Multi-Level Split Weight Learning for Multi-Cell ReRAM Computation-in-Memory

Neural networks (NNs) using binary weights have attracted attention as a hardware implementation method for lightweighting. Although the use of binary weights during training degrades the recognition performance in inference, relatively high inference performance can be achieved by generating a weight distribution using the gradient descend. In this study, towards establishing computation-in-memory architecture using an ReRAM with multi-cells, we show that inference recognition accuracy can be improved by using the weights of three and four levels.