A learning rule for CAM storage of continuous periodic sequences

Abstract

An analytic formula is used to set weights in recurrent analog networks with higher-order correlations to achieve the associative or content-addressable memory (CAM) storage of continuous pattern sequences as periodic trajectories. This learning rule allows programming of characteristics of the network vector field independently of the spatiotemporal patterns to be stored. Stability of sequences, basin geometry, and rates of convergence may be determined. A Lyapunov function in a special coordinate system governs the approach of initial conditions to the nearest stored trajectory