RESEARCH DOSSIER
Concept Formation and Categorization
Abstract. Concept formation and categorization are fundamental processes through which learners organize instances into classes that support discrimination, prediction, inference, and adaptive action. This research monograph critically synthesizes the experimental and theoretical history of category learning, including classical classification paradigms, prototypes and typicality, exemplar memory, logical rules and exceptions, similarity-based selective attention, information-integration tasks, error-driven clustering, causal knowledge, and structured Bayesian generalization. It analyzes developmental and language-mediated conceptual growth, neural and multiple-system theories, example scheduling and feedback, relational transfer, and comparative artificial category learning. Methodological attention is given to operational definitions, stimulus controls, trial-wise strategy change, psychometric and construct validity, parameter and model recovery, predictive tests, reproducibility, and the distinction between descriptive fit and mechanistic evidence. The synthesis finds convergent evidence for graded and flexible categorization while identifying persistent theoretical underdetermination between competing representations and learning systems. It proposes an experimentally discriminating research agenda grounded in primary studies and peer-reviewed research reviews.
Keywords: concept formation; categorization; category learning; concepts; prototypes; exemplars; rules; similarity; selective attention; information-integration learning; decision boundaries; typicality; causal concepts; Bayesian generalization; category-based induction; developmental categorization; transfer; neural learning systems; artificial cognition
Concept Formation as an Experimental and Theoretical Object
Concept formation is the process by which a learner acquires representations that support grouping, discrimination, prediction, and inference across encounters. Categorization is the assignment of particular instances to one or more groups under a task, context, or goal. A category is an extension of instances treated as equivalent for a specified purpose; a concept comprises the information and operations that permit such treatment and can support explanations beyond classification. These definitions prevent an observed labeling response from being mistaken for a complete account of conceptual knowledge. Experimental inquiry asks which features are extracted, which relational structures are learned, how instances and categories are represented, and why knowledge transfers to unfamiliar cases. Murphy & Medin (1985) Minda et al. (2024)
Several behaviors establish distinguishable evidential targets. Classification accuracy measures assigned labels under a response rule; category-based induction measures expectations about unseen attributes; typicality ratings probe graded internal organization; transfer tests evaluate responses to previously untrained stimuli; explanations and feature generation reveal accessible causal or semantic knowledge. An animal, a child, and an adult may achieve similar classification accuracy while using different cues. Likewise, a system may reliably group images while failing to infer the causal properties that ordinary language attaches to the group. The present monograph treats these differences as substantive empirical questions rather than methodological inconveniences. Gelman & Markman (1986) Sloutsky (2010)
The unit of explanation must be declared. A descriptive model captures category membership judgments; a process model predicts learning trials, attention changes, and errors; an informational model characterizes which features and dependencies are retained; a neural account specifies circuits that may implement these functions. The same behavioral function can be compatible with multiple representations, and a single representation can support several decision procedures. Models therefore earn explanatory strength through discriminating predictions under controlled training histories, changes in feedback, novel transfer probes, and independent measures of the representation or strategy being claimed. Nosofsky (1986) Johansen & Palmeri (2002)
The research boundary is experimental concept acquisition and its generalization. Natural-language semantics, broad associative learning, perceptual similarity geometry, and abstract reasoning enter where they explain categorical representation and use. The separate mathematical Science examines formal measurement and computational inference across psychology; here those methods are applied to category learning itself. In the Artificial Era, comparative work on computational categories adds a new testing environment: one can measure training distributions, stimulus representations, label decisions, and transfer separately across humans and artificial learners. Similar outputs establish a functional comparison and invite further tests of the mechanism responsible. Tenenbaum et al. (2006) Flesch et al. (2023)
The Historical Transformation of the Category Problem
Concept-learning experiments have roots in the study of rules, discrimination, association, and logical classification. Early laboratory problems presented sets of objects varying along a small number of binary features; an investigator could enumerate the lawful partitions and compare how readily participants learned them. The decisive methodological advance was to hold the elementary stimuli constant while varying how category membership depended on their features. This gave researchers a controlled way to distinguish difficulty attributable to the category structure from difficulty attributable to stimulus recognition or response familiarity. The enduring benchmark was established in the classification work of Shepard, Hovland, and Jenkins (1961). Shepard et al. (1961)
A subsequent line investigated whether people abstract central tendency information from distorted exemplars. Posner and Keele (1968) helped establish prototype-learning paradigms in which participants encountered structured distortions and later judged unfamiliar members or the central pattern. Research on naturally familiar categories moved the question toward graded membership and overlapping attributes. Rosch and Mervis (1975) tested family resemblance by comparing typicality with the distribution of shared and distinctive features, challenging the expectation that ordinary categories always have necessary and sufficient defining features. These experimental traditions remain distinct: a prototype benefit in a constructed stimulus space does not by itself determine the representation of an everyday biological or social concept. Posner & Keele (1968) Rosch & Mervis (1975)
Exemplar accounts subsequently formalized how classification can arise through comparison with multiple stored individual experiences. Medin and Schaffer (1978) articulated a context theory in which the similarity between a probe and remembered training cases determines categorization. Nosofsky (1986) connected identification and classification through a multidimensional similarity framework with learned attention weights. Later connectionist and cluster models combined instance sensitivity with error-driven adaptation and emergent group structure. The sequence from context theory to ALCOVE and SUSTAIN demonstrates that a historical label such as exemplar theory contains multiple computational proposals rather than a single fixed algorithm. Medin & Schaffer (1978) Nosofsky (1986) Kruschke (1992) Love et al. (2004)
Parallel traditions developed knowledge-based, developmental, logical, and neuropsychological accounts. Murphy and Medin (1985) emphasized explanatory coherence and background theories; Barsalou (1983) documented goal-dependent ad hoc categories; Feldman (2000) linked the perceived difficulty of Boolean concepts to formal complexity. Research on multiple learning systems proposed distinct explicit and procedural contributions, while Bayesian approaches examined generalization under structured priors. The history accordingly records an expansion of experimental targets, mathematical representations, and levels of explanation. A contemporary assessment must preserve the genuine achievements of each program while testing where its assumptions hold. Murphy & Medin (1985) Barsalou (1983) Feldman (2000) Ashby et al. (1998) Tenenbaum et al. (2006)
Categories, Concepts, Features, and Explanatory Coherence
A category can be defined extensionally by its members, intensionally by a feature description, or functionally by the inferences it permits. These descriptions often coincide in simple artificial experiments but can diverge in ordinary reasoning. A perceptually similar object may fall outside a biological kind, while a visually unusual object belongs because of its origin or internal causal organization. Concepts therefore have inferential dimensions that can remain hidden when evaluation uses only sorting or naming. A strong experiment measures whether learners treat an observed category as a predictive cluster and whether their generalizations change when causal information is introduced. Gelman & Markman (1986) Murphy & Medin (1985)
Feature structures range from deterministic conjunctions to probabilistic correlations and relational patterns. In a deterministic rule category, a particular property may suffice to identify membership. In a family-resemblance category, different members share overlapping but nonidentical properties, and diagnostic value depends on how often a cue also occurs in alternative categories. In a relational category, membership depends on a relation among elements, which may survive substantial changes in their physical characteristics. Typicality, feature frequency, diagnosticity, and causal centrality are distinct constructs. Their joint behavior is an empirical matter because a frequent feature need not be causally important, and an exceptional member can still be unambiguously classified. Rosch & Mervis (1975) Kurtz et al. (2013)
Prototype, rule, exemplar, and theory-based descriptions capture different kinds of information and offer different mechanisms for deploying it. A prototype can summarize the central feature profile of a learned category. An exemplar representation can preserve idiosyncratic cases and boundaries. A rule supplies an explicitly testable predicate or decision criterion. A causal theory supplies relations that make category members explanatory units rather than arbitrary bundles. None of these descriptions is sufficient simply because it generates the right average choices: the researcher must determine whether the representation also predicts memory, exceptions, generalization, explanations, and context-sensitive reclassification. Minda & Smith (2002) Zaki et al. (2003) Murphy & Medin (1985)
Knowledge structures are also shaped by goals. Barsalou's (1983) analysis of ad hoc categories shows how a temporary task can assemble objects that would rarely form a familiar taxonomic group. A person considering items to carry during an evacuation may construct a category using portability, practical value, and urgency. The objects may share little visual appearance yet display graded typicality within that goal. This property places category learning within action and decision-making: the relevance of a dimension and the membership criterion can change with consequences and intentions. An adequate theory must describe the conditions under which an established category is retrieved, modified, or constructed for a new purpose. Barsalou (1983) Love et al. (2004)
