Dynamical Memories Based On Encoding In Nonconvergent Spatio-temporal Oscillations

Igor Beliaev, Computational Neurodynamics, University of Memphis

A novel type of dynamical neural model that is strongly biologically motivated is used as an encoding device for binary patterns. This neural network, known as Freeman’s K-sets, encodes information through spatio-temporal trajectories in the state space that are shaped by Hebbian rule and habituation during learning. We estimate the practical memory capacity of this system. Theoretically number of memories coded in dynamical trajectories can be far greater than in any convergent dynamics, i.e., fixed point trajectories. In a series of experiments on recalling the distorted binary patterns we compare the capacity of the K-model with that of Hopfield network and show that K-models are capable of forming larger number of memories, which are more robust and noise-tolerant. Results also indicate that the critical load parameter, as an indicator of model’s capacity, and which is known to be 0.14 for Hopfield model, is much higher for the K-model.