Error Correction Scheme for Reliable RRAM-Based In-Memory Computing

摘要

Since conventional Hamming-code-based memory error correction scheme cannot be used for in-memory-computing circuits based on resistive random access memory (RRAM), we propose an error correction scheme using an arithmetic code of AN code that can correct the arithmetic errors by the maximum probability to improve the accuracy of dot product. When applied to a convolutional neural network on the MNIST datasets, the proposed error correction scheme largely reduces the accuracy loss.