Analog hardware may solve Internet of Things' speed bumps and bottlenecks

 

The ubiquity of smart devices—not just phones and watches, but lights, refrigerators, doorbells and more, all constantly recording and transmitting data—is creating massive volumes of digital information that drain energy and slow data transmission speeds. With the rising use of artificial intelligence in industries ranging from health care and finance to transportation and manufacturing, addressing the issue is becoming more pressing.

A research team led by the University of Massachusetts Amherst aims to address the problem with new technology that uses old-school analog computing: an electrical component known as a memristor.

"Certainly, our society is more and more connected, and the number of those devices is increasing exponentially," says Qiangfei Xia, the Dev and Linda Gupta professor in the Riccio College of Engineering at UMass Amherst. "If everyone is collecting and processing data the old way, the amount of data is going to be exploding. We cannot handle that anymore."

Xia and colleagues at Finland's Tampere University, the University of Southern California and TetraMem Inc. have developed a brain-inspired sensing system that combines a touch sensor and a smart memory chip that only reacts when necessary to greatly improve energy efficiency and computing speed. Their work is described in an article published in Nature Sensors.

"Overall, our research goal is to reduce the power consumption, latency and hardware complexity," says Xia.

The team's memristor-based haptic (touch) sensor only processes data around pixels that contain a signal while disregarding irrelevant background noise. Consider a touchscreen containing tens of millions of pixels. "When you write, it's only a very small portion that are involved," he says. "You do not have to process all the information you got from the entire screen, only those pixels you're writing on."

Their proof-of-concept sensor system can currently recognize patterns with 87%–92% accuracy, faster and more energy efficiently than traditional computational methods, Xia says.

While the paper describes a touch sensor, Xia envisions other applications of the technology, such as event-based visual sensors. Consider a camera monitoring traffic at 30 frames per second: "During the daytime, it's very busy. It makes a lot of sense," he says. "But at 2 a.m., there is less traffic. If you keep doing 30 frames per second, you are wasting a lot of resources."

Xia also recently published a second paper, on the same day, in Nature Electronics where he and colleagues demonstrate a proof-of-concept design of a memristor-based, bioinspired artificial intelligence hardware called a cellular neural network (CeNN).

First envisioned in the 1980s, the CeNN takes inspiration from the retina, which uses local connections between neurons to analyze and respond to visual stimuli. The work by Xia and his team marks the first implementation of a memristor-based CeNN.

The repeating "cells" of the hardware only connect to its nearest neighbor, not all together, like today's deep neural networks that make up AI. The memristor, in this design, acts as the synapse between the different cells.

As a result, the CeNN has simplified circuit wiring and the data transmission—which requires tremendous amounts of power—can be greatly reduced. "When we process data from an image sensor, we have tens of millions of pixels," Xia says. "Each of the cells could take care of one pixel, and then process them at the same time. That's going to give us a huge advantage in terms of latency."