Loading field map
How do minds build abstract categories from the flood of sensory experience? A child learns that a golden retriever and a chihuahua are both "dog," that a chair remains a chair whether made of wood or metal, and that fairness is a property of actions, not objects. Explaining this capacity—concept formation—has been one of cognitive science's most persistent and contested problems. Over the past century, competing frameworks have offered radically different answers, each emerging from the limitations of its predecessors and each leaving a lasting imprint on how we think about learning.
The first major clash in the study of concept learning pitted Gestalt psychology against behaviorism. Gestalt psychologists, active from the 1910s, argued that learning is not the accumulation of associations but a sudden restructuring of perceptual wholes—insight. A chimpanzee that suddenly sees a stick as a tool to reach a banana has formed a new concept through reorganization, not trial and error. Behaviorism, rising in the 1910s and dominating through the 1950s, directly opposed this view. For behaviorists like B.F. Skinner, concepts are nothing more than stimulus-response associations built through reinforcement. A child learns "dog" by being rewarded for discriminating dogs from non-dogs. Gestalt psychology emphasized innate organizing principles; behaviorism insisted that all knowledge comes from the environment. Neither framework, however, could explain the structured, generative nature of human concepts—how we can apply a concept to novel instances or combine concepts into complex thoughts.
In the 1930s–1970s, two constructivist frameworks moved beyond the behaviorist–Gestalt debate by focusing on development. Lev Vygotsky's sociocultural theory argued that concepts first appear in social interaction and are then internalized. A child learns the concept of "tool" by participating in culturally mediated activities with more knowledgeable others. Jean Piaget's genetic epistemology, by contrast, emphasized the child's active construction of concepts through assimilation and accommodation, driven by universal stages of cognitive development. These two frameworks coexisted but disagreed sharply: Vygotsky saw social context as constitutive of concept formation, while Piaget treated it as a scaffold for individual discovery. Both, however, shared a constructivist core—concepts are not given by the world or innate but built by the learner—and both would later influence embodied and enactive approaches.
The cognitive revolution of the 1950s and 1960s brought a new answer: Classical Computationalism. Inspired by digital computers, this framework treated concepts as symbolic representations—mental words that combine according to syntactic rules. Learning a concept, in this view, is a form of hypothesis testing: the learner induces a rule that defines the category (e.g., "a bachelor is an unmarried adult male"). This approach superseded behaviorism by positing internal mental representations and absorbed Piaget's stage-like development into a formal, computational model of cognitive growth. But it faced a problem: many concepts resist definitional boundaries. A "bird" includes penguins and ostriches, which don't fly—a rule-based definition struggles to capture such graded membership.
Connectionism, emerging in the 1980s, directly contested Classical Computationalism's core assumption. Instead of symbolic rules, connectionist networks represent concepts as patterns of activation distributed across many simple processing units. Learning occurs by adjusting the weights between units, gradually shaping the network to produce correct outputs. Concepts in connectionist models are graded prototypes, not definitions—a robin is a better "bird" than a penguin because it activates more typical features. This framework revived aspects of associationism (learning through co-occurrence) but with a complex, nonlinear internal structure. The symbolic–connectionist debate became a defining conflict of the 1980s and 1990s: are concepts best understood as rules or as patterns? Connectionism showed that many concept-learning phenomena, such as generalization from examples, could be modeled without explicit rules.
By the 1990s, two further frameworks emerged from perceived limits of both connectionism and classical computationalism. Bayesian Cognitive Science treats concept learning as probabilistic inference: the learner starts with a prior distribution over hypotheses and updates it based on observed data. A Bayesian model of learning "dog" would assign higher probability to hypotheses that generalize from a few examples in a rationally structured way. This framework coexists with connectionism but offers a different computational account—one that emphasizes optimality and the role of prior knowledge. Dynamical Systems Theory, meanwhile, rejected representationalism altogether. Drawing on nonlinear dynamics, it views learning as changes in the coupling between an agent and its environment, without invoking internal models or representations. A child learning to reach for a cup is not forming a concept of "cup" as a mental symbol but entering into a stable dynamical pattern of perception and action. This directly challenged the representational assumptions shared by both symbolic and connectionist frameworks.
Embodied and Enactive Approaches, also gaining traction from the 1990s, revived themes from Vygotsky and from ecological psychology. They argue that concepts are not abstract symbols but are grounded in sensorimotor experience. The concept of "chair" is tied to the affordances of sitting—what the body can do with it. Learning a concept is inseparable from acting in a physical and social world. This framework absorbs insights from Dynamical Systems Theory and directly opposes the disembodied view of Classical Computationalism. Predictive Processing, the most recent major framework (2000–present), attempts to unify perception, action, and learning under a single principle: the brain minimizes prediction error by generating hierarchical predictions about sensory input. Concept learning, in this view, is the process of refining the generative model that produces those predictions. Predictive Processing extends Bayesian ideas (inference as prediction error minimization) and connects to connectionist architectures (hierarchical neural networks). It is currently one of the most active research programs, offering a potential synthesis of Bayesian, connectionist, and even embodied insights.
Today, no single framework dominates. Bayesian Cognitive Science, Predictive Processing, and Connectionism are leading research programs, each with active communities. Dynamical Systems Theory and Embodied/Enactive Approaches remain influential, especially in developmental psychology and robotics. These frameworks agree on several points: learning involves inference and adaptation; concepts are not simply copied from the environment; and the learner brings prior structure to the task. But they disagree fundamentally on the nature of that structure. Is it symbolic, distributed, probabilistic, or non-representational? Is the body and social context merely a source of input, or constitutive of the concepts themselves? These disagreements are not signs of failure but of a healthy, pluralistic field. Each framework excels at explaining different aspects of concept formation—Bayesian models capture rational generalization, connectionist models capture graded typicality, dynamical models capture real-time sensorimotor learning, and embodied approaches capture the grounding of abstract concepts in action. The history of learning and concept formation is thus not a linear march toward a single truth but a continuing conversation, with each new framework challenging and enriching the others.