Both artificial neurons and synapses are needed to efficiently implement artificial neural networks directly in hardware. Memristors have been actively investigated to implement synaptic functions. In contrast, the implementation of artificial neurons are still limited to Si-based circuit components, while single device based artificial neuron has not been well explored. In this study, we show that VO2 two-terminal devices, driven under electrical pulses, can be used to implement the leaky, integrate and fire function of spiking neurons. Since the thermally driven metal-insulator transition of VO2 is volatile and electrical pulses can produce Joule heating, while the heating and cooling have delay times, VO2 devices exhibit characteristic firing time and relaxation time, corresponding to timing difference of the exciting pulse and the resistance switching. In addition, the integrating effect of several small pulses can also trigger the firing from high resistance to low resistance transition. These characteristics suggest the potential using the volatile phase transition of VO2 for artificial spiking neuron. Experimental measurements are further elaborated by a 2-D heterogeneous resistor network based modelling. The insulator-metal transition in VO2 under electrical bias was simulated to confirm the dynamic formation of current filament between two electrodes, and the formation of this path is a flashing process during a single pulse in the pulse train when the integrate effect of these pulses arrives the threshold.
Read full abstract