Graphene-based Artificial Dendrites for Spatio-Temporal Processing in Spiking Neural Networks

Neuromorphic computing has emerged as an important field to reduce the energy impact of artificial neural networks (ANNs). Among the types of neural networks, spiking neural networks (SNNs) rely on spatio-temporal spikes to transmit information, allowing even greater energy efficiency than the constant signals in ANNs. Due to increased complexity in the time domain, nonlinear dynamics in artificial neurons can significantly impact the performance of an SNN. These nonlinear dynamics can be mediated by dendrites, an SNN component that is relatively unexplored in neuromorphic computing, with few device candidates.<br/><br/>In this work, we propose a biocompatible graphene-based artificial dendrite that can implement tunable dendritic kernels at variable timescales, presenting a platform that can bridge between artificial and biological neural systems. We fabricated mesoscale (few mm2) and microscale (few µm2) synaptic transistor devices by interfacing graphene with a Nafion membrane. We used the dual-gate operation of the device in current mode to model alpha and Gaussian dendritic kernels, showing a transformation of input spikes to a non-spiking spatio-temporal output signal. The timing can be adjusted several orders of magnitude between microsecond to second, allowing a wide range of tunability. Device characteristics of conductance range, energy dissipation, and temperature dependence are also evaluated. A SNN constructed using the Linear Solutions of Higher Dimensional Interlayers (LSHDI) methodology is used to evaluate the performance of the artificial dendrites. We train the network offline on spoken digits from the Texas Instruments 46-word corpus using the modeled dendrites, then translate the software inputs to the dendrites into electrical signals to demonstrate functionality on a collection of hardware artificial dendrites. The results propose an artificial dendrite that is biocompatible, energy efficient, and highly tunable.