Learning And Concept Formation

Learning and concept formation is the study of how minds acquire new knowledge—how a child comes to know that a dog is not a cat, how an animal learns that a tone predicts food, how anyone sorts the endlessly varied things in the world into usable categories.

11 frameworks, from 1690 to today · Cognitive Science · Social Sciences

still a working view ended · each leads to its entry below

Overview

Learning and concept formation is the study of how minds acquire new knowledge—how a child comes to know that a dog is not a cat, how an animal learns that a tone predicts food, how anyone sorts the endlessly varied things in the world into usable categories. The stakes are large. If we can say how concepts are formed, we can say what a concept is, and that question sits at the center of cognitive science: it bears on perception, language, reasoning, and on whether a machine can be said to understand anything at all. The history of this subfield is a history of competing answers to one deceptively simple question: when you learn that some things go together, what exactly is it that you have learned?

Associationism: Learning as the Linking of Ideas

The oldest answer is associationism—the view that learning consists in forming links between ideas, so that one comes to mind when another is present. Its modern form was set out by the British empiricists. John Locke argued in the Essay Concerning Human Understanding (1690) that the mind begins as a blank slate and that complex ideas are built by combining simple ones; David Hume, in the Treatise of Human Nature (1739–40), gave the linking principles their classic statement—resemblance, contiguity in time or place, and cause and effect. The tradition's lasting contribution was the idea that learning is a matter of connections rather than of innate content, and that these connections can be described by general laws.

Associationism's weakness was that it was largely a philosophy of mind with little experimental grip. That changed when it met physiology and measurement.

Conditioning: Association Made Measurable

In the early twentieth century, Ivan Pavlov's work on dogs made association an experimental object. A dog naturally salivates to food; Pavlov showed that if a neutral stimulus—a tone or a light—reliably precedes food, the dog comes to salivate to the tone alone. This is classical conditioning: learning that one event predicts another1. Pavlov's laboratory, and the tradition that followed, treated learning as the formation of stimulus–stimulus associations, studied by counting responses.

Edward Thorndike had already approached learning from the other side2. In his puzzle-box experiments (from the late 1890s), cats placed in a box learned to escape by pressing a latch, and Thorndike observed that responses followed by a satisfying outcome became more frequent. He called this the law of effect and argued, controversially, that animals did not reason their way out but simply strengthened the connection that worked. B. F. Skinner developed this into operant conditioning: behavior is shaped by its consequences—reinforcement increases a response, punishment or extinction decreases it23. Skinner's The Behavior of Organisms (1938) and his schedules of reinforcement made the study of learning a rigorous, quantitative science of behavior.

The behaviorist program, dominant in mid-century American psychology, held that learning could be studied without appeal to internal representations at all. Its central claim was that what is learned is a connection between stimulus and response, and that complex behavior is built from these. The approach was powerful for prediction and control, and it changed education, animal training, and clinical practice. But it left a puzzle: organisms often behave as though they know more than their reinforcement history directly teaches.

The Cognitive Turn: Hypotheses, Maps, and Insight

That puzzle drove the cognitive turn. Edward Tolman's studies of rats in mazes (reported through the 1930s and 1940s) found that animals allowed to explore without reward later ran the maze efficiently once reward appeared—evidence of what Tolman called a cognitive map, a latent internal representation of the layout. The rats had learned something that no immediate stimulus–response link explained.

A parallel challenge came from the study of human concept learning. In the 1950s, Jerome Bruner, Jacqueline Goodnow, and George Austin (A Study of Thinking, 1956) presented people with cards varying on attributes—color, shape, number—and asked them to discover which combination defined a category. People did not passively accumulate associations; they formed and tested hypotheses about the rule, discarding those that failed. This hypothesis-testing view made concept learning a matter of active inference, and it fit the emerging cognitive-science picture of the mind as an information processor.

The most influential statement of the new view was Frank Rosenblatt's perceptron (1958), a simple artificial network that learned to classify patterns by adjusting connection weights. It was a concrete demonstration that a machine could learn a category from examples—and it made the connection between learning and representation explicit.

Prototypes and Exemplars: What a Category Is

The next turning point concerned the form of a concept. The classical view, inherited from philosophy, held that a concept is a definition: a set of necessary and sufficient features. But in the 1970s, Eleanor Rosch's experiments on color categories and on everyday categories like bird and furniture showed that categories have graded structure. People judge some members (a robin) as better examples than others (a penguin), and they are faster to verify typical members. Rosch proposed that categories are organized around a prototype—an abstract summary of the most typical features—and that membership is a matter of similarity to that prototype rather than of meeting a definition4. Prototype theory explained the typicality effects that the classical view could not.

A rival account, exemplar theory, argued that people do not store an abstract prototype at all but retain specific remembered instances and judge new items by their similarity to those stored examples45. Developed by Douglas Medin and Marguerite Schaffer (1978) and later by Robert Nosofsky, exemplar models could account for many of the same typicality findings while also explaining why categories can be learned from few examples and why context shifts judgments. The prototype–exemplar debate became one of the field's most productive arguments, because both sides could fit much of the same data and the disagreement forced sharper experiments.

Theory-Based Concepts: Concepts in a Web of Belief

A further challenge came from the observation that similarity alone is too unconstrained to explain categorization. In the 1980s, Susan Carey, Frank Keil, and Gregory Murphy and Medin argued that concepts are embedded in intuitive theories: people's beliefs about how the world works shape which features matter and how categories are organized. A child does not group a whale with fish because of surface similarity but because of a theory about what kind of animal it is. This theory-based view held that concepts are not isolated similarity clusters but nodes in a structured body of knowledge, and that learning a concept means acquiring a place for it in that structure.

