COGNITIVE SCIENCE

Cognitive science is the interdisciplinary study of mind, in which the concepts and methods of artificial intelligence (AI) are central (Boden forthcoming). The most prominent disciplines within the field are AI, artificial life (A-life), psychology, linguistics, computational neuroscience, and philosophy—especially the philosophy of mind and language. Cognitive anthropology is included too, though often goes unseen under the label of evolutionary psychology.

The many relevant subfields include robotics, whether classical, situated, or evolutionary; studies of enactive vision, where the organism's own movements (of eyes and/or body) provide crucial information for acting in the world; the psychology of human-computer interaction, including various aspects of virtual reality such as avatars; and computational theories of literature, art, music, and scientific discovery. Nonhuman minds are studied by computational ethology and neuroethology, and by A-life.

Who Is a Cognitive Scientist?

Not everyone working in the key disciplines is a cognitive scientist. Only those taking a computational approach to questions about mind are considered cognitive scientists.

Some AI workers, for example, are not cognitive scientists because they have no theoretical interest in human thought. Their aim is to challenge their ingenuity as computer engineers by getting a program or robot to do a task that people either cannot do or do not want to do. If hunches, or experimental evidence, about human psychology can help them achieve that goal, that is fine. But if nonhuman tricks are available, such as looking ahead in a chess game to consider all the legal possibilities for several moves, they will use them. These computer scientists are engaged in technological AI, not psychological AI. Only the latter is a proper part of cognitive science.

Even someone who does have a professional interest in human minds need not be a cognitive scientist. For instance, many social psychologists study patterns of interpersonal behavior without asking about the information processes that underlie them and make them possible. Even some cognitive psychologists insist that they are not cognitive scientists, because they follow James Gibson's (1979) affordance theory of perception—which allows for information pickup but not for information processing. (Their self-description is based on an overly narrow view of what cognitive science covers: Gibsonian insights have become prominent in various areas of cognitive science, such as enactive vision.)

Similarly, many linguists—sociolinguists and historical philologists, for instance—are not primarily concerned with just how language is generated and/or understood. But even those who do focus on these computational matters do not all agree. Chomskian linguistics, for example, was crucial in the rise of cognitive science and has deeply influenced the philosophy of mind; but non-Chomskian accounts of syntax have been developed since. In addition, theories of pragmatics have become at least as prominent as theories of syntax—and pragmatics is an aspect of situatedness, a concept of growing importance within cognitive science as a whole. As for anthropology, most anthropologists see their field as a hermeneutic enterprise, not a scientific one. They reject psychological explanations of culture in general, and computational accounts in particular.

Cognitive Science Is about More than Cognition

It includes cognitive psychology, of course: the study of language, memory, perception, problem solving, and creative thinking. What is more, most research has focused on individual human adult cognition. However, other aspects of mind are studied too: motivation, emotion, choice (Dennett 1984), development, psychopathology, interpersonal phenomena, motor control, and animal psychology.

Consider emotion, for example. The role of emotion in problem solving, attitude formation, and neurosis were topics of research in AI and computational psychology in the early 1960s. But the problems were too difficult, and were largely dropped. Interest revived later, partly because of neuroscientific work on emotional intelligence and partly because of advances in the computational theory of scheduling in multigoal systems (Sloman 1993). Interdisciplinary conferences on the psychology, neuroscience, computer modeling, and philosophy of emotion blossomed at the turn of the century, when the topic became a prominent aspect of research.

Whether the focus of attention is on development or psychopathology, emotion or motor control, the prime interest for cognitive science is in the abstractly defined computational functions that generate the behavior concerned. But the neural mechanisms that implement them are often studied too. Despite the functionalist doctrine of multiple realizability, many cognitive scientists want to know how psychological functions are actually implemented in the brain. When functionalism began in the 1960s, little attention was paid to the nervous system by philosophers or AI scientists. Since the 1980s, that has been less true.

