The “brain on a chip” contains tens of thousands of memristors and is capable of reproducing crisp images and can later be used to handle complex AI computing. (Image Credit: Peng Lin, MIT)

 

Engineers at MIT have developed a “brain on a chip” that functions similarly to an advanced neural network without being connected to the internet, relying on a supercomputer, or requiring software download. The single chip, which is smaller than a piece of confetti, contains tens of thousands of artificial brain synapses known as memristors (memory transistors). The team published their research in the journal Nature Nanotechnology. Ultimately, the goal is to develop small and power-efficient devices that can handle complex AI computing.

 

The memristors are made of alloys of silver and copper but are also created using silicon. When the researchers ran the chip through numerous visual tasks, they discovered the chip could “remember” and repeatedly reproduce stored images in very high detail. These versions were much crisper and cleaner than other types of memristor designs made of unalloyed elements. 

 

To create memristors using the new alloy, the team fabricated a negative electrode out of silicone. They then produced a positive electrode by placing a small amount of copper on it, and then finally, added a layer of silver. The two electrodes were placed between an amorphous silicone medium. By using this approach, they were able to design a millimeter-square silicone chip containing tens of thousands of memristors.

 

To test out the chip, the researchers recreated a gray-scale image of the Captain America shield and then matched each pixel to a corresponding memristor in the chip. Afterward, the team altered the conductance of each memristor so that it was relative in strength to the pixel’s color. This resulted in the chip reproducing an exact crisp image of the shield, and it “remembered” the image, allowing it to reproduce it repeatedly.

 

The researchers then ran the chip through an image processing task and programmed the memristors to alter an image of MIT’s Killian Court in a number of ways. They discovered that by using these memristors, the image could be sharpened and blurred much better than current state-of-the-art circuits.

 

“We’re using artificial synapses to do real inference tests,” says Jeehwan Kim, associate professor of mechanical engineering at MIT. “We would like to develop this technology further to have larger-scale arrays to do image recognition tasks. Someday, you might be able to carry around artificial brains to do these kinds of tasks, without connecting to supercomputers, the internet, or the cloud.”

 

Have a story tip? Message me at: cabe(at)element14(dot)com

http://twitter.com/Cabe_Atwell