The successful use of probability data in connectionist models

Reggia [10] explored connectionist models which employed “virtual” lateral inhibition, and included the activation of the receiving node in the equations for the flow of activation. Ahuja [1] extended these concepts to include summing the total excitatory and inhibitory flow into a node. He thus introduced the concept that the change of activation of a node depended on the integral of the flow into that node and not just the present activation levels of the nodes to which it is connected. Both Reggia’s and Ahuja’s models used probability data for the weights. Ahuja’s model was further extended by Alexander [2], [3], [4] in the RX model to allow both the weights and the activations of Ahuja’s model to be negative, and further, Alexander’s model included the prior probabilities of all nodes. Section 1 of this paper contains a complete listing of the RX equations and describes their development. The main result of this paper, the demonstration of the convergence of the system is presented in Section 2. Section 3 briefly describes the experiments testing the RX system and summarizes this article.