Optimization of Short-Term Relaxation Effect by Dual Error Margin Scheme for Resistive Random-Access Memory-Based Next Generation Reservoir Computing

Next Generation Reservoir Computing (NGRC) has demonstrated strong potential in hardware-based machine learning applications, leveraging RRAM for efficient feature vector generation. However, the short-term relaxation effect (STRE) in resistive random-access memory (RRAM) poses challenges to computational accuracy. This study investigates the impact of error margin (EM) on STRE and proposes a dual error margin (DEM) scheme to address the conflicting EM requirements in resistance mapping stage and feature vector computation stage. By decoupling the mapping margin (marginA) and computation margin (marginB), the DEM scheme optimizes system performance. Experimental validation using the MNIST dataset shows that DEM improves accuracy from 87.4% (SEM) to 93.7%. The proposed approach provides an effective solution for enhancing the accuracy and reliability of RRAM-based NGRC system.