Mathematical modeling of the processes of signal routing by logic matrix, information encoding and decoding in the biomorphic neuroprocessor

Abstract

At the University of Tyumen, a biomorphic hardware neuroprocessor based on a combined memristor-diode crossbar has been developed. The neuroprocessor implements a biomorphic spiking neural network with a large number of neurons and trainable synaptic connections between them. Large biomorphic neural networks make it possible to reproduce the functionality of the human brain cortical column. This provides new opportunities for information processing tasks by standalone neuroprocessor. When designing and optimizing the operation of the input and output devices, as well as the logic matrix of the neuroprocessor created based on large combined memristor-diode crossbars, physico-mathematical models are needed to simulate their work. This report presents the physico-mathematical models developed for this neuroprocessor: of the operation of a logic matrix cell built on the basis of simplified electrical models of a memristor and a Zener diode; of the process of the neurons output spikes routing of by the logic matrix to the synapses of other neurons; of processes of information encoding into biomorphic impulses and decoding of neural block output into a binary code. With the help of these models and numerical simulation, the operability of the input and output devices, as well as the logic matrix of the biomorphic neuroprocessor, is shown when processing incoming information. The originality of the models is associated with the specifics of the selected memristor-diode cell of the universal large logic matrix, which, in addition to its main work as a spikes router, is the basis of the electrical circuits of the input and output devices of the neuroprocessor. For numerical simulation of the operation of large electrical circuits containing memristor-diode crossbars, the original specialized program MDC-SPICE was used.