Long short-term memory networks in memristor crossbars

TL;DR

This paper demonstrates that memristor crossbars can implement LSTM networks, offering a low-power, high-capacity hardware solution that overcomes traditional bottlenecks in neural network computation for edge AI applications.

Contribution

The authors experimentally show that LSTM networks can be implemented in memristor crossbars, enabling in-memory computing and reducing power and latency for edge inference.

Findings

Memristor LSTM can perform regression and classification tasks.

The system exhibits low power consumption and low latency.

Memristor crossbars effectively store large model parameters.

Abstract

Recent breakthroughs in recurrent deep neural networks with long short-term memory (LSTM) units has led to major advances in artificial intelligence. State-of-the-art LSTM models with significantly increased complexity and a large number of parameters, however, have a bottleneck in computing power resulting from limited memory capacity and data communication bandwidth. Here we demonstrate experimentally that LSTM can be implemented with a memristor crossbar, which has a small circuit footprint to store a large number of parameters and in-memory computing capability that circumvents the 'von Neumann bottleneck'. We illustrate the capability of our system by solving real-world problems in regression and classification, which shows that memristor LSTM is a promising low-power and low-latency hardware platform for edge inference.