Distributed knowledge representation in fully connected networks

Abstract

Fully-connected binary networks, in addition to implementing content addressable memories, have been shown to be capable of encoding arbitrary limit cycles using synchronous dynamics. A stochastic knowledge representation paradigm is proposed, and a way to encode this knowledge form into cycles in fully-connected networks is described. This new representation format stores information in a truly distributed manner across the network, as opposed to previous schemes which store one knowledge atom per neuron.