Binary Spiking Neural Networks as Causal Models

A research paper on arXiv proposes using Binary Spiking Neural Networks (BSNNs) as causal models to explain network behavior. The authors formally define BSNNs and represent their spiking activity as binary causal models, enabling logic-based explanations. They demonstrate that SAT and SMT solvers can compute abductive explanations from these models. The BSNN was trained on the MNIST dataset, and the explanations were compared against SHAP, a popular explainable AI method. The authors claim their approach guarantees that explanations do not contain completely irrelevant features, unlike SHAP. The paper is categorized under Computer Science > Artificial Intelligence.