Analyzing Internal Activity and Robustness of SNNs Across Neuron Parameter Space

TL;DR

This paper explores the parameter space of Spiking Neural Networks to identify optimal operating regions that balance accuracy and energy efficiency, and examines their robustness to adversarial noise.

Contribution

It characterizes the operational manifold of neuron parameters in SNNs, providing guidelines for tuning to optimize performance and robustness.

Findings

Optimal parameter regions improve accuracy and energy efficiency.

Operating outside the manifold increases spike correlation and reduces robustness.

Systematic exploration across datasets reveals consistent operational principles.

Abstract

Spiking Neural Networks (SNNs) offer energy-efficient and biologically plausible alternatives to traditional artificial neural networks, but their performance depends critically on the tuning of neuron model parameters. In this work, we identify and characterize an operational space - a constrained region in the neuron hyperparameter domain (specifically membrane time constant tau and voltage threshold vth) - within which the network exhibits meaningful activity and functional behavior. Operating inside this manifold yields optimal trade-offs between classification accuracy and spiking activity, while stepping outside leads to degeneration: either excessive energy use or complete network silence. Through systematic exploration across datasets and architectures, we visualize and quantify this manifold and identify efficient operating points. We further assess robustness to adversarial noise, showing that SNNs exhibit increased spike correlation and internal synchrony when operating outside their optimal region. These findings highlight the importance of principled hyperparameter tuning to ensure both task performance and energy efficiency. Our results offer practical guidelines for deploying robust and efficient SNNs, particularly in neuromorphic computing scenarios.