Programmable Linear RAM: A New Flash Memory-based Memristor for Artificial Synapses and Its Application to Speech Recognition System

摘要

In this work, a new type of flash memory-based memristor, named programmable linear random-access memory (PLRAM), is presented to store analog synaptic weights in a single flash memory cell. A PLRAM cell with a self-calibrating program/erase scheme can provide very stable and repeatable analog memory states up to 7 bits in a single cell, which is suitable for an artificial synapse in the neural network. The physical implementation of a discrete Fourier transform on PLRAM arrays shows a remarkably good agreement with theoretical calculations. By taking nonlinearity effects on both forward propagation and back-propagation into consideration, a highly manufacturable speech recognition system, which consists of a 5-layer fully connected neural network (360k artificial synapses, 1576 neurons, and peripheral circuits) has been successfully built on 200 mm wafers. For the first time, the high accuracy of speech recognition (>90%) on 8 different Chinese speech words is demonstrated.