Multiple Level Evolutionary Learning in Neuronal Pattern Recognition

Enzymatic neurons are model neurons whose input-output behavior is controlled by internal dynamical mechanisms. In the present paper, we report on an enzymatic neuron model in which signal integration is mediated by the subcellular matrix. Readout enzymes that control the output of the neuron are activated by the integration of internal signals in space and time. The dynamics are modeled using a generalized automaton framework that allows for structure-function plasticity. The pattern recognition and generalization capabilities of the neurons are evolved by applying a variation-selection algorithm to both the signal integration dynamics and to the readout enzymes. Evolution of the signal integration dynamics means altering pathways of internal signal flow in the subcellular matrix, while evolution of readout enzymes means altering the firing behavior of the neuron in response to these signals. The neurons are presented with a small set of training inputs (both positive and negative instances) for learning and tested for their generalization properties.