Loading field map
Learning and concept formation is the branch of cognitive science that asks how minds acquire, organize, and use general knowledge from experience. It sits at the intersection of psychology, neuroscience, linguistics, artificial intelligence, and philosophy, but its core concern is specific: how does a finite set of encounters with particular things give rise to the ability to recognize, name, and reason about kinds of things—dogs, justice, triangles, breakfast—that one has never encountered before?
The field treats two intertwined problems. The first is learning: the process by which experience modifies the cognitive system. The second is concept formation: the structure that results from that modification. A concept is typically understood as a mental representation that groups entities, events, or properties together, enabling categorization, inference, and communication. Learning is the mechanism; concepts are the product. Neither can be studied entirely without the other, because the structure of concepts constrains what can be learned, and the dynamics of learning shape what concepts look like.
The field is organized around a cluster of enduring questions. How do learners generalize from limited data? A child hears a handful of examples of "bird" and can then apply the word to robins, ostriches, and penguins—but not to bats or airplanes. What licenses that leap? How do concepts get their boundaries, and why are some boundaries sharp while others are fuzzy? Why do some categories (like "prime number") have clear criteria, while others (like "game" or "furniture") resist definition yet remain usable? How do concepts change with expertise, development, or brain damage? And how do concepts relate to language—do words merely label pre-existing concepts, or do linguistic categories shape the concepts themselves?
The stakes extend beyond description. If we understand how concepts form, we can design better educational curricula, build artificial systems that learn from examples rather than being hand-programmed, diagnose and treat disorders of categorization, and address long-standing philosophical puzzles about the nature of abstract thought. The field also carries practical weight for machine learning: the question of whether neural networks form "concepts" in the human sense, or merely statistical regularities, is a live debate that borrows directly from the cognitive science of concepts.
The modern field emerged from several converging traditions. In the late nineteenth and early twentieth centuries, philosophers and psychologists treated concepts as definitions—sets of necessary and sufficient conditions. To have the concept "bachelor" was to know that something is an unmarried adult male. This view, often associated with the classical theory of concepts, dominated early experimental psychology and philosophy of mind. It had a natural ally in logic: if concepts are definitions, then categorization is a deductive process, and learning a concept is a matter of discovering the rule that generates it.
The classical view began to erode in the mid-twentieth century. The philosopher Ludwig Wittgenstein argued, through his famous discussion of games, that many everyday concepts have no single defining feature but are held together by "family resemblances"—overlapping similarities, like the features shared among relatives in a family. Around the same time, experimental psychologists found that people reliably rate some category members as "better" examples than others: a robin is a more typical bird than a penguin, even though both satisfy any reasonable definition of "bird." These typicality effects were difficult to explain if concepts were all-or-nothing definitions.
The decisive break came in the 1970s with the work of Eleanor Rosch and her collaborators, who proposed that many concepts are organized around prototypes—central, representative examples or averaged representations—with membership graded by similarity to the prototype. A robin is a prototypical bird; a penguin is a peripheral one. This prototype theory explained typicality effects, accounted for fuzzy boundaries, and resonated with the emerging computational metaphor of the mind. It also had a powerful methodological consequence: it moved the study of concepts from armchair analysis to empirical investigation of how people actually categorize.
A rival to prototype theory emerged from a different direction. In the late 1970s and 1980s, researchers studying expertise and cross-cultural cognition, notably Susan Carey and Frank Keil, argued that many concepts—especially those in science, religion, and everyday folk theories—are not organized by similarity at all but by causal-explanatory structure. The concept "water" is not a set of typical features but a substance with hidden properties; the concept "dog" is not a bundle of appearances but an animal with an internal essence that explains why it looks and behaves as it does. This theory-theory of concepts holds that concepts are like miniature scientific theories: they posit unobservable causes, support counterfactual reasoning, and change when the theory changes. A child who learns that whales are mammals, not fish, has not just moved a prototype; she has revised her theory of what whales are.
A third major approach, exemplar theory, rejected both prototypes and theories. It proposed that concepts are stored as collections of specific remembered instances. You categorize a new animal as a dog not by comparing it to an abstract prototype or consulting a theory, but by retrieving your memory of your neighbor's beagle and your childhood collie and judging similarity to those particular exemplars. Exemplar models have a distinct advantage: they can capture the fact that people sometimes use atypical examples to make judgments, and they require no abstraction process at all. They also align naturally with instance-based learning in artificial intelligence.
These three approaches—prototype, exemplar, and theory-based—are not merely competing hypotheses but different answers to the question of what a concept is. Prototype theory says a concept is a summary representation, typically a set of weighted features or a central tendency. Exemplar theory says a concept is a set of stored instances. Theory-theory says a concept is a structured piece of knowledge that explains and predicts. Each approach has strengths and weaknesses.
Prototype theory handles typicality effects elegantly and explains why categorization is fast and automatic. But it struggles with compositionality: how do you combine the prototype for "pet" and the prototype for "fish" to get the concept "pet fish," which is not a simple average of the two? It also has difficulty with concepts that have no clear prototype, such as "odd number" or "justice," and with the fact that experts often categorize on the basis of non-obvious features.
Exemplar theory avoids the compositionality problem by storing instances rather than summaries, and it naturally explains why people can be sensitive to correlations among features that a prototype would average away. But it faces the opposite problem: it seems to require storing an implausibly large number of instances, and it does not explain how people form genuinely abstract concepts like "number" or "cause" that have no concrete exemplars. It also struggles to explain why people can sometimes categorize things they have never seen anything like before.
