IBM’s Cat-Brain Breakthrough
The supercomputers being lauded at middle month of November 2009, at the SC09 high-performance computing conference in Portland, Ore., can perform tasks in hours that would take a powerful desktop computer thousands of years. But when it comes to pattern recognition, perception and other seemingly simple tasks, their computing powers pale in comparison to a three-pound computer that fills less space than a two-liter bottle and uses less power than a light bulb: your brain.
Scientists can’t understand, let alone replicate, many of the brain’s abilities. But if a supercomputer still can’t mimic an organic central nervous system, IBM hopes it can at least model one.
The tech giant plans to announce at the SC09 conference a joint project with researchers from five universities and the Lawrence Berkeley National Laboratory that it calls “Blue Matter,” a software platform for neuroscience modeling. Pulling together archived magnetic resonance imaging (MRI) scan data and assembling it on a Blue Gene P Supercomputer, IBM has simulated a brain with 1 billion neurons and 10 trillion synapses–about the equivalent of a cat’s cortex, or 4.5% of a human brain.
“This is a tool of unprecedented scale,” says Dharmendra Modha, IBM’s research lead on the project. “It allows us to probe deeper into how the brain works and how we could build something like it, and it could lead to new dynamics for computing that we’ve been pursuing for more than 60 years.”
The project, partly funded by $20 million from the Department of Defense’s research arm, aims to mine the brain for tricks that could be used to mimic its cognitive abilities and its ultra-low energy use in future chips and software. The brain’s network of neurons and their connections known as synapses, Modha points out, manages many tasks that computers still struggle with, including dealing with ambiguous data, recognizing patterns and sifting through an overwhelming load of real-world sensory information to find and focus on relevant aspects. Hence, humans’ ability to perform tasks like driving in low-visibility conditions or reading unfamiliar fonts are tricks that still elude even the world’s most advanced computers.
That ability to parse enormous streams of data in real time also aligns with IBM’s so-called “smarter planet” initiative, a method of integrating sensors into infrastructure and analyzing the data they produce to optimize systems like the electrical grid, water systems and traffic. The brain is proof that this kind of computer can be built.
IBM’s Blue Matter model will run on a Blue Gene supercomputer that uses more than 144 terabytes of memory and a cluster of around 150,000 processors, capable of about half a petaflop of processing power, or 500 trillion floating point operations a second. At that rate, it could perform in about eight hours the same work that would take a typical Intel powered laptop around 500 years.
Even with that kind of hardware, IBM’s brain-like computer is still years, if not decades, away from reality. Jim Olds, the director of George Mason University’s Krasnow Institute for Advanced Study, points out that even limiting itself to a cat brain, IBM’s simulation still only captures about 10% to 20% of the relevant information in the neurons it simulates. That could be a major hurdle, given that neuroscientists don’t yet know which parts of the brain’s physiology may turn out to be key to its function. “There are unknown unknowns,” says Olds. “We collect so much data, but this could provide the overarching theory we need to make sense of it,” he says. “We don’t have our Einstein yet.
