Bayesian Neural Network Implemented by Dynamically Programmable Noise in Vanadium Oxide

For the first time, the dynamically programmable noise on a vanadium oxide (VO
2
) device is extensively studied and exploited for implementing a Bayesian neural network (BNN). We demonstrate programming of noise in a VO
2
device with either resistance programming or temperature control. The VO
2
device achieved a 6.4 dynamic ratio on noise. We show that this ratio is sufficient to achieve ideal numerical levels of uncertainty quantification on CIFAR-100, achieving an expected calibration error of 3.7% (ECE measures the consistency between the network’s accuracy and uncertainty).