Cognitive Science

Abstract

Cognitive science, which MeSH classifies under psychology, is the interdisciplinary study of the mind and intelligence, drawing together psychology, linguistics, philosophy, computer science, neuroscience, and anthropology. It turns on a single question: how does a physical system represent and process information so as to think? Its defining premise, forged in the cognitive revolution of the 1950s, is that mental states are internal representations and that cognition is the rule-governed manipulation of those representations — a stance that displaced behaviorism's refusal to look inside the black box. The field is unified less by method than by a shared explanatory target: perception, memory, reasoning, and language treated as computation over structured mental content. It has generated rival research programs — symbol systems, connectionist networks, embodied accounts, and probabilistic models — each a different wager about the format of thought.

Keywords: cognitive science, computational theory of mind, mental representation, cognitive revolution, Bayesian cognition

The wager that thinking is computation is what let a linguist, a philosopher, a computer scientist, and a neurophysiologist find they were studying the same object. When a person hears an ambiguous sentence and settles instantly on one reading, cognitive science does not ask merely which brain regions light up or which behavior follows; it asks what problem the mind is solving, what information it is using, and what procedure could compute the answer (Marr, 1982). This article follows that question from the revolution that made it askable to the modern models that try to answer it.

The Cognitive Revolution

Cognitive science was born from a rejection. Through the first half of the twentieth century, behaviorism held that a science of mind could speak only of stimulus and response, treating whatever lay between as an unobservable and therefore unscientific black box. The cognitive revolution of the mid-1950s broke that prohibition by showing that the intervening mental processes were not only real but measurable and formally describable. George Miller's demonstration that immediate memory holds only about seven items, and that this limit reflects a fixed channel capacity for transmitting information, treated the mind as an information-processing device with quantifiable bounds (Miller, 1956). In the same years Noam Chomsky's review of B. F. Skinner's account of language argued that the unbounded productivity of human grammar could never be explained by conditioned responses, and demanded instead an internal system of rules (Chomsky, 1959).

Two further strands converged. Herbert Simon's model of bounded rationality recast human choice as computation under resource limits rather than the frictionless optimization economics assumed, giving decision-making a procedural, mechanistic character (Simon, 1955). And the new digital computer supplied both a metaphor and an existence proof: if a machine could manipulate symbols to solve problems, then symbol manipulation was a candidate account of thought itself. By 1960 these threads — memory as channel, language as rule system, choice as bounded computation, and the computer as model — had woven into a common program, later chronicled as the founding of a new science of mind (Gardner, 1985; Miller, 2003).

Channel capacity: the magical number seven

An observer judging stimuli along one dimension can identify only so many categories before errors set in. Raise the number of categories and watch the information actually transmitted rise, then flatten against a fixed ceiling of about 2.81 bits.

Below the ceiling, each added category is identified reliably, so transmitted information tracks the input.

Schematic idealization of the channel-capacity limit; empirical capacities vary by dimension.

Types of Cognitive Science

In the Medical Subject Headings vocabulary, cognitive science is a subtype of psychology and is divided into two narrower descriptors, summarized in Table 2. The division is by level of description rather than by subject matter: both study the same cognitive phenomena, but one characterizes them as neural mechanism and the other as information-processing function. The categories are therefore complementary rather than mutually exclusive — a single act of remembering is a legitimate object of both — and much of the field's most productive work lives precisely where they meet. MeSH is a classification built for indexing the biomedical literature, so its two-way split is a coarse administrative cut through a field that in practice also encompasses linguistics, philosophy, and artificial intelligence; it should be read as a filing convenience, not as an exhaustive theory of the discipline's parts.

The Computational Theory of Mind

At the field's classical core is the computational theory of mind: the claim that mental states are relations to internal representations and that thinking is the manipulation of those representations according to formal rules. Allen Newell and Herbert Simon gave the thesis its sharpest form as the physical symbol system hypothesis — the proposition that a physical system capable of storing and transforming symbol structures has the necessary and sufficient means for general intelligent action (Newell, 1980). On this view a mind and a suitably programmed computer are, at the relevant level of abstraction, the same kind of thing: both are physical devices whose behavior is explained by the symbols they carry and the operations they perform on them.

