Knowledge - Interview

Interview with Thomas Christaller

In a recent edition of Nature, scientists reported that they had built a robot capable of conducting its own research. Any chance you’ll be replaced by a machine one of these days?

Christaller: Definitely not. My colleagues have merely combined two elements: standard equipment that performs automated genetic experiments and an expert system that simulates the activities of scientists with this equipment. It’s a technology that’s been around for 30 years. Compared with doctoral candidates, computer simulation delivers better results. But that’s not very surprising, because rules for planning experiments are implicitly stored in the expert system. It would really be going out on a limb to label this system a robot scientist. This work would have been rejected as trivial by every scholarly journal for artificial intelligence (AI).

Well, AI researchers are notorious for going out on limbs with their predictions. Fifty years ago, people were saying that we would have intelligent machines today.

Christaller: That’s true, unfortunately.

Why is it that we can build chess computers capable of beating Grand Masters but can’t build robots that can find their way around in a strange environment?

Christaller: Chess is a game with rules that can be formalized and are easy to program—but our environment isn’t. As the behavior of individuals in their surroundings becomes less fixed in the genome, members of that species become less predictable. This forces individuals to put themselves in the position of others. I call this type of behavior trial behavior and believe it is the key to intelligence. Our brain is a powerful simulation machine.

When will the trial-behavior robot appear on the scene?

Christaller: I would like to be there when it’s developed. But I’m afraid it will take a generation—say 30 years. To get there, we’ll have to radically change our way of thinking.

What you’re saying is that structured knowledge—as in a database—is easy to process, but machines can’t handle everyday knowledge?

Christaller: That’s right. An encyclopedia is a collection of knowledge—but it’s not intelligent. To cope with life, a robot has to possess a lot of everyday knowledge. It must know that a chair has four legs and can be moved and carried around, while realizing that this might not be true of a cupboard. We can build robots that never run into anything, thanks to their sensors. But to recognize a chair as a chair and know what it’s used for? Not a chance.

Has a return to thinking about the abilities of the brain—the key word here is neural networks—moved AI forward?

Christaller: Even though no one has succeeded so far in making a model of intelligence using neural networks, this seems to be the most promising way forward. Traditional AI, which works with symbols, functions for codified knowledge. But if you want to create human-like behavior, you have to learn from the brain. Its complexity is created to a large extent by countless oscillators—positive feedback neural networks.

Will machines ever be able to draw conclusions in the way humans do?

Christaller: No. I’m not aware of a single convincing argument to support that premise. All previous machine solutions that outperform humans are based on different principles from those we find in the brain. If we define intelligence as the ability to make predictions about human behavior, then the chances are good that humans will be better than any artificial system. But computers will play an important role in the processing of information that can be formalized. There are no theoretical boundaries here, just technical and practical ones—like the fact that knowledge doesn’t stand still. As a result, I think that knowledge systems will require much more maintenance than normal databases do.

Will there one day be machines that have feelings?

Christaller: Feelings didn’t arise without a reason. They allow us to boil down the huge store of possible alternative actions to a very few and to write off the rest—on the basis of subjective perceptions. Otherwise, our pondering would never come to an end. If we want to build robots that function according to principles similar to those that determine how we behave, then they must have a world in which feelings have a purpose for them. But today’s robots don’t need feelings.

Rodney Brooks of the Massachusetts Institute of Technology says that the differences between humans and machines will gradually disappear over the next 50 years. Do you agree with him?

Christaller: No. That’s a strange prediction.

Interview conducted by Jeanne Rubner