Connectionism: Learning as Distributed Representation

The 1980s brought a revival of the network idea in a new form. Connectionism—the use of artificial neural networks with many simple units—was transformed by the backpropagation algorithm, popularized by David Rumelhart, Geoffrey Hinton, and Ronald Williams (1986). Unlike the perceptron, multilayer networks could learn internal representations that were not specified in advance. Rumelhart and James McClelland's Parallel Distributed Processing (1986) argued that knowledge could be stored as patterns of activation across many units rather than as discrete symbols, and that learning was the gradual adjustment of connection weights. The approach offered a unified account of learning, representation, and generalization, and it made concept formation a matter of statistical learning over distributed representations. It also reopened the old associationist question in a new technical key: was the mind, after all, a system of connections?

Bayesian Cognitive Science: Learning as Inference

The most recent major framework treats learning as probabilistic inference4. Bayesian cognitive science holds that learners combine prior expectations with observed evidence according to Bayes' rule to arrive at beliefs about the world67. In concept learning, this means treating categories as hypotheses with probabilities, updated as examples arrive. The approach, developed from the 1990s onward by researchers including Joshua Tenenbaum and Thomas Griffiths, gave a principled account of how people generalize from sparse data—how a child learns a word's meaning from a handful of examples—and it connected concept formation to rational statistical inference. It also provided a framework for comparing competing models, since different theories of concepts make different predictions about how beliefs should update.

Multiple Systems: More Than One Way to Learn

Running alongside these debates was a growing recognition that learning is not one thing. Work on category learning in the 1990s and 2000s, associated with researchers such as F. Gregory Ashby and others, distinguished multiple systems: a verbal, hypothesis-testing system that learns explicit rules, and an implicit, procedural system that gradually acquires associations through reinforcement. The distinction helped explain why some categories are learned quickly and consciously while others are learned slowly and without awareness, and why brain damage or cognitive load can impair one system while sparing the other. This pluralism now shapes much of the field: the question is less "which theory of concepts is right?" than "which learning system is doing what, under which conditions?"

The Present

Today the subfield is a set of overlapping research programs rather than a single consensus. Prototype and exemplar models remain in active use and are often combined; theory-based accounts inform work on conceptual development and on how children's intuitive theories change; connectionist and, increasingly, deep-learning models provide the computational backbone for studying representation learning; Bayesian models supply a normative standard for how learning should proceed given the evidence; and the multiple-systems framework organizes work on the neural and cognitive architecture of category learning. The open questions are the old ones in sharper form: what exactly is represented when a concept is learned, how much of concept formation is statistical versus theory-driven, and whether the brain's learning systems are best described as one inference engine or several. The history of the field is, in effect, a long argument about what kind of thing a concept is—and that argument is not settled.

References

  1. Classical conditioning, Wikipedia
  2. Operant conditioning chamber (History), Wikipedia
  3. Operant conditioning, Wikipedia
  4. Concept learning (Prototype, Exemplar, Bayesian), Wikipedia
  5. Exemplar theory, Wikipedia
  6. Bayesian cognitive science, Wikipedia
  7. Bayesian inference, Wikipedia

The frameworks of Learning And Concept Formation

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1690-1900

    Associationism

    How do ideas come to call one another to mind?

    Fly to its tower in Nousfarer

  2. 1913-1970

    Behaviorism

    How do consequences shape what an animal does?

    Fly to its tower in Nousfarer

  3. 1930-1960

    Cognitive map theory

    Do animals learn layouts they never use for reward?

    Fly to its tower in Nousfarer

  4. 1956-1980

    Hypothesis-testing theory of concept learning

    Do people test rules when sorting things into categories?

    Fly to its tower in Nousfarer

  5. 1958-1969

    Perceptron and early neural network learning

    Can a machine learn to sort patterns from examples?

    Fly to its tower in Nousfarer

  6. 1971-Present

    Prototype theory

    Why are some members of a category better examples?

    Fly to its tower in Nousfarer

  7. 1978-Present

    Exemplar theory

    Do we store summaries or particular remembered cases?

    Fly to its tower in Nousfarer

  8. 1985-Present

    Theory-theory

    Do beliefs about the world decide what counts as similar?

    Fly to its tower in Nousfarer

  9. 1986-Present

    Connectionism

    Can knowledge live in patterns across many units?

    Fly to its tower in Nousfarer

  10. 1990-Present

    Bayesian cognitive science

    How do we generalize from so few examples?

    Fly to its tower in Nousfarer

  11. 1990-Present

    Multiple systems framework

    Do separate learning systems work in the same mind?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

AssociationismBehaviorismCognitive map theoryHypothesis-testing theory of concept …Perceptron and early neural network l…Prototype theoryExemplar theoryTheory-theoryConnectionismBayesian cognitive scienceMultiple systems framework

Beyond the field

Where Learning And Concept Formation's ideas came from in other fields, and what they led to there: the strongest 12 of 52 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Learning And Concept Formation is a planet in Nousfarer

Its city has a tower for each of these 11 frameworks. Land, walk in, and climb one concept at a time.

Land on Learning And Concept Formation

Related fields

Fields its ideas are linked to

More in Cognitive Science

See where Learning And Concept Formation sits on the Map of Human Knowledge