This month, while researching neural networks for the Web-site column (see p. 45), I came across Genobyte (Boulder, CO), a company whose long-term aim is to build artificial brains, specifically a billion-neuron artificial brain by 2001. The company plans to do this using neural networks that "evolve" using genetic algorithms "grown" in field-programmable gate arrays from Xilinx (San Jose, CA).
According to Genobyte, such neural circuits can be grown and evaluated in hardware in microseconds, making possible a complete run of a genetic algorithm with tens of thousands of circuit growths and fitness measurements in less than a second. As an intermediate step in the development of the billion-neuron brain, an artificial brain with 32,000 neural network modules with up to 1152 neurons called a CBM (CAM-Brain Machine) will be built this year under contract from Advanced Telecommunications Research (ATR; Kyoto, Japan). This brain will control the behavior of a kitten robot called Robokoneko (Japanese for robot child cat) and will be built by Michael Korkin of Genobyte.
Software simulation
Initially Robokoneko will be simulated in software, using Working Model 3D, a physical-reality-simulation tool from MSC Working Knowledge (San Mateo, CA). The robot cat will be controlled using an object-linking-and-embedding automation link by a simulator software package developed at Genobyte that serves as an interface between the CBM and the Robokoneko.
"The genetic algorithm used to develop motion behavior of the robot requires tens of thousands of iterations to evolve simple motions of walking, turning, and sitting," says Hugo de Garis, head of the Brain Builder Group at ATR. "A real hardware robot would not be sufficiently reliable for this task," he says. Simulator software also allows automatic measurement and fitness evaluation of each of the thousands of iterations without human intervention.
Genetic algorithms operate on a population of the electronic equivalent of "chromosomes," which represent neural networks of different topologies and functions. Better performers for a particular function are selected and further reproduced using chromosome recombination and mutation. After hundreds of generations, this approach produces very complex neural networks with a desired functionality. According to Genobyte, this evolutionary approach can create a complex functionality without any previous knowledge about how to achieve it, as long as the desired input/output function is known.
After initial motion behaviors are evolved, a hardware robot will be built with vision, hearing, and other systems, interfaced to and controlled directly by the CBM's evolved neural modules. Because physical simulation speed is considerably lower than CBM speed, an array of PCs on a LAN could evaluate one kitten from a population of kittens to speed up evolution.
Evolvable hardware
"If the CBM is successful, it will be revolutionary because it will be the world's first significant example of functional-level, as opposed to gate-level, evolvable hardware," says de Garis. "It will replace traditional approaches to neural networks that use single nets with tens of neurons."
Instead, the CBM will allow tens of thousands of modules to be assembled into artificial brains with tens of millions of artificial neurons. "If the CBM and Robokoneko are successful in the next few years, then the CAM-Brain Project will create a new field that de Garis expects to be a trillion-dollar industry in 20 years.
Andy Wilson
Editor at Large
[email protected]