Laboratory Category Structures and Controlled Generalization
Artificial category experiments provide a precise bridge between psychological theories and observable choices. Stimuli can be described by binary properties, continuous dimensions, or complex images; category labels can be assigned according to a known rule, a noisy probabilistic structure, or a latent cluster. Investigators manipulate feature relevance, redundancy, covariance, category overlap, exposure frequencies, and reward feedback. Training trials show acquisition, whereas novel transfer items reveal which structural information supports the learned classification. The design is informative only when stimulus discriminability, response mapping, prior familiarity, and possible rule shortcuts are documented. Shepard et al. (1961) Nosofsky (1986)
The Shepard–Hovland–Jenkins six-type paradigm became important because different mappings over the same binary dimensions yielded different learning demands. A replication and extension by Nosofsky and colleagues (1994) recovered the classic relative difficulty pattern and compared process accounts. Such replication supports a reproducible property of those structures under comparable procedures, while the psychological explanation remains model dependent. A learner can initially search for one diagnostic dimension, then abandon it when exceptions accumulate. A transfer test that contrasts items near a learned exemplar with items satisfying a candidate rule can adjudicate between these possibilities more effectively than a training curve alone. Shepard et al. (1961) Nosofsky et al. (1994, replication)
The 5/4 category structure supplies a different probe: small classes share correlational patterns but contain exceptions that favor memory for particular training items. Medin and Schaffer (1978) used such structures to distinguish exemplar context predictions from additive feature-based alternatives. Subsequent model comparisons raised the question of whether apparent exemplar superiority was sensitive to response assumptions and allocation of attention. The Minda and Smith (2002) analysis and the Zaki and colleagues (2003) reassessment show why historical benchmark datasets remain scientifically productive when the precise flexibility and assumptions of competing models are made explicit. Medin & Schaffer (1978) Minda & Smith (2002) Zaki et al. (2003)
Laboratory paradigms also impose limits. Simplified dimensions are easier to measure and model than the multimodal objects encountered outside experiments. Researchers may know the correct category rule while participants infer the rule, or may know which features matter while participants develop their own. Low-dimensional classification can favor strategies seldom used for causal reasoning about biological species or institutional concepts. For this reason, experimental control must be followed by transfer across changed representations, delayed tests, different task formats, and naturalistic instances. The validity claim should be specific: successful classification of engineered shapes establishes a learning phenomenon within the tested stimulus and feedback environment. Ashby & Maddox (2005) Minda et al. (2024)
Typicality, Prototype Abstraction, and Category Central Tendency
A prototype model characterizes a category through a central or representative tendency in feature space. For continuous dimensions, the prototype may be a mean or centroid; for binary features, it may be a vector of modal values. A simple classification rule assigns higher membership strength to cases more similar to the relevant category prototype. This produces graded typicality, predicts judgments about unseen central items, and compresses training experience into a small representation. The central methodological question is whether behavior requires an abstracted summary or could be produced by storing many exemplars and averaging their influences at retrieval. Posner & Keele (1968) Rosch & Mervis (1975) Minda & Smith (2002)
Posner and Keele (1968) established an influential experimental context for studying abstraction from distortions. Rosch and Mervis (1975) connected graded typicality in familiar categories with patterns of shared attributes and between-category distinctiveness. These findings establish the importance of statistical regularity and centrality for human judgments, but their interpretation varies by task: a judged best example, a remembered typical item, and an inferred representative may emerge through different operations. Prototype learning can therefore be evaluated through delayed recognition, never-presented prototypes, transfer to new distortions, and explicit judgments of category centrality rather than classification alone. Posner & Keele (1968) Rosch & Mervis (1975)
Prototype representations are especially attractive when categories contain many related members and when compression of idiosyncratic variation aids generalization. However, a single centroid can conceal multimodal structure and fail at categories with a curved or disjoint distribution. An item can be near a mathematical average yet be unlikely under the actual category distribution. Cluster or mixture representations offer one way to retain substructure. SUSTAIN introduces clusters when prediction failures demand additional distinctions, providing a process-level account of movement between simple summaries and more differentiated representations. The resulting model has testable consequences for how unusual examples alter subsequent generalization. Love et al. (2004)
Evidence for prototype use depends on assumptions about attention, noise, decision rules, and the set of critical transfer items. Minda and Smith (2002) argued that prototype representations can outperform certain exemplar specifications in a classic paradigm; Zaki and colleagues (2003) showed that matching decision-rule flexibility materially changes such comparisons. The disagreement is informative because representation and decision processes are not observed separately by a response accuracy score. A rigorous study preregisters plausible model variants, compares held-out predictions and individual participants, and measures attention independently when the hypothesis concerns attended features. Prototype abstraction is thus an established behavioral phenomenon with multiple possible computational realizations. Minda & Smith (2002) Zaki et al. (2003) Johansen & Palmeri (2002)
Exemplar Memory, Similarity, and Instance-Sensitive Decisions
Exemplar theories explain categorization through the influence of stored individual cases. A new item activates traces of encountered instances in proportion to psychological similarity; evidence for competing categories is accumulated from the activated cases. This mechanism naturally retains exceptional members, permits a category to have several local centers, and predicts that frequency and recency can influence judgments. Medin and Schaffer's (1978) context theory made this approach experimentally testable by contrasting instance-based predictions with category-level and independent-cue alternatives. Crucially, exemplar storage is a functional modeling assumption that must be tested against recognition memory and transfer, rather than inferred directly from successful classification. Medin & Schaffer (1978)
Nosofsky's generalized context model provides a quantitative way to connect perceptual identification and categorization. Stimuli occupy a psychological space; distances are transformed into similarities, and diagnostic dimensions receive adjustable attention weights. The predicted probability of a category response depends on the similarity-weighted support supplied by the stored category members. This framework successfully captured important aspects of identification-categorization relationships in experiments with confusable perceptual stimuli. It also demonstrates a persistent methodological problem: a model can fit performance well using latent psychological distances and weights that are not independently determined. Identification tasks and scaling data help constrain those quantities. Nosofsky (1986)
Exemplar approaches can simulate graded typicality without explicitly storing a single average. A central new item may resemble many learned cases and therefore receive strong category evidence; an atypical familiar item may be classified accurately because one remembered trace contributes disproportionately. Yet unrestricted instance memory can become an overly flexible explanation when researchers allow stimulus-specific sensitivity and attention to vary freely. A meaningful test must compare how the same fitted parameters perform on unfamiliar shapes, boundary cases, exceptions, category frequency shifts, and delayed probes. The role of response noise and alternative decision rules must be reported, especially in small 5/4 learning datasets. Minda & Smith (2002) Zaki et al. (2003)
The evidence supports the psychological relevance of particular experiences, but does not settle whether human learners store complete stimulus records, transformed traces, or partial discriminative features. Memory may be selectively encoded, reconstructed, or weighted by goals. Hybrid accounts allow instances to influence early classification while abstract summaries or local clusters develop with experience. Johansen and Palmeri (2002) found changing patterns of generalization during learning, consistent with shifts in information use, while explicitly recognizing the difficulty of inferring representational architecture from transfer behavior alone. The most informative question is what trace-level information must be preserved to explain novel predictions across contexts. Johansen & Palmeri (2002) Love et al. (2004)
Rule Learning, Logical Structure, and Exceptions
Rule-based categorization describes decisions using predicates over features or relations. A learner may discover that all objects with one discriminative attribute belong to a group, that two attributes must jointly occur, or that membership depends on an exclusive relation. Rules differ in logical complexity, attentional requirements, and the number of exceptions they leave unexplained. A rule can be articulated verbally, but successful verbal report is an imperfect diagnostic of the mechanism producing choices. Experiments should measure hypothesis testing, rule articulation, response consistency, and generalization to items that separate plausible predicates. Shepard et al. (1961) Feldman (2000)
The six logical category types studied by Shepard and colleagues laid the groundwork for controlled investigations of learnability. Feldman (2000) later examined a broad range of Boolean concepts and related learning difficulty to the complexity of their logical expressions. Complexity supplies a principled alternative to the crude claim that all rules are easy while similarity-based categories are hard. The learner must discover the relevant variables, represent their combination, and maintain the rule through noisy or deceptive evidence. Logical length can predict some aspects of learning demand while working-memory limitations, instructions, and presentation order affect actual performance. Shepard et al. (1961) Feldman (2000)
