Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Psychological Similarity Spaces and Generalization

Abstract. Psychological Similarity Spaces and Generalization investigates how empirically inferred relations among stimuli, events, and concepts predict the extension of learning to novel cases. The monograph reconstructs the history of multidimensional scaling and generalization research; evaluates metric and nonmetric spaces, Shepard's generalization law, exemplar and prototype models, feature contrast, transformational similarity, Bayesian inference, and error-driven attention learning; and critically examines experimental identification, categorization, perceptual learning, and transfer. It distinguishes stimulus geometry, attentional reweighting, decision operations, and representational change, comparing behavioral experiments with recent naturalistic, neural, and artificial embedding studies. Evidence appraisal addresses identifiability, model equivalence, small and specialized samples, dependent proximity judgments, generalization to new stimulus collections, and the limits of inferring psychological mechanisms from predictive fit. The synthesis specifies independent cross-task testing, model-recovery procedures, and controlled manipulations for future research. This is a scholarly narrative synthesis of published work, not a preregistered systematic review, meta-analysis, or original experiment.

Keywords: psychological similarity spaces; psychological similarity; stimulus generalization; generalization gradients; multidimensional scaling; nonmetric multidimensional scaling; exponential generalization law; selective attention; generalized context model; exemplar theory; prototype models; decision-bound models; ALCOVE; feature contrast; transformational similarity; Bayesian generalization; perceptual learning; representational similarity analysis; artificial cognition; Artificial Era

The Research Object and the Problem of Generalization

A psychological similarity space is an empirically constrained representation of relations among stimuli, events, or concepts, in which proximity is used to account for judgments, confusions, learning, and transfer. Generalization is the application of a response, expectation, classification, or learned relation beyond the exact experiences that established it. Their connection is conditional: spatial proximity can predict transfer only when the represented dimensions, their weighting, and the response process are appropriately specified. This monograph studies that relation through perceptual measurement, multidimensional scaling, learning experiments, computational category models, and contemporary tests with naturalistic material. The research object is the inferred organization linking observed discriminations to prospective behavior, not a literal map presumed to exist independently of task and observer. Shepard (1987) Roads & Love (2024)

The organizing question is how a finite set of observed similarities supports predictions for unobserved cases. A response learned for one color, shape, sound, or object may spread to neighbors in a representational arrangement, yet the spread can be altered by attention, instructions, category membership, and prior knowledge. To describe this scientifically, the investigator must distinguish the stimuli's physical attributes, the discriminability produced by a sensory system, the psychological representation recovered from responses, and the decision rule that converts representation into action. A similar-looking pair can be classified differently because a task emphasizes a diagnostic feature, while physically different exemplars can be treated alike after training. Nosofsky (1986) Goldstone (1994)

Five questions guide the synthesis. What licenses a distance interpretation of measured similarity? Under which dimensional and attentional conditions do exponential generalization gradients appear? How do exemplar, prototype, feature-based, transformation, and Bayesian models make different predictions? Which experimental manipulations can distinguish learning a new representation from changing a criterion for responding? How far do inferences from controlled laboratories extend to natural images, semantic categories, developmental populations, or artificial representations? Answers require attention to both the geometry and the psychological procedure that generated the data. A model's predictive fit is evidence about a specified task and stimulus distribution, while causal claims about internal representations require additional converging tests. Tversky (1977) Ashby & Maddox (1993) Marjieh et al. (2024)

The method is a critical narrative synthesis of original studies, formal proposals, and selective modern reviews, with links to identifiable archival or publisher records. It is organized historically and by model family, and it considers empirical comparisons, conceptual objections, and reproducibility problems. It is not a preregistered systematic review, a meta-analysis, or a new participant experiment. When access is limited to a publisher abstract, conclusions are confined to claims established there; when a full original article or openly published research text is available, the structure of its argument is examined more closely. The resulting account distinguishes original experimental observations, parameters estimated by model fitting, explanatory conjectures, and the author's synthesis of the relationships among those levels. Wu et al. (2025) Roads & Love (2024)

