Decoupling the Contributions of Spatio-Temporal Coding: From ANNs to SNNs

Yihao Li, Hanle Zheng, Jiaxin Guo, Lei Deng

Keywords: spatio-temporal coding, Spiking Neural Networks, Artificial Neural Networks, neuromorphic computing

TL;DR: We examine the coding schemes of artificial neural networks (ANNs) and spiking neural networks (SNNs) by quantifying how information is distributed across three distinct domains: spatial, temporal, and activation.

Abstract: Artificial Neural Networks (ANNs) and Spiking Neural Networks (SNNs) represent two distinct but complementary approaches to information processing. ANNs, with their continuous activation functions, have been widely successful in tasks requiring nonlinear mapping, while SNNs provide a biologically plausible and energy-efficient alternative through their discrete, spike-based activity and spatio-temporal dynamics. To compare their coding schemes in-depth, we seek to decouple and analyze the contributions of spatio-temporal coding in these models. We introduce a novel mutual-information-based measure, the Exploitation Rate (ER), to quantify how information is distributed across spatial, temporal and activation domains. We also propose an incremental framework to analyze the transition between the two network paradigms. Our findings highlight the advantage of SNNs in leveraging rich temporal dynamics to compensate for their reduced complexity in activation values.

Submission Number: 83