Indeed, neuroscience as such has become increasingly concerned with computational questions. On the one hand, there are theories (and computer models) of specific neural circuits doing closely specified things. For instance, cells in the retina and/or visual cortex that compute particular visual features, such as light gradients or surface textures; or cells in the female cricket's brain that enable her to discriminate the song of male crickets of the same species, and to move accordingly. On the other hand, there are broad-brush theories about the computational functions carried out by large areas of the brain, where the focus is less on specific individual cells than on general neuroanatomy: the different cell types, locations, and connections of the neurons.

Developmental Issues

Most cognitive scientists study already established phenomena, although many include learning in their subject matter. Some, however, study—and model—mental development. And some do this because they believe that adult psychology cannot be properly understood without knowing how it developed. In short, they see the mind as an epigenetic system, deeply informed by its developmental history.

Epigenesis was stressed long ago by Conrad Waddington in biology and Jean Piaget in psychology. It is self-organized development, grounded in innate predispositions in continual dialectic interaction with the (internal and external) environment. For example, there are inborn dispositions to attend to broadly facelike stimuli, or to human speech-sounds. Once the attention is caught, learning can help develop the infant's pattern recognition and discriminatory powers. In some cases, such as face recognition, the neural mechanisms relevant at different stages have been largely identified.

An epigenetic view is not strictly environmentalist, nor strictly nativist either. Rather, it stresses the dialectical interplay between the two. Late twentieth-century work in developmental neuroscience and developmental psychology has therefore led to a radical reconceptualization of nativism (Elman et al. 1996). Some philosophers of biology have defined new accounts of self-organization and dynamical development accordingly (Oyama 1985).

What It Means to Say that Cognitive Science is Computational

Cognitive science employs computational models of mind in two senses.

First, the substantive concepts in its theories are computational. The mind is seen as some sort of computational system (just what sort is hotly disputed), and mental structure and mental processes are described accordingly (Haugeland 1997). So whereas many psychologists (and other scientists) use computers to express/clarify their theories, and especially to manipulate their experimental data, only cognitive scientists import computational ideas into their theories.

Second, computer modeling is often used to clarify and test computational theories of mind. Often, but not always, some work in cognitive science (in AI and psychology, not just in philosophy) employs computational concepts and insights, but with insufficient detail to allow programs to be written. When programming is possible, it provides several advantages. Even program failures can be scientifically illuminating, pointing out lacunae or mistakes in the theory, or fundamental limitations of the methodology being used. However, successes may be even more instructive. For if a program—or a robot—produces a given performance, one knows that it suffices to do so.

Whether real minds (or brains) use similar processes to produce equivalent performance is another matter: just because a program does something in a certain way, it does not follow that people do too. This question can be answered only by empirical evidence. Sometimes, a programmed theory models not only psychological phenomena at various levels, but also the details of their underlying neural base. In such cases, validation requires both psychological and neuroscientific evidence.

The references to computational ideas in the previous two paragraphs cover concepts rooted in two different intellectual traditions, namely, cybernetics and Turing computation. These were closely linked in the years when cognitive science began.

A seminal paper by Warren McCulloch and Walter Pitts (1943) prompted early work both in neural nets and in what is sometimes called GOFAI, or "Good Old-Fashioned AI." (It also influenced the design of the von Neumann computer.) McCulloch and Pitts integrated three key ideas of the early twentieth century: propositional logic, neuron theory, and Turing computation. They proved that anything expressible in the propositional calculus is computable, and can be mapped onto some specifiable neural net. In addition, they suggested that a fourth key idea—feedback, the core concept of cybernetics—could be defined in terms of these networks, in which case purpose and learning could be embodied in them too.

A few years later they published another paper, in which they argued that probabilistic networks are more like brains and can do things naturally that logic-based systems cannot (Pitts and McCulloch 1947). They still insisted, nevertheless, that their original, logical, account was correct in principle.

In short, the concept of computational systems is normally used within the field to cover both GOFAI and connectionism. (Some philosophers, however, restrict it to the former.) Cognitive science includes both.

Sometimes, the reason why a computational theory is not actually modeled is that suitable computer technology does not yet exist. By the same token, many advances in cognitive science have depended partly upon advances in computing technology. These include both increases in size (computing power) and new types of virtual machine, embodying forms of computation that were not possible previously.