Historical Formation of the Similarity Problem

Psychological comparison became a quantitative research issue before similarity was routinely represented as a numerical space. Investigators could obtain ratings of resemblance, identify which stimuli were confused, and observe whether responses learned for one case transferred to another. These procedures establish proximities: observable relations that may be used to infer an underlying organization. Fred Attneave's 1950 study of dimensions of similarity belongs to the early empirical tradition of extracting structure from such judgments. Its methodological significance lies in turning subjective comparison into data that impose constraints on a representation. A historical account must preserve the distinction between measured response patterns and the geometry later introduced to summarize them. Attneave (1950) Roads & Love (2024)

In classification research, Shepard, Hovland, and Jenkins examined how different assignments of multidimensional stimuli to categories affect learning and memorization. Their 1961 monograph connected the logical arrangement of classifications to patterns of task difficulty and supported the broader program in which generalization depends on the arrangement of examples. This program treated a classification as an experimentally controlled relation between stimuli and responses, making it possible to compare complex category structures while holding the available stimulus features constant. It also established an enduring methodological point: the complexity of a rule cannot be read directly from the number of categories because the discriminations demanded by each boundary matter. Shepard et al. (1961)

Roger Shepard's 1958 experiments explicitly investigated stimulus and response generalization with a model relating transfer to distance in psychological space. The problem was already a joint one of representation and response: distance can specify a propensity for generalization, but a decision or response mechanism must translate that propensity into observations. Later accounts frequently present spatial generalization as a universal principle; the historical record also contains the experimental effort required to estimate or test such a principle. The early work thus provides a link between discrimination, transfer, and quantitative modeling rather than simply an origin story for a famous curve. Shepard (1958) Shepard (1987)

The subsequent development of multidimensional scaling made these questions tractable for richer stimulus sets. The methods of Shepard and Kruskal sought low-dimensional configurations consistent with observed orderings of dissimilarity. This introduced the possibility that distances could be recovered even where measured proximities lacked a trusted interval scale. It also created interpretive obligations: recovery depends on assumptions about the distance family, the monotonic relation to observations, dimensionality, and noise. A recovered coordinate system is a useful scientific hypothesis about comparative structure; its rotation or reflection may carry no psychological interpretation, and its dimensional axes require further experimental grounding. Shepard (1962a) Shepard (1962b) Kruskal (1964)

Multidimensional Scaling as Inference from Behavior

Multidimensional scaling begins with a set of objects and pairwise proximity observations. In a metric formulation, the analyst seeks coordinates whose distances approximate observed dissimilarities after a specified transformation. In a nonmetric formulation, the principal constraint is the rank ordering: pairs rated more dissimilar should tend to be farther apart in the recovered space. Shepard's 1962 papers introduced computational procedures for analyzing proximities when the distance-to-observation function was not known, and Kruskal subsequently developed numerical optimization around goodness of fit. The historical innovation was to make qualitative response order informative about quantitative configuration, while keeping explicit that the fit is relative to a model family. Shepard (1962a) Shepard (1962b) Kruskal (1964)

A distance matrix can be generated by several different configurations, and inferred dimensionality must therefore be defended empirically. A single very flexible high-dimensional embedding can fit observed proximities closely without producing a useful account of new stimuli or transfer. Stress criteria, held-out comparisons, confidence intervals obtained through resampling, and sensitivity to starting configurations help determine whether an apparent structure is stable. The choice between individual-level and pooled data is equally consequential, because group averages may suppress different attention weights or representational geometries. A good-looking two-dimensional plot remains a visualization until predictions are tested on observations excluded from its construction. Kruskal (1964) Roads & Love (2024)

The interpretation of axes is not secured by an optimization routine. If a space is arbitrarily rotated, relative distances are unchanged, yet labels such as brightness, size, emotional valence, or familiarity could attach to different directions. External manipulations can identify meaningful dimensions: vary one physical attribute at a time, compare discrimination performance, or ask whether category training selectively changes proximity along the manipulated dimension. When dimensional properties are correlated in ordinary materials, experimental orthogonalization is particularly important. The relationship between geometry and sensory channels may itself depend on the stimulus integralities that determine whether attention can isolate one dimension from another. Garner & Felfoldy (1970) Nosofsky (1986)

