Title: Binary Spiking Neural Networks as Causal Models

Abstract: This study presents a causal analysis of Binary Spiking Neural Networks (BSNNs) to elucidate their operational behavior. We establish a formal definition of BSNNs, modeling their spiking dynamics through the framework of a binary causal model. This causal representation enables the interpretation of network outputs using logic-based techniques. Specifically, we demonstrate the efficacy of employing both SAT and SMT solvers to derive abductive explanations from the defined binary causal structure. To validate our methodology, we trained a BSNN on the standard MNIST dataset and utilized our SAT- and SMT-based approaches to identify abductive explanations for the network’s classifications, focusing on pixel-level features. Furthermore, we benchmarked these explanations against SHAP, a widely adopted technique in explainable AI. Our results indicate that, in contrast to SHAP, our method ensures that identified explanations exclude entirely irrelevant features.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC