Abstract:Memristor-enabled neuromorphic computing systems provide a fast and energy-efficient approach to training neural networks<a href="#ref-CR1">1</a>,<a href="#ref-CR2">2</a>,<a href="#ref-CR3">3</a>,<a href="/articles/s41586-020-1942-4#ref-CR4">4</a>. However, convolutional neural networks (CNNs)—one of the most important models for image recognition<a href="/articles/s41586-020-1942-4#ref-CR5">5</a>—have not yet been fully hardware-implemented using memristor crossbars, which are cross-point arrays with a memristor device at each intersection. Moreover, achieving software-comparable results is highly challenging owing to the poor yield, large variation and other non-ideal characteristics of devices<a href="#ref-CR6">6</a>,<a href="#ref-CR7">7</a>,<a href="#ref-CR8">8</a>,<a href="/articles/s41586-020-1942-4#ref-CR9">9</a>. Here we report the fabrication of high-yield, high-performance and uniform memristor crossbar arrays for the implementation of CNNs, which integrate eight 2,048-cell memristor arrays to improve parallel-computing efficiency. In addition, we propose an effective hybrid-training method to adapt to device imperfections and improve the overall system performance. We built a five-layer memristor-based CNN to perform MNIST<a href="/articles/s41586-020-1942-4#ref-CR10">10</a> image recognition, and achieved a high accuracy of more than 96 per cent. In addition to parallel convolutions using different kernels with shared inputs, replication of multiple identical kernels in memristor arrays was demonstrated for processing different inputs in parallel. The memristor-based CNN neuromorphic system has an energy efficiency more than two orders of magnitude greater than that of state-of-the-art graphics-processing units, and is shown to be scalable to larger networks, such as residual neural networks. Our results are expected to enable a viable memristor-based non-von Neumann hardware solution for deep neural networks and edge computing.

Memristive crossbar circuit for neural network and its application in digit recognition

Threshold-type memristor-based crossbar array design and its application in handwritten digit recognition

Analog Memristor Based Neuromorphic Crossbar Circuit for Image Recognition

Memristive Crossbar-Based Circuit Design of Back-Propagation Neural Network with Synchronous Memristance Adjustment

Memristor crossbar deep network implementation based on a Convolutional neural network

Associative learning of a three-terminal memristor network for digits recognition

Memristor‐Based Analog Computation and Neural Network Classification with a Dot Product Engine

A Fully Hardware-Based Memristive Multilayer Neural Network

Memristor Crossbar Architectures for Implementing Deep Neural Networks

A Self-Rectification and Quasi-Linear Analogue Memristor for Artificial Neural Networks

Hybrid Dual-Complementary Metal-Oxide-semiconductor/memristor Synapse-Based Neural Network with Its Applications in Image Super-Resolution

Circuit Modeling for RRAM-Based Neuromorphic Chip Crossbar Array with and Without Write-Verify Scheme.

Fully hardware-implemented memristor convolutional neural network

Fully Circuit Implementation of a Two-Layer Memristive Neural Network for Pattern Recognition

A Memristor Crossbar Based Computing Engine Optimized for High Speed and Accuracy

Memristor-Based Circuit Design for Multilayer Neural Networks

Memristor Neural Network Design

Memristor-based circuit implementation of Competitive Neural Network based on online unsupervised Hebbian learning rule for pattern recognition

Implementation of Multilayer Perceptron Network with Highly Uniform Passive Memristive Crossbar Circuits

A Circuit Implementation Method for Memristor Crossbar with On-chip Training

Long short-term memory networks in memristor crossbars