Modern representational similarity analysis applies a related logic to activity patterns from brains and computational models. Rather than requiring a one-to-one correspondence between neurons, voxels, and model units, investigators compare representational dissimilarity matrices across stimulus conditions. Kriegeskorte, Mur, and Bandettini demonstrated this approach in visual object representations, establishing a common quantitative language across neural data, behavior, and theory. The method does not remove the need for a psychologically valid dissimilarity measure: correlation distance, classification confusion, and instructed similarity judgment are different observables. Alignment between matrices also requires cross-validation and attention to dependence among pairwise entries. Kriegeskorte et al. (2008) Roads & Love (2024)

Metric Geometry, Similarity, and Scientific Interpretation

A metric distance is ordinarily nonnegative, symmetric, zero for identical points, and subject to the triangle inequality. Familiar geometric representations add assumptions about continuous dimensions and the manner in which coordinate differences combine. With separable dimensions, a city-block or weighted Manhattan form sums differences along axes; with integral dimensions, Euclidean-type combination is often used as a working approximation. Garner and Felfoldy's sorting experiments made the empirical distinction significant by showing interference or facilitation when stimulus dimensions were manipulated jointly. Those findings connect experimental selective attention to the geometry appropriate for modeling stimulus relations, without turning a laboratory classification into an invariant law for every domain. Garner & Felfoldy (1970) Nosofsky (1986)

Metric assumptions can become scientific liabilities when similarity judgments are asymmetric or strongly context dependent. If people judge A more similar to B than B to A, symmetric point-to-point distance alone cannot reproduce both judgments. Tversky's contrast model addresses common and distinctive features, allowing the weights assigned to features of each object to differ across comparison direction and context. A representation may therefore need directed comparison operations or a richer feature system rather than a revised Euclidean coordinate plot. The question is not whether similarity is globally spatial, but what structure a particular class of data licenses and which deviations reveal a different mechanism. Tversky (1977)

Task context can also produce preferences that frustrate a fixed global proximity ordering. Goldstone, Medin, and Halberstadt reported experiments with intransitivities in within-comparison similarity judgments. Such patterns matter because they can arise when the relevant comparison dimension is recruited dynamically from the alternatives on the screen. A metric space may remain useful as a local baseline while the actual response is generated by context-sensitive feature sampling or selective retrieval. Investigators should test transitivity, symmetry, reliability, and dependence on the comparison set before assuming that numerical similarity ratings directly identify a stable shared metric. Goldstone et al. (1997) Tversky (1977)

Distance has explanatory value only through a model of behavior. An estimated spacing between two stimuli can be related to recognition confusions, category probabilities, reaction times, memory substitutions, or learned-response transfer, and these variables need not share identical response functions. A unified model is substantively stronger when one representation predicts multiple tasks with limited additional fitting. Conversely, allowing a different arbitrary embedding for each dependent measure makes apparent coherence cheap. The defensible claim is therefore predictive invariance under specified transformations and contexts, tested with new data, rather than the mere existence of coordinates chosen to reproduce a training matrix. Nosofsky (1984) Ashby & Maddox (1993) Roads & Love (2024)

Generalization Gradients and the Exponential Proposal

A generalization gradient describes how the probability or strength of a learned response changes as a novel stimulus departs from one already encountered. Its empirical form depends on how distance is measured and which response criterion defines transfer. In his 1987 Science article, Shepard proposed that generalization approximately decreases exponentially with psychological distance across diverse domains, and argued that the appropriate distance metric is shaped by the organization of stimulus dimensions. The proposal is powerful because the same formal relationship can organize recognition, discrimination, and category transfer. It remains a claim subject to task- and representation-specific evaluation, not a mathematical guarantee for any observed similarity measure. Shepard (1987)

The expression g(d) = exp(−c d) conveys the core hypothesis, where d is a nonnegative distance and c controls sensitivity. For an illustrative embedding with two separable weighted dimensions, d may be calculated as a sum of weighted coordinate differences; an integral-dimensional application might instead use a square-root combination. The decay parameter and distance scale are not independently identified without further constraints: multiplying all distances by one factor and dividing c by the same factor preserves predictions. This simple fact matters when comparing fitted gradients across participant groups, tasks, or species. A claimed difference in sensitivity could reflect a difference in scale conventions rather than a psychological process. Shepard (1987) Nosofsky (1986)

