Subthreshold neuromorphic devices for Spiking Neural Networks applied to embedded A.I

Abstract

Energy autonomy is one of the major challenges of embedded Artificial Intelligence. Among the candidate technologies likely to take up such a challenge, spiking neural networks are the most promising because of both their spatio-temporal and sparse representation of the information. In this context, this paper presents a neuromorphic approach based on an industrial CMOS technology and adopting an entirely subthreshold mode of operation (supply voltage VDD lower than the MOSFET threshold voltage). The detailed topologies of fabricated artificial neurons and synapses are presented as well as experimental results, validating an energy consumption of the order of a few femto-Joules per spike. Also, an arrangement of neurons and synapses is proposed to qualify experimentally this subthreshold approach in the perspective of highly energy efficient spiking neural networks.