Cutting edge

Evolvable hardware produced by Darwinian ideas of circuit formation is making the building of brains superior to ours a reality.

One of the great economic driving forces of our age is derived from "Moore's Law", which states that the performance of integrated circuits is doubling every year or so. It is a trend that has been valid for the past 30 years, is likely to continue for a further 20 and has recently made electronic circuits capable of performing feats only dreamt of a decade ago.

One of these is the programming of hardware by individuals. Software instructions enter the chip and tell it how to switch electronic pathways to form a desired circuit.

In summer 1992, I had the idea that if such a circuit could be reprogrammed again and again, then it might be possible for it to evolve. By sending in random software instructions that form random circuits, then measuring how well they perform, one could start applying Darwinian ideas to circuit formation.

In a population of such random circuits, those that by chance get a good score survive and those with a low score are eliminated. Instructions that produce high-scoring circuits are copied to form the next generation.

Random changes or mutations are then performed on these "children" and 99 per cent of the time these decrease the score, but occasionally a mutated child outperforms its parent.

In time, this superior variant will squeeze out inferior forms until it, too, is squeezed out by a superior mutant. After hundreds of generations electronic circuits with desired functions can be evolved quite successfully.

A new research field, "evolvable hardware", is investigating the potential of the above ideas. E-Hard techniques are being used to build an artificial brain at ATR Labs in Kyoto, Japan, and Starlab in Belgium.

In 1997 the "brain builder group" at ATR contracted an American firm, Genobyte, to design a brain-building machine, a CBM. This machine is capable of evolving, E-Hard style, a neural net circuit module of more than 1,000 neurons or artificial brain cells in about one second. The elite-evolved circuit is then downloaded into a billion-byte memory.

The function of each module is specified and evolved by a human evolutionary engineer. A brain architect then designs the whole brain, by specifying how the 64,000 modules comprising it are to be interconnected.

Once all the modules are downloaded into the memory and their interconnections specified, the CBM will then be used to generate the neural signalling of the brain, and will do this fast enough for real-time control of a robot kitten. This kitten will thus be controlled by the world's first artificial brain, containing 75 million artificial brain cells. With 64,000 modules to control it, "Robokitty" should have many behaviours and be very entertaining. It should also show that brain building is possible.

The challenge for the Starlab brain-builder group in 2000 will be to create an architecture for the artificial brain. Plans are also afoot to build a second-generation CBM, "CBM-2", able to handle a billion-neuron artificial brain.

With 21st-century technologies, it will be possible to store a bit of information on a single atom, which ultimately will allow artificial brains called "Artilects" to be built with trillions of trillions of trillions of times the computational capacity of the human brain. The issue that will dominate global politics in the next century will be whether humanity should build godlike brains that one day may decide to exterminate us.

Hugo de Garis is head of the Brain Builder Group at ATR Labs in Kyoto, Japan, until February 2000, when he will continue with the same work at Starlab, in Brussels, Belgium. He is also an adjunct professor in computer science at Utah State University in the US and at Wuhan University, China.