The law's explanatory appeal comes from the relation between familiar and new cases. If two new observations are equidistant from a trained stimulus under the selected metric, the law predicts equal response propensity before decision biases or category frequencies are introduced. That produces falsifiable contrasts: construct equidistant novel test items, randomize training examples, and assess whether transfer ranks by fitted distance on a held-out set. A monotonic rank relationship is weaker evidence than a calibrated exponential function; it can be shared by several decreasing functions. Model tests therefore need to distinguish the distance-order claim, the functional-form claim, and the invariance claim rather than treating them as a single success. Shepard (1958) Shepard (1987)

The claim of generality has been revisited with richer materials. Marjieh, Jacoby, Peterson, and Griffiths tested the universal-law proposal for naturalistic stimuli, addressing the earlier concentration of evidence on simplified low-dimensional sets. Their 2024 report supplies contemporary empirical support for an appropriately construed generalization-distance relationship while underscoring the need to infer psychological representation from human behavior. Modern experiments make clear that ordinary objects and sounds create difficult correspondence problems: perceptually available features, semantics, and familiarity all influence which geometry is recovered. Robustness should be assessed over stimulus domains and comparison protocols, not inferred from a single smooth curve. Marjieh et al. (2024) Roads & Love (2024)

Exemplar Similarity and the Generalized Context Model

Exemplar-based theories predict classification from the relation between a test item and remembered instances. Medin and Schaffer's 1978 context theory explained classification performance with retrieval of information associated with similar stored examples. Their four experiments used geometric forms and schematic faces and compared exemplar-based predictions with alternatives using additive cue information. The proposal shifted explanatory emphasis from a distilled category summary to the cumulative effects of particular learning episodes. A stored exemplar in a mathematical model is an assumed functional representation, so direct claims about episodic memory implementation require evidence from independent memory measures, delay manipulation, or neural tests. Medin & Schaffer (1978)

Nosofsky's 1984 account placed context theory in relation to Luce-type choice principles and scaling results, introducing the view that similarity parameters can systematically vary with the demands of a classification task. The resulting generalized context model treats a probe as activating stored exemplars in proportion to similarity, with category evidence formed by pooling activation for the members assigned to each response. A commonly used prediction is that the probability of category A is the sum of similarities to A exemplars divided by the corresponding sums across candidate categories. This expression implies competition between responses and is contingent on choices about exemplar frequencies, sensitivity, and decisional scaling. Nosofsky (1984)

The model's most consequential departure from a fixed perceptual map is attention weighting. Nosofsky's 1986 identification and categorization study used the same confusable multidimensional stimuli in both tasks. Reported quantitative analyses showed that a representation constructed from identification could be used as a basis for categorization when selective attention was incorporated. Because the published study's central demonstration involved two intensively studied participants, its strength is fine-grained model comparison within controlled stimulus conditions, with a narrow participant sampling basis. Generality to heterogeneous populations and unrestricted materials requires separate experiments. This distinction between within-person explanatory precision and population-level reproducibility is essential to evaluating influential computational demonstrations. Nosofsky (1986)

Exemplar models offer different predictions from a category centroid when the category contains separated clusters, unusual exceptions, or regions of unequal training density. Close exemplars may exert disproportionate influence on a novel query even where its distance from the mean of all category members is large. Yet mathematical flexibility also creates risk: high numbers of remembered instances and adjustable dimension weights can fit historical responses without uniquely identifying the memory representation. Researchers should report the actual exemplar set, frequency treatment, embedding source, tuning procedure, and held-out accuracy, and compare against suitably flexible alternatives. A model that predicts transfer after learning reveals more than one that retrospectively explains the observations used to fit every parameter. Medin & Schaffer (1978) Ashby & Maddox (1993)

Selective Attention and Contextual Reweighting

Similarity depends on which dimensions enter an active comparison and how much weight each receives. For a diagnostic classification task, an observer may emphasize one subtle feature and downweight an otherwise salient but irrelevant difference. Nosofsky's 1987 study of integral stimuli explicitly tested identification and categorization learning using an exemplar generalization framework, predicting performance across categorization conditions on the basis of identification data. The research supports the value of treating attention as a process that alters effective similarity, while the perceptual integralities determine how freely weights can change. Nosofsky (1987) Garner & Felfoldy (1970)

