While computers consistently grow more powerful, their fundamental design hasn’t changed in the last 75 years or so. Still, nature has already outperformed all our best computers with the human brain. The fundamental unit of a computer is the transistor, which works more-or-less like a switch – it’s either on or off. The fundamental unit of the brain, however, can be thought of as the synapse, an electro-chemical interface connecting two neurons. One of the most important features of a synapse is that it isn’t binary (that is, it can take on more values than just on or off). This allows a single synapse to store more information than a single transistor. The “memristor” is an electrical device theorized first in 1971 and fabricated in 2008 that could do something similar to synapses. It can store past information in the strength of the signal being passed, in addition to the traditional on/off binary. The first memristors made were fairly unreliable, especially for very small signals. Now, materials engineers at MIT have used an important principle in metallurgy to make more stable memristors that could allow for a revolution in computer architecture, artificial intelligence, and data science.
Individual metals, like iron, rarely work as well, in terms of strength, toughness, or conductivity, as alloys – combinations of two or more metals, like steel. In a memristor, the same is true. In their “off” state, a non-conductive material (an insulator) separates the two electrically-conductive halves of the device, so no signal is transmitted. As the device turns on, an electrically-conductive channel forms between the two halves, as charged particles travel through the insulator and form a bridge for current to flow when a signal is applied. The size of this channel determines the size of the response (a larger channel lets current flow more easily, producing a stronger response), giving memristors their characteristic non-binary behavior. Materials that easily form conductive channels, like copper, tend to prevent the channels from ever going away when the signal is removed, so the device never turns off. In contrast, using materials that don’t form channels as easily, like silver, leads to an unreliable signal and a lot of electrical noise.
In their paper, the engineers at MIT explored several different pure and alloyed metals as the channel-forming material, and found a copper-silver combination to work the best in terms of stability over time, linearity (so an increase in voltage leads to an easily predictable increase in conductivity), and low variance. They also tested applications of their new devices, such as the ability to store an image for an extended time period and to perform tasks normally achieved by artificial intelligence algorithms, like image sharpening or edge detection.
Such a novel device could have widespread applications in computing, but nowhere moreso than in artificial intelligence. After all, machine learning software seeks to train computers to learn like the brain, but they’re all written on devices that only think in terms of on or off. Instead of trying to build the brain in software, memristor-based computing could allow for the hardware itself to do the hard part. This could potentially solve some of the most difficult challenges facing the widespread adoption of AI, such as its high energy cost and the normally massive datasets needed to properly train the algorithms. The future of image processing for medicine, molecular design, or self-driving cars may rely on these new electrical devices. And it’s possible that they may even help us understand how our own brains recognize patterns when given new information.
Hanwool Yeon and Peng Lin are post-doctoral associates in the Department of Mechanical Engineering at MIT. Chanyeol Choi is a graduate student at MIT’s Department of Electrical Engineering and Computer Science. Dr. Jeehwan Kim is an associate professor of Mechanical Engineering and Materials Science and Engineering at MIT and is the Principal Investigator of the Research Laboratory of Electronics.
Managing Correspondent: Andrew T. Sullivan
Press Articles: “Ag- and Cu-based memristor alloys on silicon as artificial synapses,” eeNews Europe
Original Journal Article: “Alloying conducting channels for reliable neuromorphic computing,” Nature
Image Credit: Pixabay