Binary Spiking Neural Networks as Causal Models
Their SAT-based method finds abductive explanations without any irrelevant pixel features.
Binary Spiking Neural Networks (BSNNs) are a class of energy-efficient neuromorphic models that communicate via binary spikes. However, their non-linear, temporal dynamics make interpretability challenging. In a new arXiv paper, researchers from CNRS (IRIT and CERCO) propose a novel framework that treats BSNNs as causal models, enabling rigorous, logic-based explanations. They formally define a BSNN and represent its spiking activity as a binary causal model, then leverage SAT and SMT solvers to compute abductive explanations—minimal sets of input features that are sufficient for a given output classification.
To validate, the team trained a BSNN on the MNIST digit dataset and compared their explanations against SHAP, a widely-used feature attribution method. The key advantage: the SAT/SMT approach guarantees that found explanations contain no completely irrelevant features. SHAP, in contrast, can assign non-zero importance to pixels that do not actually influence the decision. This work opens the door to more trustworthy explainable AI for spiking neural networks, with potential applications in safety-critical domains like autonomous driving or medical diagnosis where every feature in an explanation must be justified.
- BSNNs are modeled as binary causal systems to enable logic-based explanations.
- SAT and SMT solvers compute abductive explanations with provable minimality.
- Method guarantees no irrelevant features, unlike SHAP which can include extraneous pixels.
Why It Matters
Enables trustworthy, verifiable AI explanations for spiking neural networks in safety-critical applications.