Memristor-based in-memory computing for multilayer artificial neural networks

Abstract

Analogue in-memory computing and brain-inspired computing based on the emerging memory technology such as Resistive RAM (RRAM) and Phase Change RAM (PRAM), integrates the logic module into the storage module and has a much higher energy efficiency, providing a feasible alternative method to further improve the computing performance beyond CMOS technology. Among the devices, memristor also called ReRAM, is considered to be the most promising candidate due to its ultra-low switching voltage (less than 3V) and power consumption (~0.1pJ), fast switch speed (~10 ns), compatibility with CMOS technology, and ultra-high integration capability (~4 F2, F is the feature size of process technology). It is found that the resistance of a memristor simply depends on the applied voltage. Therefore, a memristor can mimic the work mechanism of the human brain and function as a synapse. Memristor crossbar, in which a single memristor serves as an element in the weight matrix in the forward propagation in the neural network, provide great advantages to perform as a hardware platform for both digital logic circuit application and analogue neuromorphic computing due to the programmability, non-volatility, fast write/read speed and excellent scalability of array. In this thesis, a series of work related with memristor have been illustrated and discussed. The basic physical principle of the memristor will be first explained and the electronic characteristics will be simulated with PSpice. Oxide-material based memristors are fabricated and the device characteristics are extracted for the simulation of deep learning. A multilayer Deep Neural Network (DNN) based on the memristors is established and the

accuracy performance was tested on Modified National Institute of Standards and Technology data base (MNIST). The Simulation results show high accuracy (over 90%). Additionally, based on the properties of memristors, ternary and binary neural networks were proposed. Both of these networks achieved high accuracy around 90% and high noise robustness.