Miércoles 9 de Agosto de 2006, Ip nº 165

'Brainbox' Computer Mimics Human Brain
Por Tracy Staedter

A computer with thousands of microprocessors is being built to mimic and model the function of millions of nerve networks in the brain.

The Spinnaker — short for "spiking neural network architecture" — system will not only help scientists better understand the complex interactions of brain cells, but it could also lead to fault-tolerant computers that, like the brain, work despite malfunctions in tiny circuits.

"You lose one neuron per second during your adult life. As they die, there doesn't seem to be any gross underperformance in the brain," said Steve Furber of the University of Manchester in the U.K., leader of the Spinnaker project.

The idea, said Furber, is to mimic that kind of biological robustness in components of future electronic devices — which, as they inevitably shrink to smaller and smaller sizes, are likely to experience more and more failures.

But understanding how brains achieve such resilience is still a great mystery.

Scientists frequently use technologies such as functional magnetic resonance (commonly known as MRI) to image regions of the brain, and can probe to acquire an even finer picture of specific cellular networks.

But nerve cells are so tiny and numerous that pinpointing neural networks responsible for a particular activity is nearly impossible.

Pinpointing them on a computer brain should be much easier.

Furber's electronic version will contain silicon chips equipped with 20 tiny processors each, 19 of which will be designed to behave as neurons. The other will monitor the activity of the chip.

Each processor will model about a thousand neurons, so each chip will represent about 20,000 neurons. (The brain has hundreds of billions.) The system will replicate how signals propagate to accomplish specific tasks.

For example, the computer brain could duplicate what happens in the human brain when the eye sees an object — in other words, how an array of interconnected neurons in the cortex generates an image from electrical pulses passing down the eye's optic nerve.

Getting processors on computer chips to behave like neurons is no easy task, said Jim Austin, professor of neural computation at the University of York in the U.K.

Normally, computer chips have tiny clocks that help synchronize the computing power of the processors. But biological neurons function without a clock.

Furber's approach also eliminates the clock. Instead, each spike from the artificial neuron sends a packet of information to a selected location in the communication network, where it initiates additional neural spikes.

"This is an attempt to simulate the brain at a more realistic level," said Austin.

Furber thinks his team is about two years away from a functioning system that contains 50 chips and a thousand processors. Once they successfully demonstrate the computer, they plan to build a bigger machine with more power.

  02/08/2006. Discover Magazine.