Theory-theory explains abstract and causal concepts, and it accounts for the fact that conceptual change often looks like theory revision rather than similarity adjustment. But it is less clear about how theories get learned in the first place, and it risks making concepts too intellectual—many concepts, like "red" or "pain," seem to be grounded in perception and similarity rather than explanation.
The field has largely moved beyond treating these as mutually exclusive. Contemporary research often combines them. A common synthesis holds that concepts are hybrid: they include perceptual prototypes for rapid categorization, stored exemplars for flexible judgment, and theoretical knowledge for reasoning and explanation. The relative weight of each component depends on the concept, the task, and the learner's expertise. A novice birdwatcher may rely on prototypes; an experienced ornithologist may rely on theory-laden knowledge about anatomy and ecology; both may retrieve specific memorable encounters.
A parallel development transformed the field's methods and theoretical vocabulary. Beginning in the 1950s and accelerating through the 1980s, researchers in artificial intelligence and cognitive modeling began to implement theories of concept learning as computer programs. Early symbolic models treated concept learning as rule induction: the program searched a space of logical or decision-tree rules to find one that classified examples correctly. These models were transparent and testable, and they captured the classical intuition that concepts are rules.
Connectionist or neural-network models, which became prominent in the 1980s, offered a different picture. A network learns to categorize by adjusting the strengths of connections between simple processing units, gradually shaping a distributed representation that captures statistical regularities in the input. These models have no explicit rules or prototypes; the concept is implicit in the pattern of connection weights. They have been remarkably successful at capturing graded typicality effects, learning from noisy data, and modeling developmental transitions. But they raise a deep question: if a network can categorize correctly without any explicit representation of the category, has it formed a concept? Many cognitive scientists argue that it has not, because the network cannot use the concept in flexible, context-sensitive ways—it cannot reason about counterfactuals, combine concepts in novel ways, or explain its own judgments.
This debate between symbolic and connectionist approaches is not merely technical. It reflects a fundamental disagreement about the nature of concepts. Symbolic approaches assume concepts are discrete, structured, and language-like; connectionist approaches assume they are graded, distributed, and statistical. The dispute remains unresolved, though many researchers now adopt a middle position: the brain may use both statistical learning and symbolic-like structure, with different systems handling different kinds of concepts.
A major strand of the field studies how concepts emerge and change over the lifespan. Developmental cognitive science has shown that infants as young as a few months can form categories from visual and auditory input, grouping together different dogs or different faces. But these early categories are often perceptual, based on surface similarity. Over the first years of life, children begin to form concepts that go beyond appearances—they learn that animals have insides, that objects have hidden causes, and that words refer to kinds rather than individuals.
A central finding is that language plays a dual role in concept formation. On one hand, words serve as invitations to form categories: when a child hears the same word applied to different objects, she infers that they belong to a common kind. On the other hand, language may reshape concepts themselves. The linguistic relativity hypothesis, associated with Benjamin Lee Whorf and revived in more qualified forms by contemporary researchers, holds that the categories available in a language influence how speakers perceive and remember the world. Studies of color, space, and number have shown modest but reliable effects of language on cognition, though the strong claim that language determines thought is widely rejected.
Conceptual change in childhood and in science follows similar patterns. Children initially believe that the earth is flat, that living things are defined by movement, and that weight and size determine falling speed. These naive theories are not mere collections of false beliefs but coherent explanatory frameworks. Learning science involves restructuring these frameworks, not just adding facts. This insight has direct implications for education: teaching that simply corrects individual misconceptions often fails, because the underlying theory remains intact. Effective instruction must engage and revise the theory itself.
Current research on learning and concept formation is characterized by methodological pluralism and increasing integration. Behavioral experiments remain the foundation, measuring reaction times, typicality ratings, and categorization accuracy. Neuroimaging studies have begun to identify the neural substrates of concept use, showing that different kinds of concepts—tools, animals, abstract ideas—engage partially distinct brain networks, and that the same concept can be represented in multiple formats depending on context. Computational modeling continues to test theories against behavioral and neural data, with modern deep-learning systems providing new tools for exploring how statistical regularities can give rise to category structure.
One active debate concerns the role of prediction in concept learning. Many contemporary theories, influenced by predictive-processing frameworks in neuroscience, hold that the brain is fundamentally a prediction machine: it learns by trying to anticipate sensory input and adjusting when predictions fail. On this view, concepts are not stored representations but generative models—internal simulations that can predict what a new instance will look like, how it will behave, and what it will be good for. This framework unifies perception, action, and cognition, but it remains controversial whether it can account for the full range of conceptual phenomena, especially abstract and logical concepts.
Another live question is whether concepts are best understood as individual mental states or as social and cultural products. The distributed-cognition tradition argues that many concepts are not inside individual heads at all but are carried by language, tools, and institutions. The concept of "money," for example, exists in a network of practices, symbols, and laws that no single individual fully possesses. This view challenges the methodological individualism of much cognitive science, but it has not displaced the individual-level focus of most experimental work.
The field's enduring contribution is a set of tools for thinking about thinking. It has shown that concepts are not simple definitions, not pure images, and not arbitrary labels, but complex, multi-component structures that serve multiple functions: they allow rapid categorization, support inductive inference, enable communication, and ground abstract reasoning. The challenge of explaining how such structures arise from experience remains open, and it is one of the central scientific questions about the nature of mind.