded quantised site-values (weights) to enable training of these types of neurocomputing systems to be very quick. We use three methods to enable us to do this: 1) utilises pre-calculated constrained look-up tables to train an artificial neural network; 2) pipeline the input vectors; and 3) utilises the `data parallel´ methodology to further increase the efficiency of training Sigma-pi networks with the associative reward-penalty (A_R-P) training regime. Our methodology of constrained look-up tables means that one can pre-calculate the output function of the node, the delta changes (Δ) required for the learning regime and the output error per visible node