RULEX, the rule-plus-exception model developed by Nosofsky, Palmeri, and McKinley (1994), specifies a stochastic search for simple rules supplemented by memories of exceptional cases. It predicts individual variation because learners may discover different rules and retain different exceptions while producing similar aggregate learning curves. This is a warning against treating the averaged participant as a coherent psychological subject. A population mean can combine fast rule discoverers, incremental similarity learners, and individuals who change strategies. Repeated critical transfer trials, participants' rule reports, and hierarchical analyses provide stronger tests of such heterogeneity than pooled accuracy alone. Nosofsky et al. (1994, RULEX)
Erickson and Kruschke (1998) demonstrated the value of combining rule and exemplar mechanisms in the ATRIUM model. Their experiments and modeling addressed evidence that category judgments can be controlled simultaneously by rule induction and case-specific learning. A hybrid explanation makes a stronger commitment when it predicts which component receives control as a function of category structure, instruction, error history, and feedback. Merely adding many modules can increase fit without clarifying mechanism. Model recovery experiments should therefore assess whether the available transfer conditions distinguish the intended mixture from one sufficiently flexible similarity process. Erickson & Kruschke (1998) Minda et al. (2024)
Selective Attention and the Geometry of Psychological Similarity
The similarity between two stimuli is task sensitive because dimensions can receive different attentional weights. When color reliably separates categories, the learner may progressively emphasize color and disregard shape; when the category rule changes, the weighting may shift. Formal psychological distance often combines dimension-wise discrepancies using a metric and then applies a decreasing similarity function. These choices influence predicted generalization gradients and can make the same stimulus collection appear very different under competing tasks. Category learning therefore modifies the use of a perceptual space and may affect discrimination within it. Nosofsky (1986) Goldstone (1994)
ALCOVE integrates exemplar-centered similarity with error-driven learning of attention and association weights. It can adaptively privilege diagnostic dimensions and explain complex learning phenomena, including certain reversals in accuracy across training. The model matters historically because it converts the broad idea of selective attention into a trial-by-trial mechanism that can be fitted and tested. The psychological claim is sharper than saying attention matters: feedback errors redistribute processing toward dimensions that reduce future category mistakes. The resulting predictions can be tested against eye movements, feature-relevance manipulations, and transfer to novel combinations of learned features. Kruschke (1992)
Goldstone (1994) investigated how categorization experience influences later perceptual discrimination. In experiments with category training followed by same-different judgments, acquired distinctiveness was observed when learning led participants to discriminate more sensitively along category-relevant contrasts, while certain conditions produced acquired equivalence. These findings link representation and perception, but the interpretation of a discrimination change depends on testing the contributions of response criteria, attention, and sensory coding. An improved same-different score does not isolate neural plasticity; convergent psychophysical measurements and controls for decisional bias are required to specify where change occurred. Goldstone (1994)
The dimensional attention framework also has limits for integral or relational stimuli. Some features are difficult to vary independently, and the geometrical effect of attention can be nonlinear. A new category task may require attention to interactions rather than isolated dimensions, making simple additive metrics insufficient. Investigators should assess whether learned distances generalize from measured identification confusions and whether subjects use stable dimensions across blocks. Similarity spaces can be powerful descriptive tools while remaining incomplete as explanations of semantic coherence or causal category membership. These limits motivate complementary rule, relational, and structured inference models. Minda et al. (2024) Tenenbaum et al. (2006)
Decision Boundaries and Information-Integration Learning
Decision-bound models describe categorization as assigning an item according to its location relative to a boundary in perceptual space. A unidimensional boundary supports a simple verbalizable rule, whereas a diagonal or curved boundary may require integration of several stimulus properties. This geometric formulation makes it possible to compare hypothesized strategies on the same stimulus distribution. It also places importance on perceptual noise, decision noise, and boundary stability, since the same observed error can reflect either unstable representation or a deliberately conservative decision criterion. Different observers may use distinct boundaries even when overall accuracy is similar. Ashby et al. (1998) Ashby & Maddox (2005)
A central body of multiple-systems research distinguishes rule-based tasks from information-integration tasks. The former often permit explicit hypothesis testing; the latter can favor gradual acquisition of nonverbalizable response policies. Waldron and Ashby (2001) examined category learning during a concurrent numerical Stroop task and found greater disruption of certain rule-based structures than the complex information-integration conditions they tested. This selective interference supports differentiated cognitive demands under those conditions. At the same time, dimensionality, stimulus difficulty, and dual-task interference must be matched carefully before a dissociation can be attributed to separate learning systems. Waldron & Ashby (2001)
Feedback timing provides a second experimental lever. Maddox, Ashby, and Bohil (2003) reported that delaying feedback impaired information-integration learning more than rule-based learning and that poorer information-integration performance was associated with greater use of rule-based strategies. Such results have specific causal strength because feedback timing is experimentally manipulated. The conclusion must nevertheless remain matched to the tested categories and procedures; it does not prove that all learning of complex categories shares one timing-sensitive circuit. Replication with comparable perceptual difficulty, different delays, and explicit process modeling improves the mechanistic interpretation. Maddox et al. (2003)
Decision boundaries remain useful regardless of whether multiple systems are ultimately required. They specify where a decision will be made and support precise predictions for boundary-adjacent transfer items. However, a boundary may be an emergent description of weighted exemplar comparisons, so geometric fit alone does not identify a decision-bound learning mechanism. Investigators can distinguish accounts by testing item-specific memory, trial-wise adaptation, selective interference, delayed feedback, and strategy reports jointly. The 2024 review by Minda and colleagues concluded that the single-system versus multiple-systems debate remains unresolved, making integrative experimental designs preferable to categorical declarations. Minda et al. (2024) Nosofsky (1986)
Incremental Learning, Error Signals, and Cluster Recruitment
Feedback permits learners to revise what they attend to, how they represent examples, and how confidently they assign categories. An error-driven learning rule updates a parameter according to the difference between the current prediction and the observed label. Such learning can alter association strengths while leaving stimulus representations fixed, or it can modify the representation itself. An experiment that records only the first and last blocks misses this distinction. Trial-level data, changes in diagnostic feature weighting, and responses immediately following a surprising example are needed to evaluate how categorical knowledge is reorganized. Kruschke (1992) Love et al. (2004)
ALCOVE demonstrates the feasibility of combining fixed exemplar positions with adjustable attention and category-output associations. SUSTAIN approaches the same broad problem by beginning with relatively simple clusters and recruiting additional clusters when prediction errors or surprising evidence require them. Such model architectures predict different trajectories when rare exceptions appear early versus late and when task goals emphasize inference rather than labeling. Their mathematical differences are important: a weighted sum over stored instances and a dynamically grown cluster network can make similar terminal choices while predicting different transitions, latency profiles, and adaptation after reversal. Kruschke (1992) Love et al. (2004)
Instructional regime changes the type of information acquired. Yamauchi and Markman (1998) contrasted classification learning, where participants predict the category from properties, with inference learning, where a category label supports prediction of an unobserved property. The experiments show why category formation cannot be fully characterized by supervised labeling accuracy alone. When a learner is asked to infer features, relational and within-category dependencies may become more important than the discriminative dimensions that maximize label accuracy. Cognitive goals are therefore variables in the learning process, not merely different ways to measure the same fixed representation. Yamauchi & Markman (1998)
Reversals and shifts in category distributions test how rapidly prior knowledge can be revised. A flexible learner may reweight features, recruit new clusters, retrieve context-specific exceptions, or infer that the task has changed. Similar adaptation curves can conceal sharply different informational commitments. Experimental designs should include recovery of earlier categories, delayed transfer after reversal, and changes in error cost. Computational frameworks also require explicit penalties for complexity to avoid explaining every exception by adding an unconstrained memory trace or cluster. An adequate account predicts both adaptive flexibility and the persistence of learned structure under interference. Love et al. (2004) Johansen & Palmeri (2002)
Causal Knowledge, Theory-Based Concepts, and Natural Kinds
Ordinary concepts often organize explanations of why properties coexist. A person can judge an unfamiliar animal to belong to a species even when its appearance is atypical, because ancestry and internal biological mechanisms are considered more diagnostic than surface features. Murphy and Medin (1985) argued that conceptual coherence depends on explanatory knowledge: features are interpreted through causal and functional relations rather than merely collected in a similarity list. This insight provides an empirical agenda. Investigators can hold perceptual similarity constant while varying a stated causal origin, and then measure classification, induction, and confidence. The results can reveal when explanatory knowledge changes which attributes carry evidential weight. Murphy & Medin (1985) Gelman & Markman (1986)
