A Simple Analog Neuron with Memristor Based Synapses and SATLINS Transfer Function

Memristors are unique nano-length elements, having sound memorizing and switching features and due to their small energy consumption and a good compatibility with present CMOS high-density integrated chips, they have applications in analog low-power neural nets, memory arrays, and in various electric circuits. In this paper, a simple neuron with memristor synapses is proposed, with only one memristor for each weight utilized. The saturation linear symmetric activation function (satlins) is used for it's good approximation mapping properties. The schematic is implemented in LTSPICE and on a breadboard. A simple model of memristors with activation threshold is utilized. The adder is based on two opamps, one for inputs attached to negative weights and one for positive weights. Each synaptic weight is created by only one memristor, suggesting a strongly decreased circuit complexity. Due to the reduced memristor operational voltage range a scaling is needed and scaling circuits are based on voltage dividers. The applied activation function is implemented with a humble scheme with a single op-amp. The memristor neuron is analyzed in LTSPICE and MATLAB. After a comparison, the obtained outcomes are confirmed. The offered neuron is a significant stage towards engineering of complex memristor neural nets in ultra high-density integrated chips.