Robot To Demonstrate Evolvable Neural Network Hardware

January 15, 1999

By Jon Bashor, jbashor@lbl.gov

Within a couple of years, a robotic kitten controlled by an artificial brain should be cavorting across a laboratory floor and hopefully capturing the imagination of the public, according to computer scientist Hugo de Garis of Advanced Telecommunications Research, a research laboratory in Kyoto, Japan. De Garis was a visiting guest in the Lab's National Energy Research Scientific Computing Division this month.

If all goes according to plan, de Garis says, this brain will be composed of 40 million artificial neurons, consisting of up to 32,000 neural net modules. The kitten robotó"Robokoneko"ówill be controlled by an on-body radio link to the artificial brain of some 1.25 Gigabytes of RAM. By updating the whole brain 300 times a second, researchers hope to achieve real-time control of the robot.

Remarkably, the circuits of neurons making up the brain will be "grown" by a specially designed machine called the CAM (cellular automata) Brain Machine, or CBM. The first CBM is expected to be completed in about two months, according to de Garis.

His work caught the attention of Lab computing scientists when he spoke at an annual DOE meeting on computing last April.

"When we present the machine, the project will move from science fiction to science fact," he said.

The machine, currently being built by a team in Colorado, is expected to be the world's first significant example of evolvable hardware, an idea de Garis had in 1992.

"If it works, it will blow away the traditional approach to neural networks of essentially using a single net with tens of neurons."

Not only will the brain machine randomly grow the neural networks on computer chips, but it will also test the performance of each network. "Then it will select those that perform well and get rid of the rest," de Garis says. "It's really survival of the fittest in electronics."

Copies will be made of the "keeper" circuits and mutated slightly by the brain machine, with the offspring then tested again. While most of these mutant networks will not perform as well, perhaps one percent will be better and will survive the machine's culling. "After hundreds of generations, we should end up with circuits which do quite well," de Garis says, adding that the brain machine will take about a second to evolve a functional neural network. He calls the entire process "evolutionary engineering."

Once enough of these "elite" networks have been created, they will be downloaded into a huge block of RAM as modules. The brain for the robo-kitten will contain 32,000 such modules. "My task for this year is to come up with the brain architectures, starting with one module, then 10, 50 and up to 32,000."

The true test, de Garis says, will come with the behavior of the kitten. "We want to make it as kitten-like as possibleócute and playful," he says. "If the kitten can keep an average person amused for half an hour, I'll have succeeded."

Currently, research into "brain building" is relatively small and hardly mainstream, said de Garis, adding, "The lab I'm working in is a bit fringey." But if the ideas prove successful, he predicts the field will take off and could eventually take on the momentum of other efforts conducted at national laboratories or at NASA. This could then lead to such new job titles as "brain architect" and "evolutionary engineer."

"It's getting close enough so that people will begin to take it seriously," he said.

More information can be found on de Garis' web site at http://www.hip.atr.co.jp/~degaris/.