A Lightweight Hybrid Analog-Digital Spiking Neural Network for IoT

Abstract

In the Internet of Things (IoT) domain, we introduce a novel Hybrid Artificial Neural Network (ANN)-Spiking Neural Network (SNN), where the ANN component operates digitally using Field-Programmable Gate Array (FPGA) technology, while the SNN component functions analogically with passive circuits. This hybrid architecture proves to be efficient for carrying small drones/robots. Our system, employing FPGA and analog chip integration, guarantees seamless operation while showcasing scalability and efficiency, validated via simulations. We achieve 70% accuracy in the CIFAR10 data set with our hybrid ANN-SNN, compared to 80% with pure digital ANNs. Similarly, our hybrid ANN-SNN achieves 90% accuracy in the Street View House Numbers (SVHN) dataset, compared to 96% with pure digital ANNs. The added power is less than 5% of the power consumed by the FPGA. However, due to the transition of the spiking models from digital to analog, the accuracy performance is significantly around 45% for small FPGA models.