Energy Efficiency Optimization of a Spiking Neural Network for Spatio-Temporal Point Clouds Classification

In the past decade, artificial neural networks (ANNs) and deep neural networks (DNNs) have received exponentially increasing attention across various applications. However, DNNs are becoming increasingly expensive in terms of energy consumption, highlighting the need for more energy-efficient solutions. Spiking neural networks (SNNs) represent a novel class of ANNs designed to be more biologically plausible and energy-efficient by emulating the dynamics of natural neural networks. The main difference from a standard ANN is the use of the leaky integrate-and-fire (LIF) neurons that accumulate membrane potential over time and emit spikes when a threshold is reached. This work focuses on designing an SNN architecture for a classification task on time sequences of point clouds. The proposed approach wants to be an energy-efficient alternative to traditional architectures, such as PointNet, by leveraging the strengths of SNNs. Additionally, an energy analysis is proposed aiming to find a good energy consumption metric and to further optimize the final architecture.