Mutual Information Analysis of Neuromorphic Coding for Distributed Wireless Spiking Neural Networks

TL;DR

This paper evaluates neuromorphic coding techniques for wireless spiking neural networks using information-theoretic measures to understand performance trade-offs, aiming to optimize energy efficiency and accuracy in constrained wireless systems.

Contribution

It introduces a quantitative analysis of neuromorphic impulse radio coding algorithms for distributed wireless SNNs using information-theoretic measures, highlighting their performance trade-offs.

Findings

Neuromorphic coding techniques can be effectively characterized by information-theoretic measures.

Trade-offs between energy efficiency and inference accuracy are quantifiable using these measures.

The analysis guides the selection of coding algorithms for wireless neuromorphic systems.

Abstract

Wireless spiking neural networks (WSNNs) allow energy-efficient communications, especially when considering edge intelligence and learning for both terrestrial beyond 5G/6G and space networking systems. Recent research work has revealed that distributed wireless SNNs (DWSNNs) show good performance in terms of inference accuracy and low energy consumption of edge devices, under the constraints of limited bandwidth and spike loss probability. Following this reasoning, this technology can be promising for wireless sensor networks (WSNs) in space applications, where the energy constraint is predominant. In this work, we focus on neuromorphic impulse radio (IR) transmission techniques for DWSNNs, quantitatively evaluating the features of different coding algorithms that can be viewed as IR modulations. Specifically, the main contribution of this work is the evaluation of information-theoretic measures that may help in quantifying performance trade-offs among existing neuromorphic coding techniques.