Attention weighting provides a mechanistic bridge from experimentally manipulated contingencies to changing distance. Suppose two stimuli differ in color and size but only color predicts category feedback. An attention model predicts increasing effective separation for changes on the diagnostic color dimension and reduced relevance for size differences. This is a formal directional prediction that can be tested on held-out combinations of attributes. It does not itself establish whether sensory encodings changed, whether executive control shifted sampling at the moment of choice, or whether feedback altered memory traces. Each interpretation calls for an additional design such as speeded discrimination, dual-task interference, eye tracking, or delayed categorization. Kruschke (1992) Goldstone (1994)

Observed context dependence creates a methodological challenge for multidimensional scaling. If a single distance matrix is inferred from ratings under one instruction, that matrix may fail under a new classification regime. Researchers can fit shared baseline coordinates with context-specific weights, or allow embeddings to vary, and then use cross-condition transfer to assess whether the less flexible shared-representation hypothesis survives. The crucial comparison is predictive: estimate the geometry in one condition, hold its parameters fixed where theoretically required, and test performance in another. A free geometry in every condition can capture context changes while concealing the processes that generated them. Nosofsky (1984) Roads & Love (2024)

The perceptual conditions for independent attention are themselves empirical variables. Garner and Felfoldy's classic speeded-sorting tasks distinguished patterns of interference and facilitation across integral versus separable dimensions; their study relates those effects to alternative metrics in scaling. Such findings warn against treating weights on all stimulus dimensions as unconstrained sliders. Training may change attention to chroma independently from geometric size in one apparatus, while a different stimulus family resists selective focusing. Formal similarity functions should therefore represent realistic attentional constraints and report how these constraints were independently measured. Garner & Felfoldy (1970) Nosofsky (1986)

Perceptual Learning and Category-Induced Restructuring

A learning history can change later perceptual discriminations. Goldstone's 1994 experiments trained participants on alternative category regimes, then tested same–different discrimination. His analyses reported acquired distinctiveness for distinctions that category learning had made behaviorally important and a more limited instance of acquired equivalence. The study is a key source because a changed post-training discrimination measure suggests that category learning can affect perceptual sensitivity, rather than only which response label is chosen. Even so, the same–different task contains judgmental components; isolating sensory and decisional contributions is a continuing experimental task. Goldstone (1994)

Goldstone's later review distinguished attention weighting, imprinting, differentiation, and unitization as mechanisms through which perceptual learning might develop. These proposals imply different changes in effective geometry. Weighting can stretch an already represented dimension; differentiation creates separations for previously confusable stimuli; unitization may make a complex configuration act as a functional single feature. Such mechanisms need not produce the same error patterns or transfer curves. Experiments should identify the training contingencies, pre/post sensitivity measures, generalization to unseen items, and persistence beyond a single testing session. Goldstone (1998)

Representation change and decision change are conceptually separate and empirically difficult to discriminate. Under a decision-only explanation, participants may retain the same sensory encoding while shifting the boundary used to classify uncertain observations. Under a representational-change explanation, stimulus positions or distances change before the decision rule is applied. Comparing category judgments alone often leaves these explanations confounded. Cross-task measurements, especially recognition confusions and speeded discrimination obtained before and after training, provide additional constraints. Model recovery studies should determine whether fitting procedures can actually distinguish the two architectures on datasets resembling the collected samples. Ashby & Maddox (1993) Maddox & Ashby (1993)

The practical consequence is that psychological space can be an outcome of learning as well as an input to it. A training sequence changes which distinctions will be informative in future behavior; the observer may subsequently attend more reliably to dimensions or combinations that support success. The explanatory cycle must nevertheless be specified without circularity. If a researcher infers changed geometry solely because category responses changed, and then explains those responses by changed geometry, little has been learned. An independent transfer set or cross-modal discrimination task is needed to make the causal claim informative. The most useful research designs combine formal distance predictions with measurements that vary in their sensitivity to representation and decision operations. Goldstone (1994) Nosofsky (1987)

