Computer scientists have long sought to mimic the phenomenal computing powers of the human brain. Recently, Columbia neuroscientist Stefano Fusi has helped researchers come one step closer to doing just that.
Fusi, an associate professor of neuroscience at the Medical School, is part of the SyNAPSE Project, an effort to emulate the brain’s structure electronically on a computer, a field of research known as cognitive computing.
Fusi and other researchers on the project—a collaboration with other universities funded by the government's Defense Advanced Research Projects Agency—recently released prototype cognitive computing chips.
“It’s not a simulation,” Fusi said. “You’re not transforming your problem into a symbolic problem … you really replace every element that you have in the brain with an electronic element that behaves in the same way.”
Cognitive computing offers the promise of performing complex tasks—tasks that push the boundaries of modern supercomputers—while using a fraction of the power of an average PC.
According to Fusi, this potential is made possible by “neuromorphic” hardware. Rather than building ever speedier supercomputers to run strenuous mathematical simulations of the brain’s parallel processes, scientists can use neuromorphic computing to essentially copy the brain’s processing methods.
Most computers’ processors today, Fusi said, are linear, quickly performing computations one at a time. The brain, in contrast, relies on parallel computation—it can carry out many computations at once, because each of its billions of neurons has thousands of synaptic connections to other neurons, which can all carry communications simultaneously.
Neuromorphic computing is “a very different approach from traditional computers, because you have a lot of very simple units,” Fusi said. “You don’t have a complex central processing unit.”
These simple units function like the brain’s neurons, receiving electrical impulses from each other. In the brain, neurons send synapses—electrical signals—to each other, which the research group simulated with energy-efficient electrical devices called memresistors.
Still, Fusi noted that “you can simulate whatever you want” with a supercomputer.
“Biological elements are significantly slower, a million times slower” than a traditional processor, he said.
But the artificial brain, Fusi said, would be a modern supercomputer with complex, context-driven tasks, such as pattern recognition, driving a car, or recognizing an unknown environment.
A supercomputer would be “extremely inefficient” in comparison, he added, as ever-increasing processor speeds lead to increasingly problematic usage of electric power.
“One of the main advantages of this approach is power consumption,” he said. “It’s all about power consumption.”
The prototype chips released by IBM have only 256 “neurons”—there are about 100 billion in the human brain—but that still adds up to tens of thousands of connections. Fusi said IBM will soon release a system with “hundreds, if not thousands” of 256-neuron cores. Other research teams are also pursuing neuromorphic computing, including a team at Stanford University that has succeeded in creating a million-neuron prototype.
The IBM prototype, Fusi added, will emulate the brain’s ability to learn and even rewire itself.
“There is no programming basically,” Fusi said. “You design the system in such a way that it can learn autonomously.”
Fusi, a member of Columbia’s Center for Theoretical Neuroscience, was a theoretical physicist by training. Cognitive computing, he explained, is still a developing science.
“We’re still pioneers in this field,” he said.
A previous version of this article stated that the project was funded by IBM and other universities, rather than by the Defense Advanced Research Projects Agency. Spectator regrets the error.