Random Spiking Neural Networks are Stable and Spectrally Simple

Ernesto Araya, Massimiliano Datres, Gitta Kutyniok

Keywords: Spiking Neural Networks, Stability, Simplicity Bias, Random Networks

Abstract: Spiking Neural Networks (SNNs) offer energy-efficient computation. We analyze discrete-time Leaky Integrate-and-Fire (LIF) SNNs via Boolean function theory, characterizing their \textit{noise sensitivity} and \textit{stability} under input perturbations. We show that wide LIF-SNN classifiers are on average stable, with Fourier spectra concentrated on low frequencies, implying a bias toward simple functions.

Serve As Reviewer: ~Massimiliano_Datres1

Submission Number: 23