Continuous attribution on discrete spikes: revealing the explainability paradox in spiking neural networks

Abstract

Spiking neural networks (SNNs) offer a highly energy-efficient, neuromorphic alternative to standard deep learning, but their real-world deployment necessitates trustworthy and explainable outputs. Class Activation Mapping (CAM) methods are a robust and popular way of implementing this explainability in continuous artificial neural networks, but translating these methods to the discrete, temporal domain of SNNs introduces distinct theoretical failure modes. In this work, we formalize these limitations, establishing a fundamental Sensitivity-Faithfulness Tradeoff in gradient-based attributions and expose systematic binary threshold effects in perturbation-based approaches. Building on this, we identify a severe domain mismatch when applying continuous attribution smoothing to sparse neuromorphic data. To correct the resulting out-of-distribution hallucinations, we introduce Poisson SmoothGrad, a physically-consistent noise injection method that safely marginalizes threshold volatility to yield the strongest gradient-based attributions we measure on native event data. Benchmarking six attribution methods on a spiking VGG-11 across CIFAR-10, CIFAR-10-DVS and DVSGesture, we find that no single method dominates: domain-native smoothing leads the gradient family on sparse event streams, while the activation-based spike activation map is the strongest overall. Ultimately, we measure the network-level dynamics that allow CAMs to keep working despite these neuron-level mathematical flaws, and synthesize our empirical benchmarks into a Practitioner’s Guide that provides actionable heuristics for deploying robust, domain-aware visual explanations in SNNs.