In some cases, the core ideas had already been defined long before the technology was available to test/explore them. Parallel distributed processing, for instance, was envisaged over twenty years before computers became powerful enough for it to be implemented in interesting ways. Similarly form-generating interactive diffusion equations and cellular automata were both first defined in the 1950s, but not extensively studied until the advent of large machines and computer graphics in the late 1980s. And genetic algorithms, glimpsed in the 1950s and defined in the late 1960s, were first implemented in the 1980s. Once the technology was available, further questions arose that had not been posed before.

Some Philosophical Problems

Many philosophical disputes arise within cognitive science. One dispute concerns the relative merits of the two AI approaches mentioned above: classical (symbolic) AI and connectionism, or neural networks. The latter is broadly inspired by the basic structure of the brain. (Some recent work in artificial neural networks tries to take more account of the subtleties of real neurons; even so, these models are hugely oversimplified in comparison with the real thing.) There are several types of neural networks, but the one most widely used within cog-nitive science—and the one of greatest interest to philosophers—is parallel distributed processing, or PDP.

Some researchers champion only one of these AI approaches, whereas others admit both because of their complementary strengths and weaknesses. Symbolic AI, or GOFAI, is better for modeling behaviors that involve hierarchical structure, advance planning, deliberation, and/or strict sequential order. The conscious, deliberative aspects of the mind are best suited to this approach. Connectionism, by contrast, is better for modeling the tacit learning and knowledge involved in pattern recognition, including the fuzzy family resemblances between instances of one and the same concept.

It does not follow that all unconscious mental processes are best modeled by PDP systems. Some psychoneural theories of action errors, including various clinical syndromes, employ hybrid (mixed) models in which the hierarchical aspects represent both conscious and unconscious processing.

internal representation

Another debate concerns the nature and importance of various kinds of internal representation. Connectionist representations are different from GOFAI ones, and several philosophers have argued that they are closer to the neural representations that embody concepts (Churchland 1989; Clark 1989, 1993; Cussins 1990). Computational neuroscience has described further types of representation. One example is emulator systems, which are neural mechanisms whose physical dynamics mimic the temporal changes being represented. Another, based on the anatomy of the cerebellum, is a way of representing motor behavior that is based neither on logic (GOFAI) nor on statistics (PDP), but on noneuclidean tensor geometry.

Some philosophers follow the AI community and/or the neuroscientists, in accepting that representations may take many different guises, depending on the role they have evolved to play. Others, however, argue that only formal-symbolic structures, expressed in a language of thought, are properly termed representations, and that only these can generate human conceptual/linguistic thought (Fodor 2000).

nature of computation

The nature of computation is a third topic of controversy (Scheutz 2002). Most philosophers define it as Alan Turing did in the 1930s—and his is still the only really clear definition. However, practicing AI scientists think of computation in a number of different ways, based on virtual machines whose properties are different from those of a Turing machine. Moreover, some people are trying to go beyond Turing computation by defining new forms of computers (hypercomputers), some—but not all—of which involve quantum computing. Some of these may turn out to be relevant to human brains, but others will not.

meaning in the real world

A fourth area of philosophical discussion focuses on whether—and if so, how—meaning (intentionality) can be grounded in the real world—and whether it can properly be attributed to programs and/or robots. Evolutionary theories of intentionality rule out GOFAI programs (as do many philosophers), but—arguably—allow meaning to be ascribed to some evolved robots. The grounding problem, on this view, is solved by the way in which the relevant mechanism has evolved in situated, embodied systems.

Empirical work that is closely related to the problem of intentionality includes research on theory of mind. Very young children are unable to grasp that each person is an agent with their own set of beliefs and interests, which may differ from those of the child. So although the child realizes that adults (and even other children) know many things that they do not, the child does not appreciate that someone else may believe something to be true that the child knows to be false. (This is why infants do not lie: they cannot conceive of doing so.) Normally, theory of mind develops spontaneously at around ages four or five, although in autistic children it apparently does not. In other words, inbuilt predispositions have evolved that lead the young child first to engage with other humans (maintaining eye contact, pointing to direct attention, turn-taking in communication, etc.), and eventually to attribute intentional states to them. Philosophers have asked (for instance) whether they do this by theorizing about other people's minds or by simulating, or empathizing with, them (Davies and Stone 1995).

