Mind as Machine: A History of Cognitive Science. Margaret A. Boden. Two volumes, xlviii + 1631 pp. Oxford University Press, 2006. $225.
The term cognitive science, which gained currency in the last half of the 20th century, is used to refer to the study of cognition—cognitive structures and processes in the mind or brain, mostly in people rather than, say, rats or insects. Cognitive science in this sense has reflected a growing rejection of behaviorism in favor of the study of mind and "human information processing." The field includes the study of thinking, perception, emotion, creativity, language, consciousness and learning. Sometimes it has involved writing (or at least thinking about) computer programs that attempt to model mental processes or that provide tools such as spreadsheets, theorem provers, mathematical-equation solvers and engines for searching the Web. The programs might involve rules of inference or "productions," "mental models," connectionist "neural" networks or other sorts of parallel "constraint satisfaction" approaches. Cognitive science so understood includes cognitive neuroscience, artificial intelligence (AI), robotics and artificial life; conceptual, linguistic and moral development; and learning in humans, other animals and machines.
Among those sometimes identifying themselves as cognitive scientists are philosophers, computer scientists, psychologists, linguists, engineers, biologists, medical researchers and mathematicians. Some individual contributors to the field have had expertise in several of these more traditional disciplines. An excellent example is the philosopher, psychologist and computer scientist Margaret Boden, who founded the School of Cognitive and Computing Sciences at the University of Sussex and is the author of a number of books, including Artificial Intelligence and Natural Man (1977) and The Creative Mind (1990). Boden has been active in cognitive science pretty much from the start and has known many of the other central participants.
In her latest book, the lively and interesting Mind as Machine: A History of Cognitive Science, the relevant machine is usually a computer, and the cognitive science is usually concerned with the sort of cognition that can be exhibited by a computer. Boden does not discuss other aspects of the subject, broadly conceived, such as the "principles and parameters" approach in contemporary linguistics or the psychology of heuristics and biases. Furthermore, she also puts to one side such mainstream developments in computer science as data mining and statistical learning theory. In the preface she characterizes the book as an essay expressing her view of cognitive science as a whole, a "thumbnail sketch" meant to be "read entire" rather than "dipped into."
It is fortunate that Mind as Machine is highly readable, particularly because it contains 1,452 pages of text, divided into two very large volumes. Because the references and indices (which fill an additional 179 pages) are at the end of the second volume, readers will need to have it on hand as they make their way through the first. Given that together these tomes weigh more than 7 pounds, this is not light reading!
Boden's goal, she says, is to show how cognitive scientists have tried to find computational or informational answers to frequently asked questions about the mind—"what it is, what it does, how it works, how it evolved, and how it's even possible." How do our brains generate consciousness? Are animals or newborn babies conscious? Can machines be conscious? If not, why not? How is free will possible, or creativity? How are the brain and mind different? What counts as a language?
The first five chapters present the historical background of the field, delving into such topics as cybernetics and feedback, and discussing important figures such as René Descartes, Immanuel Kant, Charles Babbage, Alan Turing and John von Neumann, as well as Warren McCulloch and Walter Pitts, who in 1943 cowrote a paper on propositional calculus, Turing machines and neuronal synapses. Boden also goes into some detail about the situation in psychology and biology during the transition from behaviorism to cognitive science, which she characterizes as a revolution. The metaphor she employs is that of cognitive scientists entering the "house of Psychology," whose lodgers at the time included behaviorists, Freudians, Gestalt psychologists, Piagetians, ethologists and personality theorists.
Chapter 6 introduces the founding personalities of cognitive science from the 1950s. George A. Miller, the first information-theoretic psychologist, wrote the widely cited paper "The Magical Number Seven, Plus or Minus Two," in which he reported that, as a channel for processing information, the human mind is limited to about seven items at any given time; more information than that can be taken in only if items are grouped as "chunks." Jerome Bruner introduced a "New Look" in perception, taking it to be proactive rather than reactive. In A Study of Thinking (1956), Bruner and coauthors Jacqueline Goodnow and George Austin looked at the strategies people use to learn new concepts. Richard Gregory argued that even systems of artificial vision would be subject to visual illusions. Herbert Simon and Allen Newell developed a computer program for proving logic theorems. And Noam Chomsky provided a (very) partial generative grammar of English in Syntactic Structures (1957).
Two important meetings occurred in 1956, one lasting two months at Dartmouth and a shorter one at MIT. There was also a third meeting in 1958 in London. Soon after that, Miller, Eugene Galanter and Karl Pribram published an influential book, Plans and the Structure of Behavior (1960), and Bruner and Miller started a Center for Cognitive Studies at Harvard. These events were followed by anthologies, textbooks and journals. "Cognitive science was truly on its way."
In the remainder of Boden's treatment, individual chapters offer chronological accounts of particular aspects of the larger subject. So, chapter 7 offers an extensive discussion of computational psychology as it has evolved since 1960 in personality psychology, including emotion; in the psychology of language; in how psychologists conceive of psychological explanation; in the psychology of reasoning; in the psychology of vision; and in attitudes toward nativism. The chapter then ends with an overview of the field of computational psychology as a whole. Boden acknowledges that "we're still a very long way from a plausible understanding of the mind's architecture, never mind computer models of it," but she believes that the advent of models of artificial intelligence has been extraordinarily important for the development of psychology.
