IBM achieves accurate brain simulation
December 2nd 2009 00:05
Signature Tune
R. Colin Johnson
EE Times
(11/30/2009 8:42 PM EST)
PORTLAND, Ore.—IBM claims to have succeeded in creating a supercomputer simulation of the human brain capable of sensation, perception, action, interaction and cognition. So far it can only emulate the number of neurons and synapses in a cat's brain—1 billion spiking neurons and 10 trillion individual learning synapses— but then again the project has only been underway for one year.
IBM's brain algorithm, called Blue Matter, is part of the Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE) initiative sponsored by the a Defense Advanced Research Projects Agency (Darpa). IBM and its university partners, which include Stanford University, University of Wisconsin-Madison, Cornell University, Columbia University Medical Center and University of California-Merced, recently received $16.1 million for the next phase of the project.
The effort is being led by IBM manager of cognitive computing, Dharmendra Modha, who claims that the team's strategy is to take a middle up/down approach, rather than the top-down approach taken by neural network researchers a decade ago, or the bottom-up approach favored by biologists.
"The top-down approach postulates a problem to be solved, then designs a brain-like network to solve that specific problem, while the bottom-up approach is too detailed for even the fastest supercomputers today," said Modha. "So we are taking the middle road by using magnetic resonance diffusion weighted imaging data that has already been collected by other researchers."
The goal is to build a cognitive computing chip that rivals the brain's low energy consumption (about 20 watts) and its compact size, but so far its simulations are running on one of the largest supercomputer in the world, the Lawrence Berkeley National Laboratory IBM Dawn Blue Gene/P with 147,456 processors and 144 terabytes of main memory.
The cognitive computing chip will duplicate the brain's computational units, neurons and synapses, using mixed-signal analog-digital, asynchronous, parallel, distributed, reconfigurable, specialized and fault-tolerant algorithms that only update artificial brain states when information changes.
| 24 |
| Vote |
Add To: 
subscribe to this blog
