MIT’s Confetti-Sized “Brain-on-a-Chip” Is Replete With Thousands of Memristors

June 10, 2020 by Robin Mitchell

While there's some debate on whether memristors even exist beyond the theoretical realm, MIT seems to have made a major breakthrough in artificial brain synapse technology.

“Imagine connecting a neuromorphic device to a camera on your car, and having it recognize lights and objects and make a decision immediately, without having to connect to the internet," says Jeehwan Kim, associate professor of mechanical engineering at MIT.

"We hope to use energy-efficient memristors to do those tasks on-site, in real-time."

This week, MIT announced that they have fit tens of thousands of artificial brain synapses, or memristors, on a single chip. The researchers claim that this "brain-on-a-chip" is neural network hardware—as opposed to software artificial synapse networks—that can advance the development of portable intelligence systems.


The memristors are made up of a silver-copper alloy.

The memristors are made up of a silver-copper alloy. Image used courtesy of Peng Lin, MIT

What exactly is a memristor and why is it important in the role with AI?


First, Memristors

According to researcher Isaac Abraham, a memristor is a theoretical component that is considered to be one of the four fundamental parts: resistors, capacitors, inductors, and memristors. The device's resistance depends on the current that once flowed through it and describes a hysteresis pattern in its IV graph.

While working examples of memristor behavior have been designed and proven, some might argue that viable memristors have not yet been developed. For instance, three-pin devices developed in laboratories can only mimic a memristor; a true memristor would be a two-pin passive part (PDF), according to researchers Francesco Caravelli and Juan Pablo Carbajal. 


Memristors in Neural Networks

So why are memristors considered to be important in AI development?

Memristors can be thought of as the electrical component equivalent of connections between neurons in a brain. If the resistance of the memristor is used to model the strength of connections between neurons, then, in theory, an array of memristors between processing nodes (as simple as a transistor or as complex as a microcontroller) can mimic weighted connections and thus simulate neural networks.

The possibility of memristors is even more significant when we consider a memristor without complex software routines or hardware circuitry to perform a weighted calculation; in such an instance, those calculations would be instantaneous. 


MIT Neuromorphic Breakthrough

One of the major breakthroughs of MIT's development is the reduced size of the memristors, which was only possible thanks to a metallurgy practice involving doping.


MIT's Killian Court

The chip successfully reprocessed (blurring and sharpening) an image of MIT's Killian Court better than existing neuromorphic chips. Image used courtesy of MIT

Typical memristors rely on the movement of conductive ions in a channel to either make the channel resistive or conductive. This is not a problem for large memristors. However, as the physical dimensions of the memristor are reduced, the channel becomes harder to work with because there are fewer ions. Therefore, the team behind the device decided to alloy the memristors with copper to form a better conductive bridge between the silver memristors and the silicon substrate. 


A Memristor Acts as a Non-Binary Transistor

The confetti-sized memristor MIT discovered acts as a transistor in a neuromorphic device. But while transistors in an ordinary circuit would relay information by switching between 0s and 1s—and only when a received electrical signal is a certain strength—a memristor doesn't work in such a binary.

Instead, the signal the MIT memristor produces varies based on the strength of the signal it receives, yielding any number of values and opening the door to more operations than a conventional transistor.


The Chip Is Put to the Test—With Captain America's Shield

In the chip's first test, it was shown a grayscale image of the Captain America shield. The researchers converted each pixel in the shield to a value. Then, they altered the corresponding memristors' resistance to match the value of the pixel and altered the conductance of each memristor to coincide with the strength of its corresponding pixel color.

The chip produced reproduced a crisp replica image of the shield and was able to remember the image, reading it back multiple times. 


MIT's silver-copper alloy memristor

MIT's silver-copper alloy memristor reproduced a crisp image of Captain America's shield more effectively than similar chips made of other materials. Image used courtesy of MIT

When the researchers tasked the chip with the image processing task of altering an image, it successfully reprogrammed such images better than existing memristor designs.


A Bright Future for Neuromorphic Devices? 

While this technology still has a long way to go—for example, programming a neural net with weighted functions—the ability of the device to retain information in a manner near identical to that of neurons shows promise in neuromorphic advancements.

The future of AI may lie in analog silicon devices that do not rely on hardware or software routines to perform mathematical execution but instead operate at near real-time speed.