Model doesn't just calculate but turns decisions into actions
View the video
A new computer simulation of the brain can count, remember and gamble. And the system, called Spaun, performs these tasks in a way that’s eerily similar to how people do.
Short for Semantic Pointer Architecture Unified Network, Spaun is a crude approximation of the human brain. But scientists hope that the program and efforts like it could be a proving ground to test ideas about the brain.
Several groups of scientists have been racing to construct a realistic model of the human brain, or at least parts of it. What distinguishes Spaun from other attempts is that the model actually does something, says computational neuroscientist Christian Machens of the Champalimaud Centre for the Unknown in Lisbon, Portugal. At the end of an intense computational session, Spaun spits out instructions for a behavior, such as how to reproduce a number it’s been shown. “And of course, that’s why the brain is interesting,” Machens says. “That’s what makes it different from a plant.”
Like a digital Frankenstein’s monster, Spaun was cobbled together from bits and pieces of knowledge gleaned from years of basic brain research. The behavior of 2.5 million nerve cells in parts of the brain important for vision, memory, reasoning and other tasks forms the basis of the new system, says Chris Eliasmith of the University of Waterloo in Canada, coauthor of the study, which appears in the Nov. 30 Science.
Input takes the form of written or typed characters, which Spaun “sees” with its vision system. The incoming information flows through the system, bouncing to and from various brain areas as it gets compressed into clear directions. Then, Spaun makes a decision about what to do. Finally, the decision gets expanded into action — it generates precise instructions on how to write out an answer. Because of the size and complexity of the system, the process is slow — in Spaun’s world, one second of work takes two real hours of computations.
After Spaun was assembled, scientists threw eight tasks at it, some of which resembled IQ test puzzlers, like a complete-the-pattern quiz. Spaun was also asked to reason, memorize and even gamble. As Spaun worked through these jobs, some curiously human quirks emerged. Just like human volunteers, Spaun was better at remembering the first and last number in a series. Also like people, Spaun took longer to count to higher numbers. “That’s the kind of thing we couldn’t program in,” Eliasmith says. To him, that similarity suggests that Spaun is performing the tasks in a way similar to the human brain.
Others disagree. Henry Markram, who leads a different project to reconstruct the human brain called the Blue Brain, questions whether Spaun really captures human brain behavior. Because Spaun’s design ignores some important neural properties, it’s unlikely to reveal anything about the brain’s mechanics, says Markram, of the Swiss Federal Institute of Technology in Lausanne. “It is not a brain model.”
Unlike the human brain, Spaun can’t learn new things outside of the selected tasks, like figuring out how to read Arabic. But scientists don’t really understand how the human brain is so good at general learning, Machens says, so it’s understandable that Spaun can’t do it. The model captures what we know about the brain so far, he says. “In that sense, I think it’s a breakthrough.”
University of Waterloo scientist Chris Eliasmith describes the abilities of Spaun, a new computer simulation of the human brain, which is shown here performing a series of four tasks.
Credit: Chris Eliasmith et al.