[AI] An ambitious plan to build a computer that works like the brain

Sanjay ilovecold at gmail.com
Fri Aug 27 06:47:32 EDT 2010


Army of smartphone chips could emulate the human brain

An ambitious plan to build a computer that works like the brain

will use "bog-standard, off-the-shelf processors of fairly modest

performance"

by Paul Marks

IF YOU have a smartphone, you probably have a slice of Steve

Furber's brain in your pocket. By the time you read this, his

1-billion-neuron silicon brain will be in production at a microchip

plant in Taiwan.

Computer engineers have long wanted to copy the compact power of

biological brains. But the best mimics so far have been impractical,

being simulations running on supercomputers.

Furber, a computer scientist at the University of Manchester, UK, says

that if we want to use computers with even a fraction of a brain's

flexibility, we need to start with affordable, practical, low-power

components.

"We're using bog-standard, off-the-shelf processors of fairly modest

performance," he says.

Furber won't come close to copying every property of real neurons,

says Henry Markram, head of Blue Brain. This is IBM's attempt

to simulate a brain with unsurpassed accuracy on a Blue Gene

supercomputer at the Swiss Institute for Technology, Lausanne. "It's a

worthy aim, but brain-inspired chips can only produce brain-like

functions," he says.

That's good enough for Furber, who wants to start teaching his

brain-like computer about the world as soon as possible. His first

goal is to teach it how to control a robotic arm, before working

towards a design to control a humanoid. A robot controller with even a

dash of brain-like properties should be much better at tasks like

image recognition, navigation and decision-making, says Furber.

"Robots offer a natural, sensory environment for testing brain-like

computers," says Furber. "You can instantly tell if it is being

useful."

Called Spinnaker - for Spiking Neural Network Architecture - the

brain is based on a processor created in 1987 by Furber and colleagues

at Acorn Computers in Cambridge, UK, makers of the seminal BBC

Microcomputer.

Although the chip was made for a follow-up computer that flopped, the

ARM design at its heart lived on, becoming the most common "embedded"

processor in devices like e-book readers and smartphones.

But coaxing any computer into behaving like a brain is tough. Both

real neurons and computer circuits communicate using electrical

signals, but in biology the "wires" carrying them do not have fixed

roles as in electronics. The importance of a particular neural

connection, or synapse, varies as the network learns by balancing the

influence of the different signals being received. This synaptic

"weighting" must be dynamic in a silicon brain, too.

To coordinate its 'neurons' the chip mimics the way real neurons

communicate using 'spikes' in voltage

The chips under construction in Taiwan contain 20 ARM processor cores,

each modelling 1000 neurons. With 20,000 neurons per chip, 50,000

chips will be needed to reach the target of 1 billion neurons.

A memory chip next to each processor stores the changing synaptic

weights as simple numbers that represent the importance of a given

connection at any moment. Initially, those will be loaded from a PC,

but as the system gets bigger and smarter, says Furber, "the only

computer able to compute them will be the machine itself".

Another brain-like behaviour his chips need to master is to

communicate coordinated "spikes" of voltage. A computer has no trouble

matching the speed at which individual neurons spike - about 10 times

per second - but neurons work in very much larger, parallel groups

than silicon logic gates.

In a brain there is no top-down control to coordinate their actions

because the basic nature of individual neurons means that they work

together in an emergent, bottom-up way.

Spinnaker cannot mimic that property, so it relies on a miniature

controller to direct spike traffic, similar to one of the routers in

the internet's backbone. "We can route to more than 4 billion

neurons," says Furber, "many more than we need."

While the Manchester team await the arrival of their chips, they have

built a cut-down version with just 50 neurons and have put the

prototype through its paces in the lab. They have created a virtual

environment in which the silicon brain controls a Pac-Man-like program

that learns to hunt for a virtual doughnut.

"It shows that our four years designing the system haven't been

wasted," says Furber. He hopes to have a 10,000-processor version

working later this year.

As they attempt to coax brain-like behaviour from phone chips, others

are working with hardware which may have greater potential.

The Defense Advanced Research Projects Agency, the Pentagon's

research arm, is funding a project called Synapse. Wei Lu of the

University of Michigan at Ann Arbor, is working on a way of providing

synaptic weights with memristors, first made in 2008 (New

Scientist, 3 May 2008, p 26).

Handily, their most basic nature is brain-like: at any one moment a

memristor's resistance depends on the last voltage placed across it.

This rudimentary "memory" means that simple networks of memristors

form weighted connections like those of neurons. This memory remains

without drawing power, unlike the memory chips needed in Spinnaker.

"Memristors are pretty neat," says Lu.

Their downside is that they are untested, though. "Synapse is an

extremely ambitious project," says Furber. "But ambition is what

drives this field. No one knows the right way to go."


Technical telepathy: 09969636745
Saints are not always saints; sinners are not always sinners.
  



More information about the AccessIndia mailing list