Jerry Fodor pressed the representational claim in two influential directions. He argued that thought is conducted in an internal, language-like system of representation — a language of thought whose symbols combine syntactically to yield the systematicity and productivity of human cognition. And he proposed that the mind's input systems are modular: fast, mandatory, domain-specific processors, informationally sealed off from central belief (Fodor, 1983). Modularity offered a concrete architecture for how a symbol-processing mind might be organized, and made testable predictions about which capacities should be encapsulated. Together the symbol-system and language-of-thought hypotheses set the agenda the rest of the field has argued with ever since: they specify a format for mental content — discrete, structured, amodal symbols — that later programs would each contest.

The strongest philosophical objection to this program targets its sufficiency claim head-on. John Searle's Chinese room argument imagines a person who manipulates Chinese symbols purely by following an English rulebook, producing fluent Chinese replies without understanding a word: the setup runs the right program yet grasps no meaning, which Searle takes to show that syntactic symbol manipulation is not sufficient for semantics, and so that a program alone cannot constitute a mind (Searle, 1980). Defenders answer that understanding belongs to the whole system rather than to its parts, and the exchange remains the reference point for arguments over whether computation can yield genuine intentionality.

Marr's Levels of Analysis

Cognitive science's most durable methodological contribution is David Marr's insistence that an information-processing system must be understood at three distinct and complementary levels, and that confusing them is a recipe for bad explanation (Marr, 1982). The computational level specifies what problem the system solves and why — the goal of the computation and the logic by which it can be achieved. The algorithmic level specifies how: the representations the system uses and the procedure that transforms input into output. The implementational level specifies the physical substrate in which that procedure is realized, whether neurons or silicon.

The levels are semi-independent, and that is the point. The same computational problem — say, recovering three-dimensional structure from a two-dimensional image — can be solved by different algorithms, and the same algorithm can run on different hardware. An explanation pitched at the wrong level explains little: describing the firing of individual neurons reveals nothing about what visual problem those neurons collectively solve, just as knowing the goal of a computation leaves its mechanism unspecified. Marr's framework thus licenses cognitive science to study the mind at the level of function and algorithm without waiting for neuroscience to finish, while insisting the levels must ultimately cohere. It remains the standard vocabulary for saying what would count as a complete explanation of a cognitive capacity, and structures debates the field is still having (Núñez et al., 2019).

Marr's three levels of analysis

One information-processing system, three complementary explanations. Choose a task, then step through Marr's levels to see how each answers a different question about the same process.

Connectionism and the Embodied Challenge

The classical symbol-processing view did not go unchallenged. Connectionism proposed a different format for thought entirely: instead of discrete symbols manipulated by rules, it modeled cognition as patterns of activation spreading across networks of simple, neuron-like units, with knowledge stored not in explicit symbols but in the weights of the connections between them (Rumelhart & McClelland, 1986). Such networks learn from examples rather than being programmed, degrade gracefully under damage, and generalize by similarity — properties that fit much of human cognition awkwardly captured by rigid rules. The dispute between symbolic and connectionist accounts became, and remains, one of the field's organizing tensions, and modern deep learning is the direct descendant of the connectionist program. Jerry Fodor and Zenon Pylyshyn sharpened that dispute into a lasting challenge: because thought is systematic — anyone who can understand 'John loves Mary' can understand 'Mary loves John' — its representations must have constituent structure, so a connectionist network either fails to explain that systematicity or explains it only by implementing a classical symbol system at a higher level (Fodor & Pylyshyn, 1988).

A second challenge questioned a shared assumption of both camps: that cognition is disembodied computation over abstract symbols, indifferent to the body and world in which it occurs. Grounded cognition holds instead that mental representations are rooted in the brain's modal systems for perception and action, so that understanding a concept partly reenacts the sensory and motor experience of it (Barsalou, 2008). The extended mind thesis pushed further still, arguing that the machinery of cognition need not stop at the skull: when a notebook or a device reliably plays the functional role a memory would, it is literally part of the cognitive system (Clark & Chalmers, 1998). See embodied cognition for the fuller development of these arguments, which relocate part of the mind into the body and environment.

