Fully parallel summation in a new stochastic neural network architecture

A space efficient fully parallel stochastic architecture is described. This stochastic architecture circumvents the main drawback of stochastic implementations of neural networks-the concurrent processing of a high number of weighted input signals-leading to a simple realization of stochastic summation. An unlimited number of stochastically coded pulse sequences can be added in parallel using only very simple and space efficient digital circuitry. Any neural network, either recurrent or feedforward, can be implemented using this scheme provided that neurons take discrete values. Design criteria are deduced from the mathematical analysis of the involved stochastic operations. Simulation results are also given