In a groundbreaking development, MIT engineers have created an innovative "brain-on-a-chip" technology, smaller than a confetti piece, incorporating tens of thousands of artificial brain synapses called memristors. These silicon-based components are designed to replicate the information-transmitting synapses found in the human brain, marking a significant leap in neuromorphic computing.
The research team applied principles from metallurgy to construct each memristor using alloys of silver and copper combined with silicon. During testing with various visual tasks, this remarkable chip demonstrated the ability to "remember" stored images and reproduce them repeatedly with enhanced clarity and precision compared to existing memristor designs made from unalloyed elements.
Published in the prestigious journal Nature Nanotechnology, their findings showcase a promising new memristor design for neuromorphic devices—electronic systems based on a revolutionary circuit type that processes information in a manner similar to the brain's neural architecture. These brain-inspired circuits could be integrated into small, portable devices, enabling them to perform complex computational tasks currently requiring today's most powerful supercomputers.
"Artificial synapse networks have primarily existed as software until now. We're dedicated to building actual neural network hardware for portable artificial intelligence systems," explains Jeehwan Kim, associate professor of mechanical engineering at MIT. "Imagine connecting a neuromorphic device to your car's camera, allowing it to instantly recognize lights and objects and make decisions without internet connectivity. We aim to utilize energy-efficient memristors to accomplish these tasks locally, in real-time."
Revolutionary Ion Movement
Memristors, or memory transistors, represent fundamental elements in neuromorphic computing. Within a neuromorphic device, a memristor functions as the transistor in a circuit, though its operation more closely resembles a brain synapse—the connection point between two neurons. The synapse receives signals from one neuron in the form of ions and transmits a corresponding signal to the next neuron.
Unlike conventional transistors that transmit information by switching between only two values (0 and 1) and only when receiving an electric current of specific strength, memristors operate along a gradient, much like brain synapses. The signal they generate varies according to the strength of the received signal. This enables a single memristor to maintain multiple values and therefore perform a significantly broader range of operations than binary transistors.
Similar to a brain synapse, a memristor can also "remember" the value associated with a particular current strength and reproduce the exact same signal when receiving a similar current. This capability ensures reliable responses to complex equations or visual object classification—tasks typically requiring multiple transistors and capacitors.
Ultimately, researchers envision that memristors will require significantly less chip space than conventional transistors, enabling powerful, portable computing devices that don't depend on supercomputers or even internet connections.
Current memristor designs, however, face performance limitations. A single memristor consists of positive and negative electrodes separated by a "switching medium" or space between them. When voltage is applied to one electrode, ions from that electrode flow through the medium, forming a "conduction channel" to the other electrode. The received ions constitute the electrical signal that the memristor transmits through the circuit. The ion channel size (and consequently the signal the memristor produces) should be proportional to the stimulating voltage's strength.
Kim explains that existing memristor designs function adequately when voltage stimulates a large conduction channel or substantial ion flow between electrodes. However, these designs become less reliable when memristors need to generate subtler signals through thinner conduction channels.
The thinner the conduction channel and the lighter the ion flow between electrodes, the more difficult it becomes for individual ions to remain together. Instead, they tend to wander away from the group, dispersing within the medium. Consequently, the receiving electrode struggles to reliably capture the same number of ions and therefore transmit the same signal when stimulated with a specific low-range current.
Innovative Metallurgical Approach
Kim and his team discovered a solution to this limitation by adopting a technique from metallurgy—the science of combining metals into alloys and studying their collective properties.
"Traditionally, metallurgists add different atoms into a bulk matrix to strengthen materials. We thought, why not modify the atomic interactions in our memristor and incorporate some alloying element to control ion movement in our medium?" Kim explains.
Engineers typically use silver for a memristor's positive electrode. Kim's team researched literature to find an element that could combine with silver to effectively hold silver ions together while allowing them to flow quickly to the other electrode.
The team identified copper as the ideal alloying element, as it can bind with both silver and silicon.
"It functions as a bridge of sorts, stabilizing the silver-silicon interface," Kim notes.
To create memristors using their new alloy, the group first fabricated a negative electrode from silicon, then created a positive electrode by depositing a small amount of copper followed by a silver layer. They sandwiched the two electrodes around an amorphous silicon medium. In this manner, they patterned a millimeter-square silicon chip with tens of thousands of memristors.
As an initial test of the chip, they recreated a grayscale image of the Captain America shield. They correlated each pixel in the image to a corresponding memristor on the chip. They then modulated the conductance of each memristor to match the intensity of the corresponding pixel's color.
The chip produced an identical crisp image of the shield and demonstrated the ability to "remember" and reproduce the image multiple times, outperforming chips made from other materials.
The team also subjected the chip to an image processing task, programming the memristors to modify an image of MIT's Killian Court in several specific ways, including sharpening and blurring the original image. Once again, their design generated the reprogrammed images more reliably than existing memristor designs.
"We're utilizing artificial synapses to conduct actual inference tests," Kim states. "We aim to further develop this technology to create larger-scale arrays for image recognition tasks. Eventually, you might carry artificial brains to perform these tasks without connecting to supercomputers, the internet, or the cloud."
This research received partial funding from the MIT Research Support Committee funds, the MIT-IBM Watson AI Lab, Samsung Global Research Laboratory, and the National Science Foundation.
