Science, Vol 304, Issue 5676, 1450-1451 , 4 June 2004
Summary of this Article
PDF Version of this Article
dEbates: ???? a response to this article

-----------------------------------------------------------------------------What Is Thought?
by Eric B. Baum
MIT Press, Cambridge, MA, 2004. 492 pp. $40, £25.95. ISBN 0-262-02548-5.

--------------------------------------------------------------------------------

A generation of prize-winning biologists, including Francis Crick and James Watson, have written about the inspiration they found in Erwin Schrödinger's What Is Life? (1)--a brilliant view of biology written from the outside, by one of the 20th century's leading physicists. Eric Baum's What Is Thought?, consciously patterned after Schrödinger's book, represents a computer scientist's look at the mind.
Baum is an unrepentant physicalist. He announces from the outset that he believes that the mind can be understood as a computer program. Much as Schrödinger aimed to ground the understanding of life in well-understood principles of physics, Baum aims to ground the understanding of thought in well-understood principles of computation.

In a book that is admirable as much for its candor as its ambition, Baum lays out much of what is special about the mind by taking readers on a guided tour of the successes and failures in the two fields closest to his own research: artificial intelligence and neural networks. To date, most work in those fields has fallen between two stools. Advocates of what the philosopher John Haugeland famously characterized as GOFAI (good old-fashioned artificial intelligence) create hand-crafted intricate models that are often powerful yet too brittle to be used in the real world. A classic example is Roger Schank's ????-understanding system, which is able to answer subtle questions about stereotyped interaction in a restaurant yet is easily fooled by even tiny changes in the scenario. At the opposite extreme are researchers working within the field of neural networks, most of whom eschew built-in structure almost entirely and rely instead on statistical techniques that extract regularities from the world on the basis of massive experience. But as Baum notes, experience can't be the whole answer: some of the most important "experience" in shaping thought comes not in the lifetime of an individual organism, but ancestrally--by way of evolution.

Too hard to program. Computer programs can compete with the best players in chess, but go still eludes their grasp. What makes the mind's program special?
CREDIT: JIM SUGAR/CORBIS

At heart, Baum makes five central claims: (i) The mind is modular, constructed of subcomponents specialized for particular computations. (ii) The initial structure of the mind is a product of the DNA. (iii) That DNA has been shaped by evolution. (iv) Mental representations are like scientific theories: the more parsimonious the better. (v) Meaning, a crucial prerequisite for thought, arises from parsimonious rede????ions.
Readers of Steve Pinker (or, for that matter, my own work) will surely recognize the first three claims. The fourth and fifth owe something to recent developments in computer science, wherein researchers have increasingly tried to formalize what it takes to make a maximally parsimonious theory. As Baum notes, if you can redescribe the world in an efficient way, there is a pretty good chance that your theory is on the right track.

Extending that notion to the mind, Baum suggests that computer programs--and by extension, minds--acquire meaning (only) when they redescribe the world compactly. By analogy, Baum points to the story of Tycho Brahe and Johannes Kepler. Brahe recorded detailed data about the locations of the planets, and Kepler developed a compact set of equations that could predict the locations of those planets. Baum's notion is that a computer program that mimicked Brahe wouldn't be meaningful, but that one that mimicked Kepler would. But would a skeptic who doubts that computer programs can think view a Kepler program as any closer to true understanding? I doubt it. More parsimonious programs may be better programs, but that doesn't make them genuinely meaningful. Meaning isn't about how compact some mental representation is but about how representations relate to the world. When it comes to understanding meaning, parsimony seems like the wrong tool for the job. (Furthermore, although compact rede????ions of the world are nice work when you can get them, it is far from obvious that we really will find them in the wetware of the brain. As a vast experimental literature shows, our intuitive theories are often a motley collection of half-truths that get us through the average day rather than true rede????ions of the world.)

That said, the book covers an enormous range, expanding from computer science into fields as diverse as economics and animal behavior, all in language that could be understood by an advanced undergraduate. Though the lay reader might be daunted at times by all the talk of complexity theory and Bayes's theorem, few other books that teach the basics of computer science range so broadly.

If there is a weakness here, it is that the approach that worked for Schrödinger may not work for Baum. Schrödinger, a physicist meditating on how biology ought to be, worked from the outside without reading that much about the actual stuff of biology; Baum does much the same for cognitive science. But whereas Schrödinger could be excused--in his day, nobody yet knew much about the exact biological basis of life---Baum's book suffers to some extent from his outsider status; huge, directly relevant literatures are scarcely mentioned. Researchers from Elizabeth Spelke and Susan Carey to Alan Leslie, Frank Keil, and Stanislas Dehaene--scientists who have spent their careers ferreting out the innate complexities of the human mind--are never mentioned, and even Pinker's well-known How the Mind Works (2), which reaches many of the same conclusions concerning the importance of evolution and innate structure, is not cited; similarly absent are cognitive scientists like Elizabeth Bates and Michael Tanenhaus who have challenged the modularity hypothesis. Nevertheless, a look from the outside is healthy for any field, and Baum clearly shares with Schrödinger an abiding concern for the big picture. If What Is Thought? can inspire a generation of computer scientists to inquire anew about the nature of thought, it will be a valuable contribution indeed.

References

E. Schrödinger, What Is Life? (Cambridge Univ. Press, Cambridge, 1944).
S. Pinker, How the Mind Works (Norton, New York, 1997).