Input multiplexing in artificial neurons employing stochastic arithmetic

Artificial neural networks employing stochastic arithmetic can under certain conditions outperform those based upon conventional radix arithmetic in reduced power dissipation, silicon area and improved fault tolerance. This paper describes limitations due to the inherent variance in the stochastic signals. We introduce and compare two stochastic multiplexing schemes, inter-count and intra-count multiplexing, for accumulating the total inputs to the artificial neurons.