ReRAM CiM Fluctuation Pattern Classification by CNN Trained on Artificially Created Dataset

Abstract

A CNN-based Fluctuation Pattern Classifier (FPC) is proposed. FPC is fully trained on the artificially created dataset with assumed fluctuation patterns such as random telegraph noise (RTN) and Oxygen Vacancy movement. FPC is applied to the measured ReRAM signals under different write conditions before read cycles and physical models are established based on the classification results. Proposed fluctuation reduction write (FRW) reduces ReRAM fluctuation rate by 35.1% to improve the inference accuracy of neural network.