Blue Brain Power: Modeling the brain with a supercomputer

IBM project uses supercomputer to model the brain

The human brain has 100 billion neurons, nerve cells that enable us to adapt quickly to an immense array of stimuli. We use them to understand and respond to bright sunlight, a honking horn, the smell of chicken frying and anything else our sensors detect.

Neurons extending from the neocortex.
Neurons extending from the neocortex.

To better understand some of those responses, researchers in Lausanne, Switzerland, recently launched an ambitious project called Blue Brain, which uses IBM's eServer Blue Gene, a supercomputer capable of processing 22.8 trillion floating point operations per second (TFLOPS). Blue Brain is modeling the behavior of 10,000 highly complex neurons in rats' neocortical columns (NCC), which are very similar to the NCCs in a human brain. The NCCs run throughout the brain's gray matter and perform advanced computing. They are 0.5mm in diameter and 2mm to 5mm in height and are arranged like the cells of a honeycomb.

The first objective of Blue Brain is to build an accurate software replica, or template, of an NCC within two to three years, says Henry Markram, the principal researcher on Blue Brain and a professor at Ecole Polytechnique Federale de Lausanne (EPFL). That first template will be modified for NCCs found in different brain regions and species, and then all the NCCs will be replicated to build a model of the neocortices of different species, he says.

Such models will shed light on how memories are stored and retrieved, Markram says. "This could reveal many exciting aspects of the [brain] circuits, such as the form of memories, memory capacity and how memories are lost."

The modeling can help find vulnerabilities in the neocortex, which is useful because that's where brain disorders often originate. "We may also be able to work out the best way to compensate and repair circuit errors," Markram says. "The model could be used to develop and test treatment strategies for neurological and psychiatric diseases," such as autism, schizophrenia and depression, he adds.

Having an accurate computer-based model of the brain would mean that some major brain experiments could be done in silicon rather than in a "wet" lab. A simulation that might take seconds on the supercomputer could replace a full day's worth of lab research, Markram estimates. Ultimately, simulated results of brain activity could be matched with recorded brain activity in a person with a disease in order to "reverse-engineer" the circuit changes in diseases, he says.

The real value of a simulation is that researchers can have access to data for every single neuron, adds IBM's Charles Peck, head of the Blue Brain project for IBM Research.

Although science "knows a lot of details about the brain, we do not know how the parts fit together and how they are related to thought and learning and perception," he says.

Peck says Markram's team will take measurements from a dozen neurons that have been sliced from rat brains and mounted on a chip. The research will examine the physical structure and the electrical properties of each neuron and how neurons affect one another.

Peck says that a model of multiple NCCs is still far from a model of the whole brain. "Once we have modeled the neocortex, we will have to include models of other brain regions, such as the basal ganglia, hippocampus, cerebellum and so on," he explains.

The Blue Gene supercomputer was installed in July, and the first simulations were run in August, with a simulation of 25,000 simple neurons that took just 60 seconds. "This was just not possible before, and even smaller networks of 1,000 neurons would take weeks to run on a cluster, so this is truly a quantum leap in size and speed," Markram says. Future simulations of 10,000 complex neurons will take much longer.

A special room was built for Blue Gene at the EPFL, and the machine sits on top of a large room that holds the cooling equipment and computer cables, Markram says. Ice-cold water from Lake Geneva is pumped in to support the cooling system. The actual computer "takes up only a small space," Markram notes, and is only about the size of four refrigerators running on four racks.

Blue Gene is an 8,096-processor supercomputer, and it will model one to 10 neurons per processor. The computer could allow simulations of as many as 100 million simple neurons, which is about half the number of neurons in a rat brain. A PFLOPS Blue Gene, which IBM says is several years away, would make it possible to simulate nearly a billion simple neurons, Markram says. "But improvements in processing speed and memory could mean the entire human brain could be simulated within a decade," he adds.

IBM's Ajay Royyuru, head of Blue Gene computing as applied to life sciences, says the supercomputer's role is another indication that "biology [has become] information science."

"The scale of this computing will reveal interesting things in biology. We need that scale to get at the complexity that biological systems have," he says. And the trickle-down effect from Blue Brain to other computing projects in science and industry will be enormous, he adds.

For example, Markram says, ASIC designs that emulate neuron network behavior might be developed for use in information processing in intelligent devices. And, more generally, he says that Blue Brain will teach lessons about real-time data processing, as opposed to off-line processing.

"There's plenty... we can learn and bring back," Royyuru says of the Blue Brain project.

Wetware, Software and Hardware

Gray Brain

100 billion = neurons in a human brain

300 million = neurons in an octopus brain

1,300 grams = adult human brain weight

90 grams = adult rhesus monkey brain weight

Blue Brain

22.8 TFLOPS peak processing speed

8,096 CPUs at 700 MHz (downgraded to handle massive parallel processing)

Two processors and four floating-point units for a combined speed of 5.6 GFLOPS

256MB to 512MB memory per processor

Linux and C++ software

100 kilowatts power con-sumption

Blue Brain is a prototype, but production machines that are similar cost "hundreds of millions of dollars," IBM says.

Copyright © 2005 IDG Communications, Inc.

7 inconvenient truths about the hybrid work trend
Shop Tech Products at Amazon