Evolution of artificial intelligence

Terrence J. Sejnowski

Artificial Minds. By Stan Franklin. MIT Press: 1995. Pp. 449. $30, £22.95.

We are living in a postmodern age in which the functionalist architecture of artificial intelligence (Al), like the functional architecture of the Bauhaus, is beginning to look rather old fashioned. Stan Franklin takes the reader on a tour of the eclectic computational styles competing to replace the traditional symbol-processing paradigm that served as the unifying framework, and straitjacket, for a generation of mindmodellers. He leads the mindful tourist on a whistle-stop tour of artificial neural networks, computational neuroethology, situated action, artificial life, chaos theory, quantum computing, neural Darwinism and a host of other hopeful approaches. Where symbolic AI was once the only game in town, there now seems to be one on every street corner.

A trend that has evolved over the past decade is a shift away from the toy worlds of early AI towards problems that arise from interaction with the real world. The first computers available to AI researchers had less computing power than a pocket calculator. So the problems attacked were highly simplified. Because computers are efficient at symbol processing, solutions were sought in the realm of logic. The world, however, is noisy and uncertain and does not come labeled by symbols, which must be grounded in sensation and learned through experience. Inductive reasoning from incomplete knowledge is as important as deductive reasoning. Ironically, it is not logic but the theory of gaming that provides the mathematical framework for inductive reasoning in an uncertain world, and George Boole, in his 1854 book An Investigation into the Laws of Thought, which introduced Boolean logic, devoted the second half to the theory of probabilities.

Artificial Minds is organized around three debates in AI. The first is whether a computer can ever be programmed to think. This is a recurring question that does not seem to have a resolution, in part because the definition of thinking keeps changing. The problem has been revived recently by Roger Penrose, who doubts the possibility that algorithms could ever think like a mathematician.

I am highly skeptical of claims that something is impossible. A small change in an innocent-looking assumption can drastically change the conclusions. Some aspects of thinking have already been automated, but there are others, such as subjective feelings, that do not have an obvious mechanical analogue. There is a vast middle ground to explore and much of it can be approached with empirical studies of brains and the invention of ever-more clever devices.

The second Al debate, led by the connectionist alternative to symbol processing, concerns Jerome Feldman's 100-step rule. The brain can solve many difficult cognitive problems in half a second. If a neuron takes 5 milliseconds to perform one operation, then it should be possible to devise algorithms for solving the same problems that require only 100 steps. Although the SOAR program of the late Allen Newell respected this constraint, most AI programs take millions of steps; connectionist algorithms, by contrast, distribute the information over many simple processing units and work in parallel. The first-generation neural network models, however, were weakest where symbol processing is strongest, in representing relationships between abstract categories. The analysis of nonlinear networks as dynamical systems is still in its infancy and I suspect that the next generation of network models, based on spiking neurons, may provide additional computational abilities that may overcome this limitation.

The third debate reaches into the heart of AI and questions the need for representations at all. Rodney Brooks incorporates a hierarchy of reactive behaviours into walking machines that inhabit the Massachusetts Institute of Technology; the emphasis is on action selection rather than highly detailed sensory representations. Of course, as the control of actions becomes more sophisticated, the burden of intelligence is simply shifted from the sensory to the motor side. Franklin declares his own bias for a theory of intelligence based on an action-selection scheme, in which tasks are solved by a variety of specialized agents. These autonomous agents remind me of programs on the World Wide Web, such as 'spider' agents that seek out and catalogue information available at Web sites. If the digital computer served as a model of the mind for a generation of students in cognitive science, will the next generation be guided by their experience surfing on the Internet?

Artificial Minds is a readable guide to the evolution of ideas in artificial intelligence. It is intended for a general audience that already has some familiarity with symbolic AI and could be used, like a tour guide, to decide what areas of nouvelle Al to visit in greater depth. My only disappointment is that cognitive neuroscience, the science of real minds, plays only a small part in the book, perhaps because this field is of less interest to the author, whose background is in mathematics and computer science. New techniques for probing brains in action are uncovering unexpected insights into the organization of human minds. But that is another tour.

Terrence J. Sejnowski is at the Salk Institute, 10010 North Torrey Pines Road, La Jolla, Califomia 92037, USA.