Neural crystal a dynamic recurrent network
Fred Jordan · 2002
Unlike feedforward MLP, recurrent networks have ability to work with temporal inputs and outputs. So the signal can be input directly to the network without preprocessing like time-windowing and without having direct dependency between signal duration and number of neurons. Crystal organisation of neurons has already been proposed by a number of neurobiologists but less work has been done to study properties of such ANNs. An important difference with recurrent multilayered networks lies in the 8 isotropic directions of the structure: it has the same topological organisation along 8 directions. The motivation for the architecture described in this paper is then to propose a model which presents the following properties: Hardware feasibility, compact matricial formalism and temporal dynamic. So we present a mathematical formalism and four examples of application in robotic and signal recognition.>