Multi-Neuron Connection Using Multi-Terminal Floating-Gate Memristor for Unsupervised Learning

Heterosynaptic plasticity in synapses has been successfully demonstrated by multi-terminal memristor and memtransistor (MT-MEMs)^1,2. However, these MT-MEMs lack the capability to mimic the membrane potential of neurons in multiple neuronal connections. In this study, we demonstrate a multi-terminal floating-gate memristor (MT-FGMEM) to emulate multi-neuron connections. The variable Fermi level (<i>E</i>_F) in graphene allows the charging and discharging of MT-FGMEM using multiple horizontally spaced electrodes. The MT-FGMEM exhibits a high on/off ratio over 10⁵ with a retention time of 1000 seconds, approximately 10,000 times higher than other MT-MEMs. The linear relationship between current (<i>I</i>_D) and floating gate potential (<i>V</i>_FG) in the triode region of the MT-FGMEM allows for accurate spike integration at the neuron membrane. The MT-FGMEM fully mimics the temporal and spatial summation of multi-neuron connections, based on the leaky-integrate-and-fire (LIF) functionality. Our artificial neuron consumes significantly less energy, approximately 100,000 times lower (150 pJ), compared to conventional neurons based on silicon integrated circuits (11.7 μJ). By integrating neurons and synapses using MT-FGMEMs, we successfully emulate spiking neurosynaptic training and classification of directional lines in the visual area one (V1), based on the LIF functionality of neurons and the spike-timing-dependent plasticity (STDP) of synapses. We achieved a learning accuracy of 83.08% on the unlabeled MNIST handwritten dataset in unsupervised learning based on our artificial neurons and synapses.