Bayesian concept learning

The numbers below are examples of a hidden concept between 1 and 100. Two hypotheses compete: powers of two (6 members) and even numbers (50 members). Add examples and watch which hypothesis a rational learner favors. The smaller hypothesis wins when the data fit both — the size principle.

Selected examples: 2, 4, 8. Red buttons (6, 10) are even but not powers of two — selecting one rules out the powers-of-two hypothesis.

All examples are powers of two. With 3 examples, the size principle favors the smaller hypothesis at 99.8% — a suspicious coincidence under “even numbers.”

Worked Example

Modern cognitive science often models a mental capacity as Bayesian inference: the mind is treated as combining prior beliefs with the likelihood of the evidence to compute a posterior degree of belief. Consider a stripped-down version of concept learning (Tenenbaum et al., 2011). A person is told that the numbers 2, 4, and 8 are examples of some concept drawn from the range 1 to 100, and must decide between two hypotheses: H1, the powers of two (the set {2, 4, 8, 16, 32, 64}, six members), and H2, the even numbers (the fifty even numbers from 2 to 100). Both hypotheses are logically consistent with the three examples, so a purely logical learner could not choose. A Bayesian learner can.

The key is the size principle: if examples are sampled at random from the true concept, then under a hypothesis of size m each example has probability 1/m, and the likelihood of n independent examples is (1/m)ⁿ. A smaller hypothesis assigns higher probability to any data it fits, because it spreads its probability over fewer possibilities. With three examples, the likelihood under the six-member H1 is (1/6)³ = 1/216 ≈ 0.00463, while under the fifty-member H2 it is (1/50)³ = 1/125,000 = 0.000008. The likelihood ratio is 125,000 / 216 ≈ 578.7 in favor of the powers of two.

Assume the learner starts with no preference, priors P(H1) = P(H2) = 0.5, so the prior odds are 1. Multiplying prior odds by the likelihood ratio gives posterior odds of about 578.7 to 1, and converting to a probability, P(H1 | data) = 578.7 / (578.7 + 1) ≈ 0.998. From three examples that a logician would find equally compatible with both rules, the Bayesian learner concludes with about 99.8% confidence that the concept is the powers of two. The suspicious coincidence that all three examples happen to be powers of two — vanishingly unlikely if the concept were merely the even numbers — is exactly what the size principle formalizes, and it is a leading account of how people generalize from a handful of cases (Griffiths et al., 2010). The interactive demonstration above adds examples and traces the resulting shift in the posterior.

Discussion

Cognitive science is unusual among sciences in being defined by a hypothesis rather than a subject matter or a method. Its unifying bet — that cognition is computation over representations — is what allowed a linguist, a philosopher, and an engineer to see themselves as colleagues, and it remains productive precisely because it is contestable: connectionism, grounded cognition, and the extended mind each accept that cognition is some kind of information processing while disputing the format and location of the representations involved. The field's history is less a march toward consensus than a sustained argument over that format, and the argument has repeatedly generated new empirical programs rather than dissolving into philosophy.

That very breadth has drawn a critique of fragmentation. A bibliometric analysis argued that the interdisciplinary core imagined at the founding never fully cohered, and that cognitive science has largely devolved into a loose federation whose contributing disciplines cite each other less than the founding vision implied (Núñez et al., 2019). Whether one reads this as failure or as healthy specialization, it sharpens the question of what still holds the field together. The most compelling answer is methodological: Marr's levels continue to give the disciplines a shared language for what would count as an explanation, and the computational stance continues to let a finding in one field constrain theory in another. The mind remains a single explanatory target approached from several directions, which is what the founders meant by a science of cognition in the first place.

Current Directions

The most active contemporary program models cognition as probabilistic inference over structured representations, uniting the classical concern with structure and the statistical learning of the connectionist tradition. On this view many capacities — categorization, causal reasoning, language understanding, intuitive physics — are well described as rational Bayesian inference given the learner's knowledge and the data available, an approach that has become the field's dominant computational framework (Griffiths et al., 2010; Tenenbaum et al., 2011). A refinement, computational rationality, folds in the cost of computation itself, treating minds and machines as making the best use of bounded reasoning resources rather than achieving unbounded optimality — an idea that traces directly back to Simon's bounded rationality (Gershman et al., 2015).

