What Is Intelligence?

This may seem to be more of a question for psychologists than physicists. But two researchers (Joseph Wakeling, jwakeling@webdrake.net, now at the University of Fribourg, Switzerland, and Per Bak, Imperial College, 011-44-20-7594-8528, p.bak@ic.ac.uk) argue that intelligence is not an abstract concept, but must be considered as a physical phenomenon.

Any definition of intelligence, they say, cannot ignore a living being's environment, including its very own body. In their view, an organism is only intelligent relative to how well it solves the problems that its surroundings throw at it.

This runs counter to many historical ideas, including the concept that the mind is separate from the body, or that it is possible to build a desktop computer that thinks like a human without having the same physical environment or body.

To explore the idea of intelligence, the researchers ran computer simulations of artificial neural networks called "minibrains." In the simulations, 251 minibrains each attempted to pick the less popular of two choices, 0 and 1, analogous to 251 motorists all trying to pick the less congested road. This "Minority Game" would be repeated over many successive rounds.

Each minibrain consisted of three layers of "neurons": "input neurons," which dictated how many past rounds it could remember, leading to an intermediary layer, which then led into an "output" layer that determined what choice was made. If the minibrain ending up making an incorrect choice, it would reduce the strength of the connections between neurons supplying the "wrong answer."

The researchers were in for a surprise when they endowed all of the minibrains with equal abilities, which would be analogous to a bunch of motorists having the same amount of decision-making skill. In this situation, no minibrains correctly guessed the minority choice with even a 50 percent success rate, which is what you'd get by making the choice with a random flip of a coin. Even an E. coli bacterium, which searches for glucose by moving in random directions in its environment, is seemingly more intelligent than this. Only when the researchers introduced a "rogue" minibrain with more intermediate neurons to analyze the past rounds did it attain more than a 50 percent success rate. Their simulations suggest that intelligence often hinges on how much one can make use of the data in its physical environment. (Wakeling and Bak, Physical Review E, November 2001.)