Background-Pattern-Dependency-Tolerant Weight Transfer Method for Accurate NAND-Flash-Based Spiking Neural Networks

This study focuses on the background pattern dependency (BPD) issue of NAND flash, where threshold voltage (V
TH
) or bit line current (I
BL
) change due to the differences between verify and read operations. In NAND-flash-based spiking neural networks (SNNs), which utilize I
BL
for analog multiply-accumulate operations, the impact of BPD is more significant than in conventional memory operation. A more-BPD-tolerant weight transfer method for accurate NAND-based SNNs is proposed and evaluated by using experimental data.