The parallel rise of deep learning has reopened the oldest questions in a new key. A prominent argument holds that machines will learn and think like people only when they incorporate the structured, causal, compositional knowledge that human cognition brings to a problem, rather than relying on pattern statistics alone (Lake et al., 2017). Reinforcement learning has become a shared vocabulary for biological and artificial agents, with a distinction between slow, incremental learning and fast, model-based inference mapping onto both neural systems and machine architectures (Botvinick et al., 2019). At the same time the implementational level has grown richer: network neuroscience characterizes cognition as a property of the brain's connectivity rather than of isolated regions (Bassett & Sporns, 2017), and cognitive computational neuroscience proposes to close Marr's loop by fitting task-performing computational models directly to neural and behavioral data (Kriegeskorte & Douglas, 2018). The founding ambition — one account of mind spanning function, algorithm, and mechanism — is, if anything, more concretely in reach than before.

Common Misconceptions

Cognitive science is just another name for cognitive psychology.

Cognitive psychology is one contributing discipline; cognitive science is the broader interdisciplinary field that also draws on linguistics, philosophy, computer science, neuroscience, and anthropology, unified by the computational study of mind (Gardner, 1985).

The computer is only a loose metaphor for the mind.

For the classical program it is a literal hypothesis: the physical symbol system claim holds that symbol manipulation is necessary and sufficient for intelligence, making mind and computer the same kind of system at the level of computation (Newell, 1980).

Explaining the brain's biology would explain cognition.

Marr's levels show that implementational detail alone does not reveal what problem a system solves or by what algorithm; a full explanation must also specify the computational goal and the procedure (Marr, 1982).

Glossary

Bounded rationality.

Simon's account of decision-making as optimization under limits of information and computation, rather than the frictionless maximization assumed by classical economics.

Channel capacity.

The maximum amount of information, measured in bits, that a communication channel can reliably transmit; Miller argued immediate memory has a fixed capacity of about seven items.

Cognitive revolution.

The mid-twentieth-century shift that replaced behaviorism's stimulus-response framework with an account of the mind as an information-processing system.

Computational theory of mind.

The thesis that mental states are relations to internal representations and that thinking is the rule-governed manipulation of those representations.

Connectionism.

An approach that models cognition as patterns of activation across networks of simple units, storing knowledge in connection weights learned from examples rather than in explicit symbols.

Embodied cognition.

The view that cognitive processes are shaped by and grounded in the body's sensory and motor systems, rather than being disembodied symbol manipulation.

Extended mind.

The thesis that external tools playing the functional role of a mental process are literally part of the cognitive system that includes them.

Grounded cognition.

The claim that mental representations reuse the brain's modal systems for perception and action, so that concepts are partly reenactments of experience.

Language of thought.

Fodor's hypothesis that thinking occurs in an internal, syntactically structured symbol system whose compositionality explains the productivity of thought.

Levels of analysis.

Marr's distinction between the computational, algorithmic, and implementational descriptions of an information-processing system, each answering a different question.

Modularity.

The proposal that the mind's input systems are fast, mandatory, domain-specific processors that are informationally encapsulated from central cognition.

Physical symbol system.

Newell and Simon's term for a system that stores and transforms symbol structures, hypothesized to have the necessary and sufficient means for general intelligence.

Probabilistic model of cognition.

A model that treats a mental capacity as Bayesian inference, combining prior knowledge with the likelihood of evidence to compute a posterior belief.

Representation.

An internal state that stands for something in the world and over which cognitive processes operate; the shared posit whose format the field's programs dispute.

Size principle.

The Bayesian rule that, when examples are sampled from a concept, smaller hypotheses assign higher likelihood to the data they fit, driving generalization from few examples.

