Ultra-Compact, Scalable, Energy-Efficient $VO_{2}$ Insulator-Metal-Transition Oxide Based Spiking Neurons for Liquid State Machines

Spiking neural networks, inspired by biological neural systems, could process immense volumes of spatio-temporal data by representing them as spikes. Here, we propose to implement compact, scalable, energy-efficient spiking neurons based on the unique insulator-metal transition in Vanadium dioxide
$(VO_{2})$
which interact through memristive synapses, to emulate a Liquid State Machine (LSM). Further, we demonstrate the implementation of this recurrent neural network as a temporal auto-encoder, and adaptive channel equalizer for application in neuromorphic signal processing. Our approach provides a pathway to reduce component count
$(50\ -100X)$
and improve energy efficiency
$(> 50X)$
over conventional CMOS based implementations.