A correlation-based network for hardware implementations

Abstract

An architecture and learning rules for a correlation-based network are proposed. Hidden activity predictors dynamically compute local temporal receptive field centres through a decorrelation process. Temporal feedback loops between units in the hidden layer are then used to synchronise the activities of similar near by units. The simultaneous activation of different topologically overlapping unit groupings results in a continual reorganisation of units in the hidden layer: the dependence of hidden intra-layer communication on cross-correlations gives it the image of an analogue spiking neural network. The predominantly feedforward nature of the architecture makes it attractive for implementation in parallel hardware. Some suggestions on how this can be accomplished are also proposed, together with some software simulation results on a problem of instantaneous separation of two sine waves with different phases.