Rajit Manohar, associate dean for academic affairs at Cornell Tech and professor of electrical and computer engineering, pioneered the design methodology for a new IBM computer chip inspired by the human brain.
“After years of collaboration with IBM, we are now a step closer to building a computer similar to our brain,” Manohar said.
He is co-author on a paper in the journal Science that explains the “brain chip” – the first production-scale “neuromorphic” chip, with a design based on neuroscience. The breakthrough is featured on the journal’s Aug. 8 cover. Two of Manohar’s former students, Nabil Imam and Filipp Akopyan, now researchers at IBM, are also co-authors. The chip, called TrueNorth, is made of 5.4 billion transistors and 1 million “neurons” as computation elements. From robotics to computer vision, the entirely digital chip could revolutionize computing as the world knows it.
The idea of neuromorphic computing has been around since the late 1980s, Manohar said, but the TrueNorth’s computing power is unprecedented, he said. Manohar has been collaborating with IBM researchers on improving neuromorphic chip design since 2008.
Manohar’s expertise is in asynchronous circuits, which are logic circuits that use signals instead of clocks to indicate completion of an operation. Most digital devices use synchronous circuits, but Manohar’s asynchronous varieties use little power and are designed to integrate well into larger systems.
TrueNorth uses parallel processing – conducting many operations at once that rely not on the speed of the network, but on its complexity, more like a brain and less like a very fast assembly line. Parallel processing greatly lowers energy usage, too – according to IBM, the chip uses power equivalent to that of a hearing aid.
Because the chips can be tiled in a two-dimensional array, the design can be scaled easily, according to the paper. This makes it suitable for any application using complex neural networks, like object detection and classification – something today’s computer processors find difficult.
Today’s laptops are optimized for numerical computation but not pattern recognition.
The TrueNorth chips could be put in computers that can handle complex tasks like image processing and voice recognition. The design of the chip allows a computer, like a brain, to engage in pattern recognition – identifying a human, or a cat, or a bicycle – as easily as people can. “This chip won’t be good at running Excel,” Manohar said, but excellent for computing that involves classification and finding patterns in data.