Temporal sequence learning with a history-sensitive probabilistic learning rule intrinsic to oxygen vacancy-based RRAM

摘要

Widely spread and low value resistance distributions inhibit the use of filamentary resistive RAM (RRAM) at low currents for deep learning training and inference. An entirely different approach which employs RRAM as active computational elements is proposed. For this means, the history-sensitive probabilistic reset in Tantalum-Oxide (TaOx)-based RRAM is characterized and explained. This intrinsic RRAM effect is used as a local learning rule in a novel temporal sequence learning algorithm.