Heteroscedastic Gaussian Process Regression for ReRAM Device Modeling

Jump tables are useful modeling approaches for emerging memory devices, e.g. ReRAM, and their use in neural network simulations because they rely only on experimental data. While binning is traditionally used for such modeling, this work proposes the use of Heteroscedastic Gaussian Process Regression (hetGP) to estimate signal mean and standard deviation simultaneously and develop a robust jump table model. The binning and hetGP approaches are verified using Kolmogorov-Smirnov (K-S) and maximum mean discrepancy (MMD) test.