Chapter 8 discusses the very minor role of anthropology as the "missing," or "unacknowledged," discipline of cognitive science. Here Boden touches on the work of the relatively few anthropologists who do fit into cognitive science.
Chapter 9, the last in volume 1, describes Noam Chomsky's early impact on cognitive science, discussing his famous review of B. F. Skinner's book Verbal Behavior, his characterization of a hierarchy of formal grammars, his development of transformational generative grammar and his defense of nativism and universal grammar. Boden notes that psychologists, including Miller, lost interest in transformational grammar after realizing that the relevant transformations were ways of characterizing linguistic structure and not psychological operations.
As Boden mentions, many people, including me, raised objections in the 1960s to Chomsky's so-called nativism—his view that certain principles of language are innate to a language faculty. She seems unaware that Chomsky's reasons for this view became clearer as time went on and formed the basis for the current, standard principles-and-parameters view, which explains otherwise obscure patterns of differences between languages.
Perhaps the heart of Boden's story is her account of the development of artificial intelligence, broadly construed. There were two sorts of artificial intelligence at the beginning: One treated beliefs and goals using explicit languagelike "propositional" representations, whereas the other—the connectionist approach—took beliefs and goals to be implicitly represented in the distribution of excitation or connection strengths in a neural network.
The proposition-based approach, outlined in chapter 10, initially developed programs for proving theorems and playing board games. These were followed by studies of planning, puzzle problem solving, and expert systems designed to provide medical or other advice. Special programming languages were devised, including LISP, PROLOG, PLANNER and CONNIVER. Systems were developed for default reasoning: For instance, given that something is a bird, assume it flies (in the absence of some reason to think it does not fly); given that it is a penguin, assume it does not fly (in the absence of some reason to think it does fly).
There were difficulties. One was "computational complexity"—almost all methods that worked in small "toy" domains did not work for more realistic cases, because of exponential explosions: Operating in even slightly more complex domains took much longer and used many more resources. Another issue was whether "frame" assumptions (such as that chess pieces remain in the same position until captured or moved) should be built into the architecture of the problem or should be stated explicitly. This became a pressing issue in thinking about general commonsense reasoning: Is it even possible to explicitly formulate all relevant frame assumptions?
On the other side was the connectionist neural-net approach, considered in chapter 12, which seeks to model such psychological capacities as perception, memory, creativity, language and learning, using interconnected networks of simple units. Connectionism was especially concerned with rapidly recognizing and classifying items given their observed characteristics, without having to go through a long, complicated chain of reasoning.
In the simplest case of a single artificial perceptron, several real-number inputs represent the values of selected aspects of the observed scene, and an output value (the activation of the perceptron in question), possibly 1 or 0, indicates yes or no. The perceptron takes a weighted sum of the input values and outputs 1, or yes, if the sum is greater than some threshold value; if not, the output is 0. Perceptrons can be arranged in feed-forward networks, so that the output of the first layer goes to perceptrons in the second layer, whose outputs are inputs to a third layer, and so on until a decision is made by a final threshold unit. Given appropriate weights and enough units, a three-layer network can approximate almost any desired way of classifying inputs. Relevant weights do not need to be determined ahead of time by the programmer. Instead, the network can be "trained" to give desired outputs, by making small corrections when the network's response is incorrect.
There are other kinds of connectionist networks. For example, in certain sorts of recurrent networks, the activations of the units settle into a more or less steady state.
Boden describes these developments in loving detail, along with bitter disputes between proponents of proposition-based research and those who favored the connectionist approach. The disagreements were fueled by abrupt changes in U.S. government funding, which are noted in chapter 11. Much of the government money available was provided in the expectation that artificial intelligence would prove to be militarily useful. In the 1980s, funders decided to switch their support from proposition-based artificial intelligence to connectionism. They did so both because of perceived stagnation in the proposition-based approach (mainly due to the difficulties mentioned above), and because connectionism became more attractive with the discovery (or rediscovery) of back-propagation algorithms for training multilayer networks.
More recent developments are described in chapter 13. These include virtual-reality systems, attempts to construct societies of artificial agents that interact socially, and CYC—a project aimed at explicitly representing enough of the commonsense background to enable an artificial system to learn more by reading dictionaries, textbooks, encyclopedias and newspapers. Chapter 14 is a rich account of computational and cognitive neuroscience. Topics touched on include challenges to the computational approach, theories of consciousness and philosophy of mind. In chapter 15, Boden describes the origins of artificial life and then discusses reaction-diffusion equations, self-replicating automata, evolutionary networks, computational neuro-ethology (computational interpretation of the neural mechanisms that underlie the behavior of an animal in its habitat) and work on complex systems. Chapter 16 reviews philosophical thinking about mind as machine. Is there a mind-body problem? If a robot simulation of a person were developed, would it be conscious? Would it suffer from a mind-body problem? Would it be alive? A very brief final chapter lists promising areas for further research.
This is, as far as I know, the first full-scale history of cognitive science. I am sure that knowledgeable readers may have various quibbles about one or another aspect of this history (like my own objection above to the discussion of Chomsky's work in linguistics). But I doubt that many, or in fact any, readers will have the detailed firsthand knowledge that Boden has of so much of cognitive science. Future histories of the subject will have to build on this one.