Modeling and Demonstration of Oxygen Vacancy-Based RRAM as Probabilistic Device for Sequence Learning

The joint device-algorithm development of a resistive RAM (RRAM)-based sequence learning system is presented. The low-voltage gradual RESET of oxygen-vacancy-based RRAM was characterized and modeled by kinetic Monte Carlo simulations based on the hour-glass model. In the low-voltage regime, the RESET becomes stochastic and depends on the SET-history of the device. The stochastic RESET is detrimental to deep neural network training, which requires precise weight updates. However, this intrinsic RRAM effect can be employed as a local learning rule for brain-inspired computing. Therefore, a novel and non-deep learning sequence learning approach that employs RRAM as active computational elements is proposed. Two applications using this algorithm are explored in both simulations, employing a parameterized learning rule and a hardware demonstration: sequence denoising and gait authentication. This article shows promise for the applicability of RRAM in non-deep learning applications targeting low-power online learning at the edge.