Vertical NAND Flash Memory‐Based Bayesian Neural Networks With Intrinsic Gaussian Weight Sampling

Bayesian neural networks (BNNs) provide improved robustness and well‐calibrated uncertainty estimates compared with deterministic networks, but their hardware realization remains challenging because Gaussian weight sampling typically requires random number generators, substantial memory overhead, or additional circuitry. Here we experimentally characterize commercial vertical NAND (V‐NAND) flash memory and show that a fully hardware‐compatible BNN can be realized without modifying the native array structure. We demonstrate that although individual ISPP‐programmed V‐NAND cells exhibit non‐Gaussian conductance distributions, the differential current between adjacent cells naturally forms a zero‐mean Gaussian distribution. The variance of these Gaussian weights can be tuned using only the read voltage, introducing no extra storage or peripheral circuitry. Using this hardware‐aware mapping, we construct a Bayesian last‐layer on top of a ResNet‐34 classifier and evaluate it in hardware‐calibrated simulations on the Tiny‐ImageNet‐C dataset. The resulting V‐NAND BNN achieves up to a 9.0% point improvement in accuracy over a deterministic counterpart under severe corruptions, along with enhanced calibration, ROC/PR performance, and separation of epistemic and aleatoric uncertainty. These results establish V‐NAND flash memory as a practical and scalable platform for uncertainty‐aware in‐memory computing and provide a blueprint for hardware BNNs.