Graphene\u2013ferroelectric transistors as complementary synapses for supervised learning in spiking neural network

Yangyang Chen,Yue Zhou,Bobo Tian,Fuwei Zhuge,Yuhui He,Yi Li,Xiang Shui Miao,Mengge Yan

doi:10.1038/s41699-019-0114-6

Abstract

The hardware design of supervised learning (SL) in spiking neural network (SNN) prefers 3-terminal memristive synapses, where the third terminal is used to impose supervise signals. In this work we address this demand by fabricating graphene transistor gated through organic ferroelectrics of polyvinylidene fluoride. Through gate tuning not only is the nonvolatile and continuous change of graphene channel conductance demonstrated, but also the transition between electron-dominated and hole-dominated transport. By exploiting the adjustable bipolar characteristic, the graphene–ferroelectric transistor can be electrically reconfigured as potentiative or depressive synapse and in this way complementary synapses are realized. The complementary synapse and neuron circuit is then constructed to execute remote supervise method (ReSuMe) of SNN, and quick convergence to successful learning is found through network-level simulation when applying to a SL task of classifying 3 × 3-pixel images. The presented design of graphene–ferroelectric transistor-based complementary synapses and quantitative simulation may indicate a potential approach to hardware implementation of SL in SNN.

Highlights

By mimicking the plasticity of brain, neuromorphic computing is capable of self-learning, while with revolutionary speed and energy efficiency, and is regarded as a promising candidate to generation computing.[1]
In a typical supervised learning (SL) task, the two-terminal memristors implement algorithms with iterative read-and-write operations: during the forward step the outputs are obtained through multiplying voltages from input neurons by the conductance of memristive synapses, while during the update step the conductance of memristors is delicately tuned in order to minimize the error between real outputs and the desired ones
It should be noted that the back gate does not cause memorable modulation of the channel conductance not suitable for artificial synapses

Summary

Introduction

By mimicking the plasticity of brain, neuromorphic computing is capable of self-learning, while with revolutionary speed and energy efficiency, and is regarded as a promising candidate to generation computing.[1]. 1234567890():,; Fig. 1 Schematic view of a 3-terminal memristive synapses for supervised learning (SL) in spiking neural network (SNN), where b the field effect transistors with graphene as channels and organic ferroelectric polyvinylidene fluoride (PVDF) as gate dielectric (GrFeFET) mimic the synaptic functions.

Results

Conclusion