Knowledge-based categorization requires a careful distinction between the theorist's description of causal structure and the participant's actual understanding. A subject may reproduce a label taught by an experimenter without representing the explanation that links category members. Conversely, children can use a category label to project an unseen property without articulating their reason. In studies of children's induction, Gelman and Markman (1986) demonstrated that category information can guide property projections even when appearances conflict. The important finding is a dissociation between perceptual resemblance and some category-based inferences, subject to the stimuli, age group, and assumptions tested. Gelman & Markman (1986) Keil (1989)
Natural-kind concepts have been central to the developmental study of how people organize biological and physical knowledge. Keil (1989) investigated conceptual development as a transformation in the explanatory structures children attribute to kinds. Such development is not equivalent to accumulating more instances or learning a longer vocabulary. A child's new understanding of what makes an animal a particular kind can reorganize the relative importance of surface features, internal composition, and origins. Experiments assessing transformations, anomalies, and property inheritance are useful because they differentiate category membership from ordinary resemblance judgments. Their interpretation nonetheless depends on background knowledge and on the kinds of explanations made available to children. Keil (1989) Sloutsky (2010)
Theory-based accounts are strongest when they specify a mechanism for selecting hypotheses and revising beliefs under counterevidence. Merely attributing every surprising judgment to an unspecified background theory would make the account difficult to test. Researchers can instead contrast causal diagrams, category-feature correlations, and interventions designed to change one feature while preserving another. Predictions concern which unobserved attributes participants infer and how willing they are to generalize after learning a dependency. A causal relation can make a rare feature central, whereas an appearance-based procedure might assign it little weight. This contrast defines a tractable problem for combining causal knowledge with statistical category learning. Tenenbaum et al. (2006) Murphy & Medin (1985)
Bayesian Generalization, Priors, and Sampling Assumptions
Bayesian models represent category learning as inference over hypotheses given observed examples. A hypothesis may specify an extension, a generative process, a feature structure, or a relation among instances. Prior probability encodes a learner's initial expectation over those hypotheses; likelihood describes how probable the evidence would be if a candidate hypothesis were true. The posterior integrates the two. This framework clarifies why a handful of examples may support narrow or broad generalizations depending on the learner's assumptions about how examples were selected. A mathematically successful posterior prediction remains a computational-level account unless additional evidence identifies the algorithm and representations used by a participant. Tenenbaum et al. (2006) Xu & Tenenbaum (2007a)
Xu and Tenenbaum (2007) analyzed word learning with Bayesian models and experiments involving generalization from observed labeled examples. Their account related the breadth of a learner's extension to the examples supplied and to assumptions about sampling. When examples are assumed to be sampled deliberately from the target concept, repeated examples from a narrow subcategory can carry stronger evidence for a restricted meaning than examples assumed to be encountered incidentally. The critical empirical feature is the manipulation of sampling structure and category membership rather than the bare fact that people generalize. Related developmental work examined children's sensitivity to sampling processes directly. Xu & Tenenbaum (2007a) Xu & Tenenbaum (2007b)
Pedagogical examples add a social source of information. An instructor can select cases intending to reveal a generalization, and the learner may reason about that intention. Shafto, Goodman, and Griffiths (2014) presented a rational model of teaching and learning from deliberately selected examples. The framework predicts how evidence can become more informative when the learner assumes a knowledgeable teacher is choosing it. This does not imply that all educational examples are chosen optimally or that learners reliably infer the teacher's aims. A testable design contrasts random and pedagogically selected samples, checks the learner's knowledge of the selection mechanism, and measures subsequent transfer to novel cases. Shafto et al. (2014)
Structured priors can extend beyond familiar taxonomic hierarchies. Kemp and Tenenbaum (2008) investigated the inference of structural forms such as hierarchies, chains, and other relational arrangements from observations. This illustrates how the space of possible concepts can shape generalization before individual feature associations are learned. The research challenge is identifying which structural constraints a participant brought to the experiment and whether those constraints are learned, supplied by instruction, or imposed by the modeler. Predictive evaluation should compare plausible priors on withheld generalization cases, test sensitivity to new data, and separate descriptive mathematical adequacy from evidence of a psychologically implemented probabilistic computation. Kemp & Tenenbaum (2008) Tenenbaum et al. (2006)
Category-Based Induction, Relational Structure, and Analogy
Categorization allows a learner to extend information from known to unknown cases. Category-based induction can be studied by telling participants that one member has an unfamiliar property and asking whether another member possesses it. This outcome differs from membership classification: it requires inferring what the category predicts, and may depend on causal beliefs, diversity of observations, category hierarchy, and the property being projected. A learner can correctly label a member while giving an unjustified projection. Such dissociations reveal how concepts participate in reasoning, rather than only how labels are attached to stimuli. Gelman & Markman (1986) Tenenbaum et al. (2006)
Relational categories depend on common structures connecting elements rather than on shared individual attributes. Two object pairs may share a relation such as greater-than, same-function-as, or cause-of while differing visually. The capacity to identify the common relation often benefits from explicit comparison of examples. Kurtz, Boukrina, and Gentner (2013) found that comparison can promote learning and transfer of relational categories. The mechanism proposed involves alignment of corresponding roles and the extraction of structure shared across cases. A discriminating experiment uses new objects at test so successful transfer cannot be explained solely by recognition of trained appearances. Kurtz et al. (2013)
Inductive generalization may be sensitive to which cases are selected as evidence, not only to how many are observed. An example set concentrated in one corner of a category can mislead a learner when sampling is assumed representative; deliberate selection from a teacher may have a different evidential interpretation. Bayesian sampling research makes this distinction explicit. Analogical alignment supplies another route: two examples that look different can illuminate a relational invariant if the task encourages comparison. These routes interact when the hypothesized category is relational and the examples are pedagogically chosen. Experimental evaluation should independently manipulate sample diversity, alignment opportunities, and cues to the teacher's selection process. Xu & Tenenbaum (2007b) Shafto et al. (2014) Kurtz et al. (2013)
Category-based reasoning depends on the direction of inquiry. Classification asks for a label given features; induction asks for features or outcomes given a label; explanation asks why the observed configuration occurs. Yamauchi and Markman (1998) showed that inference and classification learning can yield different representations and generalization patterns under controlled tasks. This makes the format of training an important causal variable. Researchers should compare identical stimulus statistics under each objective, measure remembered feature correlations, and test both kinds of judgment after learning. A concept is functionally richer when it can support several forms of inference without relying on new memorization in each task. Yamauchi & Markman (1998) Murphy & Medin (1985)
Infant and Child Category Development
Developmental research asks how categories change as learners acquire perceptual experience, words, causal explanations, and opportunities for structured comparison. Infants can discriminate statistical regularities and group stimuli under restricted laboratory conditions, while older children can explicitly name, explain, and generalize categories. Different measures therefore index different capacities: preferential looking and habituation reveal sensitivity to distinctions; object choice and novel-label extension reveal task-specific inferences; explanations probe accessible conceptual knowledge. Researchers should avoid treating an infant's perceptual grouping response as direct evidence for the complete semantic organization observed in adults. Sloutsky (2010) Gelman & Markman (1986)
Gelman and Markman's (1986) studies demonstrated the role of categories in children's induction of nonobvious properties. Such tasks can pit a familiar kind label against superficial resemblance, demonstrating that children's reasoning sometimes privileges kind membership. However, the strength of label-based inference depends on the properties involved, the task's pragmatic assumptions, and participants' knowledge of the kinds under study. A child may respond to an unfamiliar label as an invitation to infer shared properties without understanding its biological significance. Distinguishing semantic knowledge from the social and statistical effects of naming remains central to developmental experimentation. Gelman & Markman (1986) Xu & Tenenbaum (2007a)
Sloutsky (2010) reviewed the transition from perceptual categorization to richer concepts and emphasized interactions among attention, similarity, language, and conceptual development. The resulting picture permits both continuity and change: experience modifies the weighting of visible cues while teaching and domain knowledge open possibilities for inference about hidden causes. An empirical developmental account should specify the sequence of capabilities and avoid presupposing that all children pass through identical discrete stages. Cross-sectional comparisons can reveal age-group differences, but longitudinal observation and carefully matched training designs are needed to establish the processes that generate those differences. Sloutsky (2010) Keil (1989)
