A probabilistic logic neuron network for associative learning

There is a developing interest in parallel systems where properties emerge from some aspect of their connections. These are called connectionist systems or parallel distributed processors. Although most of these systems use a cell that is related to the weight-sum-and-threshold neuron model first proposed by McCulloch & Pitts, the element in this paper is based on the random-access memory (RAM) model of the neuron, first proposed by one of the authors. In fact, the element described in this paper is a probabilistic version of the RAM model which responds with a randomly generated output for inputs on which it has not been trained. We call these probabilistic logic neurons (PLN). The multi-layer associative network described here uses PLNs to provide learnt image-to-image transformations with design rules and an algorithm which allow the designer to achieve such transformation with a selected degree of accuracy.