Device Stochasticity Directly Implements Bayesian Sampling in Spiking Neurons

Device stochasticity is considered a non-ideality in neuromorphic hardware, although stochastic neural activity supports probabilistic computation in biological systems. We show that intrinsic stochastic switching of magnetic tunnel junctions (MTJs) can be mapped to stochastic-threshold spiking neurons, providing a physical sampling primitive for Bayesian spiking inference. We develop a spiking Bayesian neural network (SBNN) with neuronal threshold distributions learned during training and sampled during inference. We also introduce a rate-estimation method that propagates firing probabilities rather than explicit spike sequences, achieving up to a 22-fold speedup over surrogate-gradient training. Compared with representative Bayesian neural-network baselines, SBNN achieves favorable accuracy-calibration trade-offs, reaching high accuracy across static and event datasets and low negative log likelihood in both MNIST and Fashion-MNIST. On MNIST, SBNN retains $96.20\%$ accuracy under a normalized weight-noise standard deviation of 0.20 and $81.30\%$ under an input-noise standard deviation of 0.25, compared with $27.42\%$ and $70.27\%$ , respectively, for the spiking neural network (SNN). Hardware-in-the-loop validation using physical MTJ switching for sampling and software for network arithmetic closely matches software inference in spiking behavior and classification performance. This suggests that intrinsic device stochasticity can be transformed from a hardware non-ideality into a Bayesian computational resource for compact and noise-tolerant neuromorphic systems.