How Much Power Does It Take To Simulate The Human Brain?

Illustration for article titled How Much Power Does It Take To Simulate The Human Brain?

Kwabena Boahen, a computer scientist at Stanford University, believes that it would require 10 megawatts to power a processor as smart as the human brain. His new "Neurogrid" supercomputer might be able to do it on only 20 watts.

Advertisement

To put that in perspective, 10 megawatts is the kind of energy a small hydroelectric plant produces—20 watts is only enough juice to power up a weak light bulb. Amazingly, your physical brain runs on this minuscule amount of power, and it's not very efficient. However, embracing this inefficiency could be the key to creating computers that mimic the human brain.

It sounds cockamamy, but it is true. Scientists have found that the brain's 100 billion neurons are surprisingly unreliable. Their synapses fail to fire 30 percent to 90 percent of the time. Yet somehow the brain works. Some scientists even see neural noise as the key to human creativity. Boahen and a small group of scientists around the world hope to copy the brain's noisy calculations and spawn a new era of energy-efficient, intelligent computing. Neurogrid is the test to see if this approach can succeed.

Most modern supercomputers are the size of a refrigerator and devour $100,000 to $1 million of electricity per year. Boahen's Neurogrid will fit in a briefcase, run on the equivalent of a few D batteries, and yet, if all goes well, come close to keeping up with these Goliaths.

Advertisement

So far Boahen has managed to squeeze a million neurons onto his new supercomputer compared to only 45,000 last year. By 2011, he hopes to have 64 million up and running, bringing the project to the equivalent of a mouse's brain.

Ditching reliability and efficiency in favor of organized chaos flies in the face of everything that an engineer holds dear, but the approach does make sense—and reducing the power consumption is the key to upholding Moore's law. But how will this development change our perception of what an artificially intelligent robot might become? Instead of some cold, logical machine that can think for itself, we might end up with robots that are just as stupid and flawed as we are. In other words. it could be a robot on that episode of future Cops running through the bushes with no shirt on after trying to rob a convenience store with a plastic lightsaber. Think about it. [Discover Mag via PopSci]

Share This Story

Get our newsletter

DISCUSSION

People have been trying to build chips with large numbers of neurons for decades (this includes my last company). It's not the number of neurons that's the problem though. It's our lack of understanding of the organization and the training algorithms.

Traditional training algorithms like Back Propagation start to really grind as the network and pattern space grow and other modified approaches don't do that much better. I don't hear Boahen addressing this.

Building more neurons that lack the required organization and learning algorithms is like putting a human brain in a blender, turning it on for a few seconds and then declaring that the jug has the processing power of Albert Einstein!

DARPA has been sinking money in to this kind of work for decades with very little to show for it (although not as much as they've wasted on "traditional" AI). A lot of the groups that continue to receive funding make extremely bold claims, but they can't substantiate any of them (a certain group at MIT springs to mind).

In case you think I'm just talking out of my rear, I say all this having worked on a DARPA Neuromorphic Computing project until a couple of years ago. I also originally started my Ph.D. in the implementation of neurons on silicon chips before switching to another subject.

Of course this kind of work is important and should be funded because one decade a may lead to something. However, it should be able to exist and to have academic freedom (and funding) without having to make outrageous claims. #neurogrid