Novel Weight Mapping Method for Reliable NVM based Neural Network

Abstract

Errors of vector-matrix-multiplication induced by interconnect resistance become a crucial reliability challenge in non-volatile memory (NVM) based neural network. Here, we propose a novel weight mapping method, called weight mapping correction (WMC), to mitigate the deviation of weight represented by the conductance of NVM array without time-consuming retraining and circuit overheads. Simulation results show that accuracy is recovered significantly when WMC is applied to various size arrays consisting of mainstream NVM in advanced technology nodes.