Theses and Dissertations

Mitigating Non-Ideal Characteristics of Memristor-Based DNN Hardware

Date of Award

12-2022

Document Type

Thesis

Degree Name

M.S.

Department

Electrical and Computer Engineering

Committee Chair

Jinhui, Wang, Ph.D.

Advisor(s)

Dr. Mohamed Shaban, Dr. Silas Leavesley

Abstract

Memristor based hardware development has recently received increased attention in academia and industry for edge computing applications, owing to its power efficiency and low latency in performing complex analog in-situ matrix-vector multiplication which is critical to operation in deep neural networks (DNNs). However, the Stuck at Fault (SAF) and nonlinearity behavior of memristors makes such memristor based DNNs inefficient in practice. For SAF, a mapping transformation (MT) method is proposed to increase reliability by mitigating SAF in memristor based DNNs. The weight distribution for the VGG8 model using the CIFAR10 dataset is analyzed and investigated firstly. The MT method is then utilized to obtain inference accuracies ranging from 0.1% to 50% SAFs. The results of the experiments reveal that the MT method can restore memristor based DNNs to their original accuracies. For nonlinearity, the programmed weight for nonlinearity method is proposed to alleviate weight update error in the training process and enhance the inference accuracy even at extreme nonlinear conditions. This method is verified for different technology nodes, device variations, and neural network architectures. Finally, both techniques provide great advantages for improving the overall performance of memristor based DNNs when compared with the state-of-the-art.

This document is currently not available here.

Share

COinS