consciousness in computational terms

A fifth philosophical puzzle concerns whether consciousness could be explained in computational terms—or in any other scientific, naturalistic, manner (Heil 2004, Newell 1980, Searle 1993). Research in various disciplines within cognitive science has shown that there is no such thing as the problem of consciousness; rather, there are many problems of consciousness, because the term is used to make many different distinctions. Some of these are much better understood than they were twenty years ago, thanks to computational work in AI, psychology, and neuroscience. Reflective self-consciousness, for example, and the bizarre dissociations of consciousness typical of multiple personality, are intelligible in terms of recursive processing, guiding procedures, and access limitations within complex hierarchical structures for perception, memory, and action.

Considerable controversy, however, still attends the problem of qualia. Some cognitive scientists argue that qualia can be analyzed in terms of complex dispositions for making discriminatory computations (Dennett 1991). Others see them as aspects of an irreducible informational feature of the universe, applying not only to human brains but to atoms as well (Chalmers 1996). Still others make further suggestions, including several based on quantum physics. In short, there are many theories of qualia, and no agreement about what a successful theory might look like.

opposition to orthodox cognitive science

A sixth controversy—or rather, batch of controversies—arises from recent work that opposes orthodox (neo-Cartesian) cognitive science (Cliff, Harvey, and Husbands 1993; Port and van Gelder 1995; Wheeler 2005). This involves both empirical theory/modeling and philosophical discussion. In general, it draws on the traditions of phenomenological philosophy and/or autopoietic biology, rather than Cartesianism. It rejects both symbolic and connectionist AI, and the concept of representation. It highlights embodied systems (not abstract simulations), embedded in their environment and responding directly to it. Examples include situated robotics in AI, dynamical systems theory, ecological psychology, and A-life studies of evolution and coevolution.

Philosophies inspired by these empirical researches include the theory of extended mind (Clark 1997). This starts from the position that minds must necessarily be embodied and that memory storage lies largely outside the skull (ideas familiar within phenomenology and GOFAI, respectively) and goes on to argue that an individual person's mind is extended over the surrounding cultural artifacts: language, customs, and material objects—from palaces to pencils. The claim is that mind is not merely deeply influenced by these things, but it is largely constituted by them.

Philosophical questions associated with A-life include whether evolution is a necessary characteristic of life, and whether the concept of autopoiesis captures the essence of life (Bedau 1996; Maturana and Varela 1987). If living things are defined as autopoietic systems—whose physical unity, boundaries, and self-maintenance are attained by self-organized metabolic processes—then questions about the origins of life take on a different color, as do questions about the possibility of strong A-life (life in computer memory)—so called by analogy to strong AI.

Philosophers of A-life consider not only the nature of life as such, but how and why it is related to mind. Must all minds be evolved, for example? Autopoietic theorists define all life as involving cognition, while insisting that only linguistic life (i.e., adult humans) involves representations. But questions remain about whether, and if so why, life really is essential for mind. By the same token, questions remain about whether the study of A-life is essentially unrelated to cognitive science or fundamental to it.

culture and cognitive science

Finally, culture-directed research in cognitive science raises philosophical questions too. One concerns the nature of group mind, or as it is more commonly called, distributed cognition (Hutchins 1995). Can one identify aspects of cognition that cannot be attributed to any single individual, but only to a team of enculturated persons acting in concert—and if so, can one model such phenomena in computers? Two more such questions concern the evolution of information-processing mechanisms that underlie important cultural phenomena—religion or aesthetic appreciation, for example—and the evolution of culture as such.

See also Computationalism; Neuroscience; Psychology.

Bibliography

Boden, Margaret A. The Creative Mind: Myths and Mechanisms. 2nd ed. London: Routledge, 2004.

Boden, Margaret A. Mind as Machine: A History of Cognitive Science. Oxford: Oxford University Press, forthcoming.

Bedau, Mark A. "The Nature of Life." In The Philosophy of Artificial Life, edited by Margaret A. Boden. Oxford: Oxford University Press, 1996.

