Vertically Integrated Dual-memtransistor Enabled Reconfigurable Heterosynaptic Sensorimotor Networks and In-memory Neuromorphic Computing

Abstract: Neuromorphic in-memory computing requires area-efficient architecture for seamless and low latency parallel processing of large volumes of data. Here, we report a compact, vertically integrated/stratified field-effect transistor (VSFET) consisting of a 2D non-ferroelectric MoS$_2$ FET channel stacked on a 2D ferroelectric In$_2$Se$_3$ FET channel. Electrostatic coupling between the ferroelectric and non-ferroelectric semiconducting channels results in hysteretic transfer and output characteristics of both FETs. The gate-controlled MoS$_2$ memtransistor is shown to emulate homosynaptic plasticity behavior with low nonlinearity, low epoch, and high accuracy supervised (ANN - artificial neural network) and unsupervised (SNN - spiking neural network) on-chip learning. Further, simultaneous measurements of the MoS$_2$ and In$_2$Se$_3$ transistor synapses help realize complex heterosynaptic cooperation and competition behaviors. These are shown to mimic advanced sensorimotor neural network-controlled gill withdrawal reflex sensitization and habituation of a sea mollusk (Aplysia) with ultra-low power consumption. Finally, we show logic reconfigurability of the VSFET to realize Boolean gates thereby adding significant design flexibility for advanced computing technologies.