Toward Trustworthy Neuromorphic AI: A Bayesian Framework for Uncertainty‐Aware Spiking Neural Networks

Spiking Neural Networks (SNNs) offer a biologically plausible and energy‐efficient paradigm for processing temporal data, particularly suited for neuromorphic computing platforms. However, their deterministic nature limits their deployment in safety‐critical applications where reliable uncertainty quantification is essential. This paper introduces a novel Bayesian Spiking Neural Network (BSNN) framework that integrates variational inference with surrogate gradient learning to enable robust uncertainty estimation while preserving the efficiency of SNNs. We formulate a scalable Bayesian framework using mean‐field variational approximation over network weights and biases, enabling full predictive uncertainty quantification through Monte Carlo sampling. To address the non‐differentiability of spiking dynamics, we employ a smooth surrogate gradient method based on sigmoidal derivatives during backpropagation through time. A tailored Poisson encoding scheme ensures rich temporal input representation, while a KL annealing strategy stabilizes training by gradually increasing the regularization pressure from prior distributions. Comprehensive experiments on the N‐MNIST dataset (Digits 1, 2, 3) demonstrate competitive classification accuracy (96.85%) alongside well‐calibrated uncertainty estimates, as evidenced by strong calibration curves (expected calibration error [ECE] = 0.0066) and high area under receiver operating characteristic (AUROC) curve scores (> 0.994). The proposed BSNN achieves superior calibration and computational efficiency compared to deterministic, MC Dropout, and ensemble SNN baselines, with training completing in 671 s (11.18 min) upon early stopping at epoch 20. Our approach bridges the gap between efficient neuromorphic computation and trustworthy artificial intelligence (AI), offering a pathway toward deployable, uncertainty‐aware edge intelligence systems.