ReSta: Recovery of Accuracy During Training of Deep Learning Models in a 14-nm Technology-Based ReRAM Array

In this article, we propose an electrical bias technique to recover the accuracy of a degraded ${\text {HfO}}_{x}$ -based resistive random access memory (ReRAM) array in deep neural network (DNN) training. We simulate degradation through the application of ${\sim } {10}^{{4}}$ pulses having high pulse amplitude, resulting in a fatigued ReRAM array that fails to converge during training. We propose a novel technique, recovery stabilization (ReSta), which can recover the array accuracy up to the level it was before the fatigue was introduced. After using the proposed controlled recovery technique, we obtain an accuracy of 98% on the reduced Modified National Institute of Standard and Technology (MNIST) classification task, approaching a floating point baseline. This work demonstrates a viable pathway to recover the performance of the fatigued ReRAM crossbar arrays in in-memory DNN training.

I. Introduction

Analog in-memory accelerators based on resistive random access memory (ReRAM) are complementary metal–oxide–semiconductor (CMOS) compatible with unprecedented potential for energy-demanding artificial intelligence applications [1], [2], [3], [4], [5], [6], [7], [8]. By performing the computation directly within the nonvolatile memory, analog in-memory accelerators eliminate the significant energy requirement of data movement between compute and memory unit of the traditional von-Neumann architecture [9], [10], [11], [12].