The linguistic and cultural environment shapes which category contrasts are named and how examples are presented. Such influences invite investigation of category acquisition across different languages and populations, with methods that separate differences in prior exposure from differences in task comprehension. Researchers should equate materials for familiarity and allow appropriate local meanings rather than assume that translated English labels define the same concept. An apparent developmental or cultural gap can arise when a task privileges one education history. The strongest generalizations combine comparable experimental structure with domain-appropriate stimuli, open reporting of sampling contexts, and tests of how concepts are used outside a single sorting task. Sloutsky (2010) Murphy & Medin (1985)
Language, Labels, and Instruction in Concept Learning
A word may mark a category, direct attention to shared structure, or convey that a teacher intends the grouped objects to form a meaningful kind. These are different mechanisms. A label can improve classification by serving as a stable cue while contributing little explanatory knowledge; it can also become the entry point into a system of relational, causal, and social expectations. Consequently, vocabulary acquisition and conceptual acquisition overlap without being identical. Word-learning experiments are particularly informative when labels are matched across conditions but example selection, property distributions, and explicit instructions vary independently. Xu & Tenenbaum (2007a) Gelman & Markman (1986)
Language provides access to categories that are difficult to infer from isolated appearances. Causal explanations, taxonomic descriptions, counterfactual statements, and analogies can supply relations absent from short training sets. Theory-based approaches predict that verbally taught background knowledge should alter classification and induction even when the perceptual stimuli remain constant. Yet instructional effects can also reflect demand characteristics and verbal memory. A rigorous design asks whether learning survives a response-format change or generalizes to cases on which the taught phrase itself provides no direct answer. Such transfer is stronger evidence for usable conceptual knowledge than repetition of an explanation. Murphy & Medin (1985) Keil (1989)
The structure of examples matters as much as the presence of a verbal explanation. A tutor who selects informative contrasting cases can signal an intended rule, while randomly sampled examples may require the learner to estimate a distribution. Shafto and colleagues (2014) formalized this pedagogical interaction through reciprocal assumptions about knowledgeable teaching and receptive learning. The framework helps analyze why examples chosen for instruction are often more useful than their raw frequency would suggest. Its empirical application demands direct verification of the learner's interpretation of pedagogical intent and tests that distinguish teaching effects from simple exposure. Shafto et al. (2014) Xu & Tenenbaum (2007b)
Conceptual instruction can further exploit comparison: paired examples reveal similarities, differences, and correspondences that a single example conceals. Relational-category experiments demonstrate transfer benefits when learners align structurally similar cases. Effective experimental inquiry should ask which comparison was induced, how long it persisted, and whether a new task requires extending the relationship to unfamiliar objects. Instructional conclusions require attention to learners' prior knowledge and to the possibility that the same explanation helps one group but diverts another toward unhelpful features. The science of instruction thus depends on the same careful control of features, rules, examples, and transfer that underpins basic category-learning research. Kurtz et al. (2013) Kornell & Bjork (2008)
Neuroscience of Category Acquisition and Memory Systems
Neural approaches seek to identify brain systems participating in feature selection, feedback use, memory retrieval, and category decisions. A prominent account proposes partially differentiated learning systems for hypothesis-guided, explicit categorization and incremental information integration. Ashby, Alfonso-Reese, Turken, and Waldron (1998) developed a neuropsychological theory linking these functions to interacting cortical and subcortical processes. Such models contribute a mechanistic research program because they specify how behavioral dissociations may relate to neural organization. Evidence for a brain region's activation during a categorization task alone, however, establishes involvement under the tested conditions rather than a unique representation of a concept. Ashby et al. (1998)
Poldrack and colleagues (2001) examined category-learning performance and brain activity during a weather-prediction-type probabilistic classification task, studying interactions between memory systems as learning proceeded. Their work is relevant because category acquisition can involve both distributed declarative knowledge and response-sensitive learning, whose relative contributions change with task conditions. The interpretation of an imaging result requires attention to the specific training demands, feedback contingencies, and contrast tasks. A region's greater activity may reflect memory retrieval, strategic control, novelty, or the costs of performing the task rather than category representation itself. Behavioral and neural predictions must therefore be jointly specified. Poldrack et al. (2001) Ashby & Maddox (2005)
Selective experimental manipulations strengthen system claims. Dual-task interference can preferentially disrupt a category task that requires executive hypothesis testing, while delayed feedback can impair a learning regime relying on close temporal association between response and outcome. Waldron and Ashby (2001) and Maddox, Ashby, and Bohil (2003) provide classic tests of these ideas. Their outcomes contribute evidence for differences in learning demands and processing pathways but must be reconciled with alternative explanations involving task difficulty, verbal accessibility, feedback utility, or strategy shifts. A neural theory should predict which distinctions remain after those variables are controlled. Waldron & Ashby (2001) Maddox et al. (2003)
Memory systems and category-learning processes need not map one-to-one. Explicit retrieval can support a supposedly procedural classification task, and a learner may use similarity without deliberate recall. Neural activity may be shared across mechanisms, and experimentally defined categories often recruit several processes simultaneously. The current theoretical debate therefore concerns the necessity, organization, and interaction of multiple systems, rather than whether many brain regions participate in learning. Minda and colleagues (2024) review how behavioral, computational, and neuroscientific findings can support competing single- and multiple-system models, motivating experiments that compare distinctive mechanistic predictions across convergent measures. Minda et al. (2024) Poldrack et al. (2001)
Transfer, Retention, and Category Generalization
Transfer is the application of acquired categorical knowledge to cases beyond the training examples or context. Near transfer may involve modest changes to familiar features, whereas more demanding transfer changes dimensions, modalities, task objectives, or response requirements. Researchers should specify exactly what changed and what structure remained invariant. Successful labeling of a distorted known prototype differs from detecting a relational category in objects with entirely new appearances. In category experiments, withheld test items are essential because memorization of the training set can produce apparently impressive learning while supplying little evidence about the breadth of the acquired concept. Posner & Keele (1968) Kurtz et al. (2013)
Generalization depends on psychological similarity, learned decision rules, causal beliefs, and assumptions about how the examples were sampled. A central unseen stimulus may be readily classified under both prototype and exemplar accounts, whereas a rule-violating but perceptually similar item can separate them. Likewise, a causal transformation can alter category membership with relatively little visible change. These tests are stronger when alternative hypotheses generate predictions before data collection and when learners have enough exposure for each model's proposed mechanism to operate. A single transfer score should not be treated as a generic measure of abstractness because each probe reflects a specified generalization relation. Nosofsky (1986) Murphy & Medin (1985)
Retention raises additional questions about what survives after the immediate training context disappears. Recognition of individual examples may weaken while category-level centrality persists, or a learned explicit rule may remain articulate even as speed and accuracy deteriorate. Such patterns do not uniquely determine a representational transition; they may reflect differences in retrieval accessibility or response thresholds. Longitudinal testing should preserve the distinction between retention of item memory, retention of category decisions, and retention of inferential relations. A design that combines recognition, identification, induction, and withheld transfer offers more diagnostic coverage than retesting the original labels. Johansen & Palmeri (2002) Love et al. (2004)
Novel contexts can expose fragile assumptions in both people and models. A classifier that learned a texture cue under one stimulus distribution may fail when texture is decorrelated from class membership, while a learner who acquired a causal or relational invariant may transfer more successfully. Computational simulations allow these distinctions to be defined explicitly and tested against human responses. Flesch, Saxe, and Summerfield (2023) review continual learning across natural and artificial agents, highlighting the significance of task structure and transfer. Such cross-system comparisons should record differences in available training examples, embodiment, feedback, and task instructions before attributing divergent performance to general properties of intelligence. Flesch et al. (2023) Kemp & Tenenbaum (2008)
Effects of Training Regime, Interleaving, and Feedback
The arrangement of learning examples alters what comparisons a learner can make. Blocked study presents repeated members of one category before switching, encouraging within-category similarity and perhaps prototype extraction. Interleaved study mixes different categories, potentially increasing opportunities to discriminate them. Kornell and Bjork (2008) reported that interleaving paintings by different artists supported better later classification of new paintings than blocking under their experimental conditions. The result illustrates the potential value of contrasts between categories, although its educational generalization depends on material complexity, task goals, prior knowledge, and retention interval. Kornell & Bjork (2008)
