A biomorphic neuron model and principles of designing a neural network with memristor synapses for a biomorphic neuroprocessor

This paper presents an original biomorphic neuron model, which differs from common IT models by a more complex synapse structure and from biological models by replacement of differential equations that describe the change in potential over time with explicit recurrence expressions by approximation of experimental data in the cortical neuron, and therefore, by transition from the spiking information coding to the coding using the average frequency of action potentials per a simulation step. This approach ensures sufficiently simple and efficient calculation of an ultra-large neural network in the stand-alone hardware with limited computing resources. The model consists of three separate functional parts: dendrites, soma, and axon, which allows implementing any connections between functional parts of different neurons, thus making the neural network architecture more flexible. To perform functional testing of the neuron model, the test neural network performing simple association and constructed as a consequent stack of functional blocks with primary connections organized using experimental neurophysiological data was simulated. It is shown that encoding of the information transmitted by the impulses, similar to biological ones, allows using memristors for calculating recurrence expressions that describe the change in the quantity of neurotransmitter receptors of the dendrite membrane. The elaborated biomorphic neuron model, defined conceptual principles of a neural network construction based on it, as well as replacement of synapses in the neural network with memristors will allow building an ultra-large biomorphic neural network that simulates the functioning of a separate brain cortical column in the stand-alone hardware—a biomorphic neuroprocessor.