LANL: Computer just might reveal the secrets of the brain
6/19/2008
Scientists know it's there, but speaking it and understanding it is a problem of monumental scale.
Nobody really knows how it works. All they know is that it absorbs shapes, light and colors and defines them for us as objects such as "beer," "burrito" or "green chile."
That may change in coming years though, as Roadrunner, the new extremely fast supercomputer at Los Alamos National Laboratory, starts the complex task of unlocking that language by simulating the brain's inner-workings, said Steven Brumby, a Los Alamos scientist.
Creating a model like that only became possible this month, when a test proved Roadrunner could operate at the petaflop scale — which is the speed definition of a computer that can do a million billion, or a quadrillion, calculations per second.
Roadrunner is the first machine to ever break that petaflop barrier. It hit 1.026 petaflops in early June, and less than a week later, the machine broke its own record, hitting 1.144 petaflops on a preliminary simulation of the visual cortex cooked up by Brumby, Louis Bettencourt and other lab scientists.
The petaflop scale is essential for scientists to make a realistic model of the billion or so neurons and trillions of synapses in the human visual cortex.
And with that model may come answers to some very deep questions about who we are as humans, and why we function the way we do.
"In terms of the brain, it's sort of like we're in a land before Isaac Newton," Brumby said. "We can see the brain doing things, and we can see how it functions. But we lack an understanding of the fundamental laws of how it does what it does."
Looking at a giraffe or a picture of a giraffe will trigger a certain chain of events in the brain. It first goes to a section of the visual cortex called "V1," where it's processed and moved along to another section called "V2," where even more processing happens. After that, it moves through a section called "V4," where it is processed again, Brumby said.
"Don't ask me what happened to V3," he added with a laugh. "But after it goes through 'V4' it goes to another region called 'IT,' which is located above your ear. And in part of that area, if you look at a giraffe anywhere in your field of vision, a certain set of cells will trigger."
The brain creates a series of cells like that with every visual object we define. But the set of cells the brain uses differs from person to person — although generally similar objects are mapped in similar areas, Brumby said.
"One of the interesting things in brain science is the first things they started studying were insect brains, and they're so small that the entire circuit layout of an insect brain is encoded in their DNA," Brumby said. "Mammal brains, though, like ours, are not designed that way. The DNA holds a generic blueprint. But after you're born and start observing the world, the brain starts to wire itself."
What that means is that everybody's brain is wired a little differently, and every brain stores information in different places. But the process of storing — or the way wires send data from one part of the brain to encode information in another area — is done by that underlying brain language that we're all born with in our DNA, he said.
The visual cortex is the largest and most complex part of the brain, which makes it an interesting place to begin modeling efforts that could reveal how that system works, he said.
"We're studying the visual cortex first to learn about how the brain works because it's the best-studied part of the brain," Brumby said. "We hope it can teach us the rules of the brain and how those neurons wire together to define objects. We suspect those rules are common throughout the brain, though. So if we can figure them out in the visual cortex, we may be able to better understand them elsewhere."
So far, scientists have only done a few test runs of part of the cortex while IBM was testing Roadrunner at its facility in Poughkeepsie, NY, but the Los Alamos team hopes to do many more brain simulations once the machine gets to the lab this fall, Bettencourt said.
"For us it's just very exciting," Bettencourt said. "It's a challenge to do new science at this scale, but it's also just like getting your first piece of candy."
If scientists can understand the visual cortex, and eventually the brain, it could help them build even faster, smarter computers. That's because the human brain is an organic computer far superior to anything we've built with technology, Bettencourt said.
"If you think about our visual system, we can look at a visual field and without really thinking, reduce things there into objects, then add qualities to those objects," Bettencourt said. "And we do that in a fraction of a second."
Teaching computers to see things like humans do has been especially tricky. Things like facial recognition software so far really haven't been very effective because computers have a hard time defining shapes.
"So in the best visual programs, the computer makes mistakes about 10 percent of the time," Bettencourt said. "That sounds good at first, but it's not so good. If you think about something like crossing the street several times, and 10 percent of the time you'd get run over by a car, you can understand that the rate isn't that good."
If scientists can understand the visual cortex, though, it's possible to build computers that can see like a human does, he said.
It's possible that full-scale brain models could also one day help doctors treat illnesses like phantom limb syndrome, dementia or possibly schizophrenia, because knowing that underlying language could help medical scientists learn how to reprogram the brain.
That sort of thing is still pretty far off in the future, though, Brumby said.
The petaflop scale is super-fast, but a model of just the visual cortex will push it to its limit.
Modeling the entire brain will probably require the exaflop scale — or a machine 1,000 times more powerful than Roadrunner, Brumby said.
"Maybe in 20 years we'll have the brute force computing power to model an entire brain," Brumby said. "But we still have to learn how to program it, and that's something we're just starting on now."
And it's likely, even after full scale models exist, there will still be many questions to answer, such as how emotions work, how we build memories or how we come to have a consciousness of ourselves, Brumby said.
"There are many, many mysteries of the mind," Brumby said. "The science of that is only just starting."
Contact Sue Vorenberg at 986-3072 or svorenberg@sfnewmexican.com.