Bioengineering Professor Challenges Blue Gene/L, Mimics Brain on Chip

Remember Deep Blue, the IBM supercomputer that beat the reigning World Chess Champion Garry Kasparov back in 1997? It only made it through the tournament by sheer brute force, evaluating some 200 million potential moves a second, while Kasparov only considered about three. It's not obvious Deep Blue would beat the Champ a second time; human brains are much better than computers at performing specific tasks such as vision, hearing, pattern recognition, and learning. One reason why human brains excel is that different areas in the brain are highly specialized and cleverly wired to the sensors feeding them.

This fact inspired Kwabena Boahen, an associate professor in the Department of Bioengineering, Stanford University, to explore ways of how to improve computer performance through mimicking brains functionality.

The brain's method of executing the likes of digital instructions is to chemically activate links between neurons, so called synapses. The transmission of one such signal takes a sluggish thousandth of a second. But through every second the brain activates 10 quadrillion (10^16) of those connections, and does so by using only 10 watts of power.

"A computer as powerful as the human brain would require around a million high-end general-purpose Intel Pentium Processors and a gigawatt of power to juice them up. Only to simulate the activity of 4000 neurons during one second would take 20 minutes for a nice PC. To do it in real time for one million neurons would require a capacity of 500 Teraflops", says Professor Boahen.

Such a computer simply doesn't exist-- yet. IBM?s Blue Gene/L-- presently the fastest computer on the globe--'only' sustains just over 280 teraflops with its 131,000 processors. And it's about half the size of a tennis court and consumes about 1.6 megawatts too.

However, Boahen claims that, within the next couple of years, he and his colleagues will build a brain in silicate that can beat the world?s fastest supercomputers in every way worth mentioning: smartness, speed, size and power consumption. That's quite sturdy talk, but cocky as it sounds it has a firm link with reality. The team has already developed key components of their brain--silicon retinas to restore vision.

And they have also developed circuits that can learn: In response to changing stimuli or circumstances, the circuits can form new connections to their silicon brain's vision neighborhood. Boahen calls this the "cells-that-fire-together-wire-together"-chip.

"It actually works, it isn't science fiction. Thanks to the silicon retina we know how to model neurons and synapses; we can do that for the fly brain already. That is about 100,000 neurons and from them we're trying to make a multiple-chip network that gets up to about one million neurons. This will enable us to model what the different cortical areas are doing and how they are talking to each other."

"If we could pave the way for implants to help people who suffered from a stroke, for instance, to restore their brain's function, that would really be a great achievement", Boahen says. But he doesn't believe in the technology bringing sci-fi to life. For instance, you won't be able to pick up French just by hooking yourself up to a memory stick in the future. "The knowledge of a language is complex," he explains. "Many areas of the brain are at work both when you learn and use it."

"But think about it. It's no wonder Google or Yahoo Search is text based. Today there is no well developed, widely used way for computers to recognize actual pictures and patterns, or for that matter to make queries for such entities. The person who comes up with the solution will become very rich," Boahen says.

But we will have to wait for at least another 30 years before we're going to see a silicon brain as clever as one of human tissue. And the business model for such a machine is still to be figured out. For example: How much and for what purpose should you train it before you sell it?