Key Researchers

Lawrence W. Barsalou (living). Cognitive psychologist at the University of Glasgow; his grounded-cognition and perceptual-symbol-systems work is the leading challenge to the classical view that cognition runs on amodal, disembodied symbols. ORCID · Wikipedia

Noam Chomsky (living). Linguist at the University of Arizona and MIT (emeritus); his 1959 review of Skinner's Verbal Behavior is often taken as the opening argument of the cognitive revolution against behaviorism. Google Scholar · Wikipedia

Jerry Fodor (1935-2017). Philosopher at Rutgers University; he framed the modularity of mind and the language-of-thought hypothesis, defining the computational-representational theory of mind that cognitive science set out to test. Wikipedia

Thomas L. Griffiths (living). Psychologist and computer scientist at Princeton University; he develops probabilistic models of cognition that link human inductive inference to rational analysis and to the limits shared by human and machine intelligence. ORCID · Wikipedia

David Marr (1945-1980). Neuroscientist at MIT; his three levels of analysis remain cognitive science's standard framing of what it means to explain an information-processing system. Wikipedia

George A. Miller (1920-2012). Psychologist at Princeton University and co-founder of the Harvard Center for Cognitive Studies; his channel-capacity paper and later work made him a defining figure of the cognitive revolution. Wikipedia

Ulric Neisser (1928-2012). Psychologist at Cornell University; his 1967 textbook Cognitive Psychology named the field and set its agenda, earning him the label father of cognitive psychology. Wikipedia

Allen Newell (1927-1992). Computer scientist at Carnegie Mellon University; co-author of the physical symbol systems hypothesis and of the Logic Theorist, the General Problem Solver, and the Soar architecture. Wikipedia

Steven Pinker (living). Experimental psychologist and linguist at Harvard University; his work on language, computation, and the mind carried cognitive-science arguments to a broad readership. ORCID · Wikipedia

Herbert A. Simon (1916-2001). Polymath at Carnegie Mellon University; he introduced bounded rationality and, with Newell, the physical symbol systems hypothesis, winning the Turing Award and the Nobel Prize in Economics. Wikipedia

Joshua B. Tenenbaum (living). Computational cognitive scientist at MIT; he leads the Bayesian program in modern cognitive science, modeling learning and reasoning as probabilistic inference over structured representations. ORCID · Wikipedia

Paul Thagard (living). Philosopher of cognitive science at the University of Waterloo; his textbook Mind: Introduction to Cognitive Science framed the field's competing representational approaches for a generation of students. ORCID · Wikipedia

Frequently Asked Questions

What is cognitive science? It is the interdisciplinary study of the mind and intelligence, uniting psychology, linguistics, philosophy, computer science, neuroscience, and anthropology around the idea that cognition is the processing of information over mental representations (Gardner, 1985).

How is cognitive science different from cognitive psychology? Cognitive psychology is one of the contributing disciplines, focused on the experimental study of mental processes; cognitive science is the broader field that integrates it with linguistics, philosophy, artificial intelligence, and neuroscience under a shared computational framework (Miller, 2003).

What was the cognitive revolution? It was the mid-1950s shift that overturned behaviorism by showing that internal mental processes are measurable and formally describable, exemplified by Miller's work on memory capacity and Chomsky's critique of behaviorist accounts of language (Miller, 1956; Chomsky, 1959).

What is the computational theory of mind? It is the claim that mental states are internal representations and that thinking is the manipulation of those representations by formal rules, given its strongest form in Newell and Simon's physical symbol system hypothesis (Newell, 1980).

What are Marr's three levels of analysis? They are the computational level (what problem is solved and why), the algorithmic level (what representations and procedure are used), and the implementational level (how the procedure is physically realized), which together specify a complete explanation of a cognitive system (Marr, 1982).

What is connectionism? It is an approach that models cognition as activation spreading across networks of simple units, storing knowledge in connection weights learned from examples rather than in explicit symbolic rules, and it is the ancestor of modern deep learning (Rumelhart & McClelland, 1986).

What is embodied cognition? It is the view that cognition is grounded in the body's perceptual and motor systems, and in some versions extends into the environment, challenging the classical assumption that thought is disembodied symbol manipulation (Barsalou, 2008; Clark & Chalmers, 1998).

What role does Bayesian modeling play in cognitive science? It has become the dominant computational framework, treating mental capacities such as learning and reasoning as probabilistic inference that combines prior knowledge with evidence, and linking human cognition to rational analysis and machine intelligence (Tenenbaum et al., 2011).

References