The July 2022 issue of IEEE Spectrum is here!

Close bar

Brainwave: If Memristors Act Like Neurons, Put Them in Neural Networks

Newfound “edge AI” applications for device that integrates memory and computing, has randomness built in, sips battery power

3 min read
The memristors (the top grey-and-black sandwich stuctures) in a new array can exploit randomness to perform machine-learning tasks.
The memristors (the top grey-and-black sandwich stuctures) in a new array can exploit randomness to perform machine-learning tasks.
Image: Image: CEA-Leti

A low-power and non-volatile technology called the memristor shows initial promise as a basis for machine learning. According to new research, memristors efficiently tackle AI medical diagnosis problems, an encouraging development that suggests additional applications in other fields, especially low-power or network “edge” applications. This may be, the researchers say, because memristors artificially mimic some of the neuron’s essential properties. 

Memristors, or memory resistors, are a kind of building block for electronic circuits that scientists predicted roughly 50 years ago but only created for the first time a little more than a decade ago. These components, also known as resistive random access memory (RRAM) devices, are essentially electric switches that can remember whether they were toggled on or off after their power is turned off. As such, they resemble synapses—the links between neurons in the human brain—whose electrical conductivity strengthens or weakens depending on how much electrical charge has passed through them in the past.

In theory, memristors can act like artificial neurons capable of both computing and storing data. As such, researchers have suggested memristors could potentially greatly reduce the energy and time lost in conventional computers shuttling data back and forth between processors and memory. The devices could also work well within neural networks, which are machine learning systems that use synthetic versions of synapses and neurons to mimic the process of learning in the human brain.

One challenge with developing applications for memristors is the randomness found in these devices. The level of electrical resistance or conductivity seen in memristors depends on a handful of atoms linking up two electrodes, making it difficult to control their electrical properties from the outset, says study lead author Thomas Dalgaty, an electrical engineer at Grenoble Alpes University in France.

Now Dalgaty and his colleagues have developed a way to harness this randomness for machine learning applications. They detailed their findings this month in the journal Nature Electronics.

Memristors are programmed by cycling through high-conductance on states and low-conductance off states. Usually the level of electrical conductivity seen in memristors can vary between one on state and the next due to intrinsic random processes within the devices.

However, if memristors are cycled on and off enough, the electrical conductivity of each memristor follows a pattern—“a bell curve,” Dalgaty says. The scientists revealed they could implement an algorithm known as Markov chain Monte Carlo sampling that could actively exploit this predictable behavior to solve a number of machine-learning tasks.

When compared with the performance of conventional digital CMOS electronics, the researchers’ memristor arrays achieved a stunning five order of magnitude reduction in energy. This, Dalgaty says, is because the memristors did not need to shuffle data back and forth between processors and memory. For context, that 100,000-fold discrepancy is equivalent to “the difference in height between the Burj Khalifa, the tallest building in the world, and a coin,” he explains.

One potentially exciting application for memristors would be devices capable of learning, adapting and operating at the far ends of a network (a.k.a. its “edge”), where low-power devices like embedded systems, smart home gear and IoT nodes sometimes reside. Indeed, Dalgaty says, memristors could help make edge learning devices a reality.

“Currently edge learning is not possible because the energy required to perform machine learning with existing hardware is far greater than the energy that is available at the edge,” he explains. “Edge learning [using memristors] ... can potentially open up completely new application domains that were not possible before.”

For example, the researchers used an array made of 16,384 memristors to detect heart rhythm anomalies from electrocardiogram recordings, reporting a better detection rate than a standard neural network based on conventional, non-memristor electronics. The team also used their array to solve image recognition tasks such as diagnosing malignant breast-tissue samples.

Potential future edge learning memristor applications might include implanted medical early-warning systems that can adapt to a patient’s state as it changes over time. “We are looking towards these really energy-constrained edge applications that maybe don’t or can’t exist yet because of energy [restrictions],” Dalgaty says.

The next big challenge, Dalgaty says, “will be putting all of this functionality together onto a single integrated chip that can be applied outside of the laboratory.” It may take a few years before such a chip exists, he says.

The Conversation (0)

How the U.S. Army Is Turning Robots Into Team Players

Engineers battle the limits of deep learning for battlefield bots

11 min read
Robot with threads near a fallen branch

RoMan, the Army Research Laboratory's robotic manipulator, considers the best way to grasp and move a tree branch at the Adelphi Laboratory Center, in Maryland.

Evan Ackerman

“I should probably not be standing this close," I think to myself, as the robot slowly approaches a large tree branch on the floor in front of me. It's not the size of the branch that makes me nervous—it's that the robot is operating autonomously, and that while I know what it's supposed to do, I'm not entirely sure what it will do. If everything works the way the roboticists at the U.S. Army Research Laboratory (ARL) in Adelphi, Md., expect, the robot will identify the branch, grasp it, and drag it out of the way. These folks know what they're doing, but I've spent enough time around robots that I take a small step backwards anyway.

This article is part of our special report on AI, “The Great AI Reckoning.”

The robot, named RoMan, for Robotic Manipulator, is about the size of a large lawn mower, with a tracked base that helps it handle most kinds of terrain. At the front, it has a squat torso equipped with cameras and depth sensors, as well as a pair of arms that were harvested from a prototype disaster-response robot originally developed at NASA's Jet Propulsion Laboratory for a DARPA robotics competition. RoMan's job today is roadway clearing, a multistep task that ARL wants the robot to complete as autonomously as possible. Instead of instructing the robot to grasp specific objects in specific ways and move them to specific places, the operators tell RoMan to "go clear a path." It's then up to the robot to make all the decisions necessary to achieve that objective.

Keep Reading ↓Show less