Feedback may provide a category label, a reward signal, a correction, or explanatory information. Its psychological effects depend on which uncertainty it resolves. In a rule-search task, an error can eliminate a hypothesis; in an incremental associative model, the same event can change connection weights; in a cluster model, it may motivate a new subcategory. Manipulations of feedback delay are particularly useful because they reveal temporal requirements for learning, but the interpretation must consider whether a learner can maintain the relevant stimulus and response representations across the delay. The delayed-feedback experiments by Maddox and colleagues (2003) are one influential comparison. Maddox et al. (2003) Kruschke (1992)
The order and diversity of examples may also influence the discovery of relations. Two examples presented close together can reveal a shared structure, whereas repeated examples from one subcategory may encourage an overly narrow boundary. Comparison-based relational learning and Bayesian sampling analyses provide complementary explanations for these effects. Researchers can cross example order with category overlap and explicit teaching, then measure feature weighting and induction separately. The value of an instructional arrangement should be evaluated by delayed transfer and the ability to infer new properties, not solely by immediate training speed. Faster acquisition can arise from shortcuts that later reduce flexibility. Kurtz et al. (2013) Xu & Tenenbaum (2007b)
Training design research should distinguish interventions that are effective for a specific category family from those that generalize across domains. Interleaving can aid discriminative learning when the main difficulty is distinguishing confusable classes, while other tasks may benefit from periods of concentrated within-class study. Theories make more progress by predicting such interactions than by treating a single scheduling rule as universally optimal. Evidence should include random assignment where feasible, training-equated exposure, predeclared outcomes, and transfer items that reflect actual intended use. The robust result sought is a conditional mechanism: how example arrangement changes the information available for category abstraction. Kornell & Bjork (2008) Yamauchi & Markman (1998)
Individual Variation, Strategy Shifts, and Cultural Scope
Category-learning experiments frequently display meaningful differences among participants. Some learners discover a compact rule early, others accumulate exemplar evidence, and still others alternate strategies as feedback changes. A model fitted to the mean acquisition curve can disguise those distinct trajectories. RULEX explicitly permits variability in rule search and exception learning; Johansen and Palmeri (2002) used trial-wise generalization data to investigate representational change across learning. These programs show why within-person trajectories and critical transfer items matter. Individual learning profiles should be reported alongside aggregate outcomes rather than treated as unexplained error around a supposedly universal strategy. Nosofsky et al. (1994, RULEX) Johansen & Palmeri (2002)
Variability can arise from the perceived dimensions of a stimulus, prior familiarity, attentional control, memory retrieval, and expectations about the task. A participant experienced with a natural domain may possess a rich explanatory concept before entering the laboratory; a novice may rely on visible similarity. For this reason, investigators should distinguish learning the experimental response mapping from building a genuinely new conceptual organization. Baseline assessments of prior knowledge, item-level memory, and verbalizable rules help interpret performance. The relative influence of rule and exemplar information may change within the same individual depending on available time and incentives. Erickson & Kruschke (1998) Murphy & Medin (1985)
Many classic category-learning datasets were gathered with relatively small and homogeneous samples performing brief visual classification tasks. Their internal control is scientifically valuable, but generalizing to different educational histories, languages, ages, and cultural experiences requires renewed testing. The target categories themselves may be familiar or unfamiliar in different communities. Experimental materials should be selected and translated with attention to local conceptual distinctions; a supposedly universal feature may carry different diagnostic or semantic meaning elsewhere. Formal models should be tested for parameter stability and predictive calibration across populations rather than assuming that group-level fits imply invariance. Sloutsky (2010) Minda et al. (2024)
Researchers should also examine how strategy reports relate to actual behavior. A learner can accurately describe a rule yet employ stored exceptions, or can develop a stable decision boundary without being able to explain it. A verbal report is evidence about accessible knowledge, not a transparent recording of the entire process. Converging measures might combine response time, eye movement, post-training explanation, neural data, and model-based predictions for held-out items. When these measures disagree, the conflict can reveal a layered division of labor between representations and response strategies. Explaining such divergence is an important part of the psychology of concepts. Ashby & Maddox (2005) Nosofsky (1986)
Artificial Category Models and Human–Artificial Comparison
Artificial systems provide computational instantiations of category representations whose training procedures and internal parameters can sometimes be inspected directly. Prototype classifiers, exemplar-style memory models, discriminative networks, Bayesian models, and large representation-learning systems differ in what information they preserve and how they generalize. Their comparative significance comes from explicit implementation: an algorithm can be trained on precisely specified examples, tested on withheld categories, and analyzed for how learned boundaries respond to distribution shift. Functional resemblance to human categorization is a research result when independently demonstrated under comparable tasks, while claims about shared processes require additional mechanistic evidence. Kruschke (1992) Flesch et al. (2023)
An artificial categorizer may outperform humans on a high-volume perceptual sorting benchmark yet fail at explanations about causal origin or relational structure. Conversely, a language-trained system may express a plausible semantic rule without robustly applying it to unfamiliar instances. These asymmetries show why classification accuracy alone cannot establish conceptual competence in the broad sense defined in this monograph. A comparative evaluation should separate new-instance accuracy, novel-feature induction, sensitivity to causal interventions, robustness to changed exemplars, and explicit interpretation of labels. The tasks should be selected before observing results and should prevent memorized benchmark answers from substituting for new inference. Murphy & Medin (1985) Kemp & Tenenbaum (2008)
Continual learning adds the temporal dimension. When category structures change, an artificial model may overwrite earlier distinctions or preserve them through replay, parameter separation, or modular organization. Natural learners may similarly show interference, context dependence, and asymmetric retention, though the physical mechanisms differ. Flesch and colleagues (2023) discuss the relationship between continual task learning in natural and artificial agents. A useful research program characterizes the training history, adaptation costs, return to earlier tasks, and cross-task generalization for each agent, thereby making claims of adaptation and conceptual stability quantitatively comparable. Flesch et al. (2023) Love et al. (2004)
Within the Artificial Era framework, concept formation becomes a test case for the general scientific comparison of cognitive functions across biological and artificial systems. The relevant constructs are observable generalization, representational differentiation, inferential coherence, stability, and adaptation. These can be operationalized without presupposing identical implementation or settling philosophical questions about subjective experience. Model architecture can be investigated alongside behavioral evidence, while the philosophical vocabulary of Artificial and Artificial Sapiens remains identifiable as authorial context. The scientific task is to discover which capacities occur in which systems, under what conditions, with which mechanisms, and with what boundaries of transfer. Tenenbaum et al. (2006) Minda et al. (2024)
Experimental Designs and Measurement Validity
A valid categorization study begins with a specified target construct and an experimental contrast capable of distinguishing explanations. If the question concerns prototype abstraction, a suitable design includes previously unseen central items and controls for their similarity to individual training examples. If the question concerns rule search, probes must separate competing candidate predicates. If the question concerns causal concepts, new instances should manipulate explanatory origins independently of surface resemblance. The study must specify what participants know before training, how category feedback is delivered, and which responses will count as evidence of representation rather than repetition of a practiced label. Posner & Keele (1968) Nosofsky et al. (1994, RULEX) Murphy & Medin (1985)
The basic observables include choice accuracy, response time, confidence, generalization gradients, typicality, feature predictions, and individual trajectories. Each measures a different aspect of the phenomenon. An accuracy difference between conditions can arise from sensitivity, response bias, guessing, criterion movement, or differential familiarity with the stimuli. Response latency may reflect perceptual encoding, decision conflict, preparation, or motor execution. Even typicality ratings depend on explicit instructions and the set of compared exemplars. Construct validity improves when investigators use several measures that converge under the hypothesized mechanism and diverge when the competing mechanism predicts a different pattern. Nosofsky (1986) Goldstone (1994)
The factorial manipulation of training structure, feedback regime, stimulus similarity, and test format can isolate interactions but may also generate complex designs with limited power for individual predictions. Investigators should predefine primary contrasts and collect enough trials to estimate each participant's learning trajectory. When categorical outcomes are repeated within participants and stimuli, hierarchical statistical models help account for both sources of variation. Items and participants should be sampled from the populations to which conclusions are intended to generalize. Reporting the full set of exclusion rules, stimuli, order randomization, and analysis decisions permits later investigators to evaluate whether an effect depends on a particular stimulus list or analytic choice. Johansen & Palmeri (2002) Minda et al. (2024)
Some experiments use existing natural categories, while others create wholly artificial ones. These choices have different advantages: familiar concepts reveal real semantic and causal knowledge, whereas invented stimuli permit control of feature statistics and prior exposure. Combining them is particularly informative when the same proposed mechanism is evaluated across conditions. Transfer tests should include both items near the trained distribution and genuine extrapolation probes; a model that succeeds only by interpolation has a narrower scope. Researchers should publish the stimulus-generating procedure and, where licensing allows, the actual stimulus set so an independent laboratory can reconstruct the decisive comparisons. Rosch & Mervis (1975) Medin & Schaffer (1978) Tenenbaum et al. (2006)