Chalmers, David J. The Conscious Mind: In Search of a Fundamental Theory. Oxford: Oxford University Press, 1996.

Churchland, Paul M. A Neurocomputational Perspective: The Nature of Mind and the Structure of Science. Cambridge, MA: MIT Press, 1989.

Clark, Andy J. Associative Engines: Connectionism, Concepts, and Representational Change. Cambridge, MA: MIT Press, 1993.

Clark, Andy J. Being There: Putting Brain, Body, and World Together Again. Cambridge, MA: MIT Press, 1997.

Clark, Andy J. Microcognition: Philosophy, Cognitive Science, and Parallel Distributed Processing. Cambridge, MA: MIT Press, 1989.

Cliff, David; Harvey, Inman; and Husbands, Philip. "Explorations in Evolutionary Robotics." Adaptive Behavior 2 (1993), 73–110.

Cussins, Adrian. "The Connectionist Construction of Concepts." In The Philosophy of Artificial Intelligence, edited by Margaret A. Boden. Oxford: Oxford University Press, 1990.

Davies, Martin, and Tony Stone, eds. Folk Psychology: The Theory of Mind Debate. Oxford: Blackwell, 1995.

Dennett, Daniel C. Consciousness Explained. Boston: Little Brown, 1991.

Dennett, Daniel C. Elbow Room: The Varieties of Free Will Worth Wanting. Cambridge, MA: MIT Press, 1984.

Dennett, Daniel C. The Intentional Stance. Cambridge, MA: MIT Press, 1987.

Elman, Jeffrey L., Elizabeth A. Bates, Mark H. Johnson, et al. Rethinking Innateness: A Connectionist Perspective on Development. Cambridge, MA: MIT Press, 1996.

Fodor, Jerry A. The Mind Doesn't Work That Way: The Scope and Limits of Computational Psychology. Cambridge, MA: MIT Press, 2000.

Gibson, James J. The Ecological Approach to Visual Perception. Boston: Houghton Mifflin, 1979.

Haugeland, John, ed. Mind Design II: Philosophy, Psychology, Artificical Intelligence. 2nd ed. Cambridge, MA.: MIT Press, 1997.

Heil, John, ed. Philosophy of Mind: A Guide and Anthology. Oxford: Oxford University Press, 2004. See especially 3–6, and 9–10.

Hutchins, Edwin L. Cognition in the Wild. Cambridge, MA: MIT Press, 1995.

McCulloch, Warren S., and Walter H. Pitts. "A Logical Calculus of the Ideas Immanent in Nervous Activity." In The Philosophy of Artificial Intelligence, edited by Margaret A. Boden. Oxford: Oxford University Press, 1990. Essay first published in 1943.

Maturana, Humberto R., and Francisco J. Varela. The Tree of Knowledge: The Biological Roots of Human Understanding. Boston: New Science Library, 1987.

Newell, Allen. "Physical Symbol Systems." Cognitive Science 4 (1980): 135–183.

Oyama, Susan. The Ontogeny of Information: Developmental Systems and Evolution. Cambridge, U.K.: Cambridge University Press, 1985.

Pitts, Walter H., and Warren S. McCulloch. "How We Know Universals: The Perception of Auditory and Visual Forms." In Embodiments of Mind, edited by Seymour Papert. Cambridge, MA: MIT Press, 1965. Essay first published in 1947.

Port, Robert F., and Timothy J. van Gelder, eds. Mind as Motion: Explorations in the Dynamics of Cognition. Cambridge, MA: MIT Press, 1995.

Scheutz, Matthias, ed. Computationalism: New Directions. Cambridge, MA: MIT Press, 2002.

Searle, John R. The Rediscovery of the Mind. Cambridge, MA: MIT Press, 1993.

Sloman, Aaron. "The Mind as a Control System." In Philosophy and the Cognitive Sciences, edited by Christopher Hookway and Donald Peterson. Cambridge, U.K.: Cambridge University Press, 1993.

Wheeler, Michael W. Reconstructing the Cognitive World: The Next Step. Cambridge, MA: MIT Press, 2005.

Margaret A. Boden (1996, 2005)