Spaun, a new software model of a human brain, is able to play simple pattern games, draw what it sees and do a little mental arithmetic. It powers everything it does with 2.5 million virtual neurons, compared with a human brain’s 100 billion. But its mistakes, not its abilities, are what surprised its makers the most, said Chris Eliasmith, an engineer and neuroscientist at the University of Waterloo in Canada.
Ask Spaun a question, and it hesitates a moment before answering, pausing for about as long as humans do. Give Spaun a list of numbers to memorize, and it falters when the list gets too long. And Spaun is better at remembering the numbers at the beginning and end of a list than at recalling numbers in the middle, just like people are.
“There are some fairly subtle details of human behavior that the model does capture,” said Eliasmith, who led the development of Spaun, or the Semantic Pointer Architecture Unified Network. “It’s definitely not on the same scale [as a human brain],” he told TechNewsdaily. “It gives a flavor of a lot of different things brains can do.”
Eliasmith and his team of Waterloo neuroscientists say Spaun is the first model of a biological brain that performs tasks and has behaviors. Because it is able to do such a variety of things, Spaun could help scientists understand how humans do the same, Eliasmith said. In addition, other scientists could run simplified simulations of certain brain disorders or psychiatric drugs using Spaun, he said.
Researchers have made several brain models that are more powerful than Spaun. The Blue Brain model at the Ecole Polytechnique Fédérale de Lausanne in France has 1 million neurons. IBM’s SyNAPSE project has 1 billion neurons. Those models aren’t built to perform a variety of tasks, however, Eliasmith said.
Spaun is programmed to respond to eight types of requests, including copying what it sees, recognizing numbers written with different handwriting, answering questions about a series of numbers and finishing a pattern after seeing examples.
Spaun’s myriad skills could shed light on the flexible, variable human brain, which is able to use the same equipment to control typing, biking, driving, flying airplanes and countless other tasks, Eliasmith said. That knowledge, in turn, could help scientists add flexibility to robots or artificial intelligence, he said. Artificial intelligence now usually specializes in doing only one thing, such as tagging photos or playing chess. “It can’t figure out to switch between those things,” he said.
In addition, artificial intelligence isn’t built to mimic the cellular structure of human brains as closely as Spaun and other brain models do. Because Spaun runs more like a human brain, other researchers could use it to run health experiments that would be unethical in human study volunteers, Eliasmith said. He recently ran a test in which he killed off the neurons in a brain model at the same rate that neurons die in people as they age, to see how the dying off affected the model’s performance on an intelligence test.
Such tests would have to be just first steps in a longer experiment, Eliasmith said. The human brain is so much more complex than models that there’s a limit to how much models are able to tell researchers. As scientists continue to improve brain models, the models will become better proxies for health studies, he said.
There’s one major way Spaun differs from a human brain. It takes a lot of computing power to perform its little tasks. Spaun runs on a supercomputer at the University of Waterloo, and it takes the computer two hours to run just one second of a Spaun simulation, Eliasmith said.
So Eliasmith’s next major step for improving Spaun is developing hardware that lets the model work in real time. He’ll cooperate with researchers at the University of Manchester in the U.K. and hopes to have something ready in six months, he said.
In the far future, people may find Spaun’s humanlike flaws deliberately built into robot assistants, Eliasmith said. “Those kinds of features are important in a way because if we’re interacting with an agent and it has a kind of memory that we’re familiar with, it’ll more natural to interact with,” he added.
Eliasmith and his colleagues published their latest paper about Spaun today (Nov. 29) in the journal Science.
( via news.discovery.com)