Computational Model Fitting, Recovery, and Predictive Tests
Formal modeling converts competing explanations into probability distributions over responses, response times, or feature judgments. For an exemplar model, central parameters may include attention weights, similarity sensitivity, response determinism, and the memory strength of learned instances. Prototype models specify how central tendencies and noise generate choices. Rule models specify hypothesis generation, exceptions, and search costs; clustering models specify recruitment, competition, and adaptation. Models should be fitted to the outcomes they purport to explain, and parameters must be constrained by the actual experimental design. Flexibility is scientifically useful when it yields new predictions, but uncontrolled flexibility can mimic several mechanisms without identifying any of them. Nosofsky (1986) Kruschke (1992) Love et al. (2004)
A benchmark model comparison is informative only when alternatives have a fair opportunity to explain the data. The prototype-versus-exemplar debate illustrates this requirement: Minda and Smith (2002) and Zaki and colleagues (2003) reached different assessments partly because model specification and assumptions about attention and decisions matter. A minimum comparison should fit plausible implementations of each theory using equivalent outcome definitions and transparently described parameter constraints. Held-out stimuli, held-out participants where appropriate, and transfer to changed category structures test prediction beyond the training observations. Better training likelihood alone is insufficient to establish a richer psychological explanation. Minda & Smith (2002) Zaki et al. (2003)
Parameter recovery asks whether simulated data generated by a model with known parameters allow those parameters to be estimated accurately under the proposed experimental schedule. Model recovery asks whether the analysis can distinguish datasets generated by one candidate model from datasets generated by alternatives. These simulation-based checks are crucial when the experimental structure is small and multiple accounts fit the same responses. Researchers should explore recovery across realistic ranges of accuracy and noise rather than only at idealized parameter settings. Failure to recover a mechanism indicates that the experiment or estimation procedure cannot support a strong claim about that mechanism, regardless of the apparent precision of a fitted point estimate. Minda et al. (2024) Johansen & Palmeri (2002)
Model comparison should be joined to experimental design rather than postponed until a dataset is complete. Before collecting new responses, a researcher can simulate which probe items create the largest differences between candidate model predictions. An exemplar account may be tested on exceptions close to known instances, a prototype account on untrained central cases, and a causal account on visually similar objects with altered generating mechanisms. A design optimized for discriminating predictions is more informative than one consisting only of easy category members. Calibration, predictive uncertainty, individual differences, and qualitative errors should be evaluated alongside average accuracy. Medin & Schaffer (1978) Tenenbaum et al. (2006) Nosofsky et al. (1994, RULEX)
Evidence Appraisal and Reproducibility
The evidential base spans controlled laboratory experiments, formal simulations, developmental studies, neuroimaging research, behavioral interventions, and theoretical syntheses. Each method supports a different kind of inference. Randomized manipulation of example order or feedback timing can support a causal claim about the effect of that manipulation on measured learning. A model's good fit demonstrates quantitative compatibility with observed behavior but is insufficient to establish the uniqueness of its internal mechanism. Neuroimaging identifies patterns of association between activity and task performance unless a more discriminating design supports a specific mechanistic account. These distinctions are central to evaluating the claims of category-learning science. Maddox et al. (2003) Poldrack et al. (2001) Minda et al. (2024)
Several historically important findings have been reexamined or replicated within the research tradition. Nosofsky and colleagues (1994) replicated and extended the classic category-structure results of Shepard, Hovland, and Jenkins (1961), illustrating how a benchmark can test subsequent models. The continuing prototype–exemplar debate shows why data replication and explanatory settlement should be distinguished: the same reliable behavior can remain compatible with alternative decision and representational processes. Independent labs can advance the issue by specifying the theory-relevant contrasts, testing new stimuli, and publishing unsuccessful model discriminations along with positive results. Shepard et al. (1961) Nosofsky et al. (1994, replication) Zaki et al. (2003)
Evidence quality also depends on sampling and reporting. Small artificial-category studies may provide precise within-experiment contrasts while offering limited evidence about generalization to large, naturalistic category spaces. Analysis of aggregate averages can conceal heterogeneity, especially when subjects switch from one strategy to another. A scientifically complete report should include descriptive distributions, uncertainty intervals for main contrasts, participant-level trajectories where available, exact model specifications, and sensitivity to different exclusion and fitting decisions. For reviews of the literature, findings should be connected to the original design rather than aggregated into the misleading impression of one undifferentiated effect. Johansen & Palmeri (2002) Minda et al. (2024)
This monograph is a critical source-based synthesis of selected original research and reviews. It does not represent a protocol-registered systematic search, effect-size meta-analysis, new experiment, or independently reproduced dataset. Its conclusions are calibrated to the published designs and independently interpretable comparisons. Wherever original records were available, historical attributions, publication details, and the logic of the experimental manipulations are identified by direct source links. The resulting synthesis provides a framework for asking better questions and designing studies that discriminate rival mechanisms, while quantitative claims about specific effect sizes require direct verification of the corresponding original analyses. Ashby & Maddox (2005) Minda et al. (2024)
Competing Explanations and Unresolved Scientific Debates
One long-running debate concerns the number and kind of representations supporting categorization. Prototype theories emphasize central tendencies; exemplar theories preserve cases; rule theories specify predicates and exceptions; causal and Bayesian approaches make explanatory and structural assumptions explicit. The same learner may recruit several representations, so a good theory must specify when each contributes and how they interact. It must also identify conditions that would falsify its interpretation. A general model that can flexibly imitate every result may offer useful prediction but weak causal explanation unless its key operations are measured or manipulated independently. Minda & Smith (2002) Zaki et al. (2003) Murphy & Medin (1985)
A second debate concerns whether multiple neural or cognitive systems are required. Rule-based and information-integration task dissociations, including dual-task interference and delayed feedback effects, support claims of differentiated processing under the tested conditions. Single-system or hybrid models can sometimes reproduce differences by varying attention, task demands, learning rates, or accessible rules. The challenge is to derive predictions for circumstances in which the accounts meaningfully diverge and then measure neural and behavioral outcomes together. The 2024 review by Minda, Roark, Kalra, and Cruz locates this question at the center of contemporary research rather than treating it as resolved. Waldron & Ashby (2001) Maddox et al. (2003) Minda et al. (2024)
A third debate concerns the relation between similarity, language, and causal knowledge. The existence of graded typicality shows that category structure is often nonuniform, but it does not demonstrate that all concepts are stored as perceptual gradients. Category-based induction in children, ad hoc goal-derived categories, and relational transfer require attention to what is inferred and why. Theories must specify whether apparently abstract performance can be accounted for by extensive statistical experience or needs additional representational machinery. These questions can be tested through cases in which perceptual overlap and explanatory relations provide opposing guidance. Rosch & Mervis (1975) Gelman & Markman (1986) Barsalou (1983) Kurtz et al. (2013)
A fourth debate concerns generalization across tasks, development, and artificial systems. A short categorization experiment may demonstrate successful local induction yet leave open whether the resulting representation supports explanations, future prediction, or robust adaptation after a change in the environment. Neurobiological evidence and machine-learning implementations can constrain accounts but must not be treated as automatic translations of psychological constructs. Progress depends on convergent tests across task formats, longitudinal histories, structured interventions, and well-calibrated mathematical comparisons. Categories are practical engines of generalization, but the conditions of successful generalization remain the central object of explanation. Flesch et al. (2023) Sloutsky (2010) Tenenbaum et al. (2006)
Research Agenda and Integrative Scientific Synthesis
The available evidence supports an integrated research agenda organized around representation, learning, decision, inference, and transfer. First, experiments should measure not only which labels participants assign but also the structure they use in property prediction, exception handling, and explanation. Second, candidate formal models should be compared on new instances selected to discriminate their predictions. Third, tests should track how representations evolve across a known sequence of examples and feedback rather than infer dynamic mechanisms exclusively from terminal accuracy. These requirements connect the traditional psychological study of concepts with formal modeling and developmental inquiry. Medin & Schaffer (1978) Love et al. (2004) Gelman & Markman (1986)
