Adaptive Gateway Element Based on a Recurrent Neurodynamical Model

Abstract

Dynamic model of a recurrent neuron with a sigmoidal activation function is considered. It is shown that with the presence of a modulation parameter its activation characteristic (dependence between input pattern and output signal) varies from a smooth sigmoid-like function to the form of a quasi-rectangular hysteresis loop. We demonstrate how a gateway element can be build using a structure with two recognizing neurons and one output neuron. It is shown how its functional properties change due to changes in the value of the modulation parameter. Such gateway element can take the output value based on a weighted sum of signals from the recognizing neurons. On the other hand it can perform a complex binary-like calculation with the input patterns. We demonstrate that in this case it can be used as a coincidence detector even for disjoint-in-time patterns. Futhermore, under certain extreme conditions it can be triggered even if only the one input pattern was recognized. Also the results of numerical simulations presented and some directions for further development suggested.