An especially informative design crosses category geometry with instructional regime. The same category structure could be learned through ordinary classification, feature inference, pedagogically chosen examples, or comparisons of relational cases. Models would make distinct predictions about attention weights, memory for exceptions, acquisition trajectories, and generalization. A delayed follow-up could test stability, while a controlled change in feature–label correlations could measure adaptation. This design links the mechanisms articulated by Yamauchi and Markman (1998), Shafto and colleagues (2014), and Kurtz and colleagues (2013) without presuming that the outcomes of their separate paradigms are interchangeable. Yamauchi & Markman (1998) Shafto et al. (2014) Kurtz et al. (2013)
Model identifiability should be treated as part of scientific discovery. Investigators can simulate the proposed experiment under exemplar, prototype, rule, clustering, and causal structures, check whether the parameters and models can be recovered, and revise the choice of test items if competing mechanisms remain observationally indistinguishable. Independent perceptual identification data can constrain similarity geometry; explicit feature predictions can constrain inferred category structure; neural or dual-task manipulations can test processing commitments. The strongest explanatory conclusions emerge when several independent observations favor the same mechanism while plausible alternatives make different, unsupported predictions. Nosofsky (1986) Minda et al. (2024) Poldrack et al. (2001)
The same research logic can inform comparison across biological and artificial agents. Artificial systems enable controlled manipulation of training distributions and computational components, while human experiments reveal attentional, inferential, and social processes not captured by a simple classification benchmark. Structured functional comparisons can establish which operations support stable categories, including under distribution shift and continual learning. A scientifically useful conclusion would specify the tested function and its evidence, rather than importing broader philosophical commitments into the experimental result. In the Artificial Era, this discipline of comparison can expand the empirical map of categorization while preserving the historical depth of the psychological research tradition. Flesch et al. (2023) Kemp & Tenenbaum (2008)
Conclusion
Concept formation and categorization organize experience into representations that support discrimination, prediction, inference, and action. A category may be learned through central tendencies, remembered instances, explicit rules, dynamic clusters, or knowledge of causal and relational structure. These mechanisms produce overlapping patterns of classification and can interact within the same learner. The central accomplishment of the experimental tradition is to turn broad conceptual questions into controllable distinctions: feature relevance, category structure, feedback, sample selection, similarity, delay, and transfer can each be manipulated to reveal how categorical knowledge is acquired and used. Shepard et al. (1961) Rosch & Mervis (1975) Love et al. (2004)
The reviewed evidence establishes graded typicality, sensitivity to individual exemplars, learnability differences among rules, feedback-dependent adaptation, effects of attention and training schedules, and the importance of inferential goals. At the same time, the explanatory interpretation of many familiar results remains contested, particularly when alternative representations can fit the same final choices. Research on children, natural kinds, language, and neural mechanisms shows why a complete account must extend beyond the classification of simplified visual stimuli. Formal models become especially powerful when used to design experiments that could prove their distinctive commitments wrong. Gelman & Markman (1986) Ashby & Maddox (2005) Minda et al. (2024)
The mature scientific task is to explain when and how a learner preserves particular experiences, compresses shared structure, discovers a rule, infers an unseen cause, and transfers knowledge under changed conditions. Such explanations require experimental precision, representational clarity, longitudinal evidence, and transparent predictive comparison. Human–artificial research offers additional opportunities to inspect computational mechanisms and test functional parallels under controlled conditions. Concept formation remains a foundational problem because it joins perception, memory, learning, language, development, reasoning, and agency in the production of usable knowledge. Its theoretical significance is measured by how effectively a model predicts those connections across new tasks and new evidence. Tenenbaum et al. (2006) Flesch et al. (2023) Minda et al. (2024)
References
- Ashby, F. G., Alfonso-Reese, L. A., Turken, A. U., & Waldron, E. M. (1998). A neuropsychological theory of multiple systems in category learning. Psychological Review, 105(3), 442–481. https://doi.org/10.1037/0033-295X.105.3.442
- Ashby, F. G., & Maddox, W. T. (2005). Human category learning. Annual Review of Psychology, 56, 149–178. https://doi.org/10.1146/annurev.psych.56.091103.070217
- Barsalou, L. W. (1983). Ad hoc categories. Memory & Cognition, 11(3), 211–227. https://doi.org/10.3758/BF03196968
- Erickson, M. A., & Kruschke, J. K. (1998). Rules and exemplars in category learning. Journal of Experimental Psychology: General, 127(2), 107–140. https://doi.org/10.1037/0096-3445.127.2.107
- Feldman, J. (2000). Minimization of Boolean complexity in human concept learning. Nature, 407, 630–633. https://doi.org/10.1038/35036586
- Flesch, T., Saxe, A., & Summerfield, C. (2023). Continual task learning in natural and artificial agents. Trends in Neurosciences, 46(3), 199–210. https://doi.org/10.1016/j.tins.2022.12.006
- Gelman, S. A., & Markman, E. M. (1986). Categories and induction in young children. Cognition, 23(3), 183–209. https://doi.org/10.1016/0010-0277(86)90034-X
- Goldstone, R. L. (1994). Influences of categorization on perceptual discrimination. Journal of Experimental Psychology: General, 123(2), 178–200. https://doi.org/10.1037/0096-3445.123.2.178
- Johansen, M. K., & Palmeri, T. J. (2002). Are there representational shifts during category learning? Cognitive Psychology, 45(4), 482–553. https://doi.org/10.1016/S0010-0285(02)00505-4
- Keil, F. C. (1989). Concepts, kinds, and cognitive development. MIT Press. https://mitpress.mit.edu/9780262111317/concepts-kinds-and-cognitive-development/
- Kemp, C., & Tenenbaum, J. B. (2008). The discovery of structural form. Proceedings of the National Academy of Sciences, 105(31), 10687–10692. https://doi.org/10.1073/pnas.0802631105
- Kornell, N., & Bjork, R. A. (2008). Learning concepts and categories: Is spacing the “enemy of induction”? Psychological Science, 19(6), 585–592. https://doi.org/10.1111/j.1467-9280.2008.02127.x
- Kruschke, J. K. (1992). ALCOVE: An exemplar-based connectionist model of category learning. Psychological Review, 99(1), 22–44. https://doi.org/10.1037/0033-295X.99.1.22
- Kurtz, K. J., Boukrina, O., & Gentner, D. (2013). Comparison promotes learning and transfer of relational categories. Journal of Experimental Psychology: Learning, Memory, and Cognition, 39(4), 1303–1310. https://doi.org/10.1037/a0031847
- Love, B. C., Medin, D. L., & Gureckis, T. M. (2004). SUSTAIN: A network model of category learning. Psychological Review, 111(2), 309–332. https://doi.org/10.1037/0033-295X.111.2.309
- Maddox, W. T., Ashby, F. G., & Bohil, C. J. (2003). Delayed feedback effects on rule-based and information-integration category learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 29(4), 650–662. https://doi.org/10.1037/0278-7393.29.4.650
- Medin, D. L., & Schaffer, M. M. (1978). Context theory of classification learning. Psychological Review, 85(3), 207–238. https://doi.org/10.1037/0033-295X.85.3.207
- Minda, J. P., & Smith, J. D. (2002). Comparing prototype-based and exemplar-based accounts of category learning and attentional allocation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 28(2), 275–292. https://doi.org/10.1037/0278-7393.28.2.275
- Minda, J. P., Roark, C. L., Kalra, P., & Cruz, A. (2024). Single and multiple systems in categorization and category learning. Nature Reviews Psychology, 3, 536–551. https://doi.org/10.1038/s44159-024-00336-7
- Murphy, G. L., & Medin, D. L. (1985). The role of theories in conceptual coherence. Psychological Review, 92(3), 289–316. https://doi.org/10.1037/0033-295X.92.3.289
- Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of Experimental Psychology: General, 115(1), 39–57. https://doi.org/10.1037/0096-3445.115.1.39
- Nosofsky, R. M., Palmeri, T. J., & McKinley, S. C. (1994). Rule-plus-exception model of classification learning. Psychological Review, 101(1), 53–79. https://doi.org/10.1037/0033-295X.101.1.53
- Nosofsky, R. M., Gluck, M. A., Palmeri, T. J., McKinley, S. C., & Glauthier, P. (1994). Comparing models of rule-based classification learning: A replication and extension of Shepard, Hovland, and Jenkins (1961). Memory & Cognition, 22(3), 352–369. https://doi.org/10.3758/BF03200862
- Poldrack, R. A., Clark, J., Paré-Blagoev, E. J., Shohamy, D., Creso Moyano, J., Myers, C., & Gluck, M. A. (2001). Interactive memory systems in the human brain. Nature, 414, 546–550. https://doi.org/10.1038/35107080
- Posner, M. I., & Keele, S. W. (1968). On the genesis of abstract ideas. Journal of Experimental Psychology, 77(3), 353–363. https://doi.org/10.1037/h0025953
- Rosch, E., & Mervis, C. B. (1975). Family resemblances: Studies in the internal structure of categories. Cognitive Psychology, 7(4), 573–605. https://doi.org/10.1016/0010-0285(75)90024-9
- Shafto, P., Goodman, N. D., & Griffiths, T. L. (2014). A rational account of pedagogical reasoning: Teaching by, and learning from, examples. Cognitive Psychology, 71, 55–89. https://doi.org/10.1016/j.cogpsych.2013.12.004
- Shepard, R. N., Hovland, C. I., & Jenkins, H. M. (1961). Learning and memorization of classifications. Psychological Monographs: General and Applied, 75(13), Whole No. 517. https://worthylab.org/wp-content/uploads/2021/08/shephardetal1961_classificationlearning_psychologicalmonographs.pdf
- Sloutsky, V. M. (2010). From perceptual categories to concepts: What develops? Cognitive Science, 34(7), 1244–1286. https://doi.org/10.1111/j.1551-6709.2010.01129.x
- Tenenbaum, J. B., Griffiths, T. L., & Kemp, C. (2006). Theory-based Bayesian models of inductive learning and reasoning. Trends in Cognitive Sciences, 10(7), 309–318. https://doi.org/10.1016/j.tics.2006.05.009
- Waldron, E. M., & Ashby, F. G. (2001). The effects of concurrent task interference on category learning: Evidence for multiple category learning systems. Psychonomic Bulletin & Review, 8(1), 168–176. https://doi.org/10.3758/BF03196154
- Xu, F., & Tenenbaum, J. B. (2007). Word learning as Bayesian inference. Psychological Review, 114(2), 245–272. https://doi.org/10.1037/0033-295X.114.2.245
- Xu, F., & Tenenbaum, J. B. (2007). Sensitivity to sampling in Bayesian word learning. Developmental Science, 10(3), 288–297. https://doi.org/10.1111/j.1467-7687.2007.00590.x
- Yamauchi, T., & Markman, A. B. (1998). Category learning by inference and classification. Journal of Memory and Language, 39(1), 124–148. https://doi.org/10.1006/jmla.1998.2566
- Zaki, S. R., Nosofsky, R. M., Stanton, R. D., & Cohen, A. L. (2003). Prototype and exemplar accounts of category learning and attentional allocation: A reassessment. Journal of Experimental Psychology: Learning, Memory, and Cognition, 29(6), 1160–1173. https://doi.org/10.1037/0278-7393.29.6.1160