Feature Matching and Directional Similarity

Spatial distances formalize how much two stimulus representations differ along dimensions. A feature-matching theory asks which discrete or conceptual properties the two representations share and which belong uniquely to either member of the comparison. Tversky's 1977 contrast model expresses similarity through a weighted balance of shared features and distinctive features. Its distinctive explanatory contribution is directional comparison: if one object is a salient or complex reference, its unique features may be weighted differently from the unique features of a less prominent comparison object. The theory therefore predicts systematic asymmetries and changes in judgment when comparison roles or contexts are reversed. Tversky (1977)

Feature contrast is especially relevant when the compared items are natural concepts rather than controlled patches of color or geometric shapes. A participant may compare two vehicles by shared purpose, engine type, size, or social association, and the diagnostically relevant set will depend on the question being asked. A latent geometric embedding can reproduce some feature differences but may need many dimensions or contextual parameters to represent directional judgments. Feature-based formulations make the retrieval and relevance of features part of the explanation. They face their own measurement problem: the experimenter must specify how features are discovered, how commonality is scored, and whether the selection process is constrained before fitting the judgments. Tversky (1977) Rips (1989)

The asymmetry test illustrates a general principle of model discrimination. A symmetric metric alone predicts equal distance from A to B and from B to A; contrast models allow directional weights, although additional assumptions are needed to predict response magnitudes. Researchers can randomize which item appears as the standard, manipulate category typicality, and estimate individual-level directional effects. Repeated comparisons test whether an asymmetry exceeds measurement noise or position bias. Whenever a contextual feature account is proposed, feature ratings or verbal descriptions collected independently can constrain the analysis. A model comparison that simply grants the feature theory any set of free weights offers little evidence about its psychological mechanism. Tversky (1977) Goldstone et al. (1997)

Feature accounts and spatial models can coexist as different levels of description when each is linked to a specified task. A distance representation may summarize discriminability among finely graded perceptual stimuli, while feature retrieval may better predict verbal comparisons among semantic concepts. There is no requirement that the same representational format govern all uses of similarity. Conversely, multiple formats may be invoked within one task, as when an observer first identifies an object through perceptual cues and then evaluates its functional resemblance to another object. The research program is to state when each account predicts a divergence and to test those divergences on stimuli outside the fitting set. Roads & Love (2024) Hahn & Chater (1998)

Similarity as Transformation and Structural Alignment

A different tradition characterizes similarity by the operations necessary to transform one representation into another. Hahn, Chater, and Richardson proposed that readily transformable representations are judged more similar than representations requiring more transformations. Their 2003 report described three experiments and argued that transformation distance can explain effects difficult for simple feature or spatial models. The proposal changes the unit of analysis: similarity depends on a sequence of permitted operations and the cost assigned to those operations. For patterned images, transformations may involve translation, rotation, insertion, substitution, or reorganization; the permitted operator set is itself a substantive psychological assumption. Hahn et al. (2003)

Transformation models become important when relational configuration matters more than a list of separate properties. Two drawings can share several local components while differing in an arrangement that is costly to reorganize; alternatively, they may share few pixels yet be related by a simple rotation. A geometric embedding can be constructed from transformation costs, but it may obscure the structure of the operations that gives the distances psychological meaning. A direct experiment varies the number or complexity of transformations independently from crude feature overlap, then tests whether judgments and reaction times track the predicted operation cost. Hahn et al. (2003) Roads & Love (2024)

The transformation approach is also vulnerable to underdetermination. If the analyst is free to invent any operation and assign any cost after observing the data, the model can reproduce almost any ordering. A credible test defines the operator vocabulary before fitting and uses tasks that probe whether the predicted transformations are psychologically accessible. Transfer to novel exemplars, changes under time pressure, and errors involving partial transformations can provide additional evidence. Its principal scientific value is to show that the organization of resemblance may depend on representational computation, not only on static pairs of locations. Hahn et al. (2003) Hahn & Chater (1998)

Structural alignment enters the question whenever common relations among components guide comparison. Transformational and alignment accounts should be separated from purely low-level matching even if all three can be implemented computationally. A relational description of two objects supports generalization across differences in physical surface, provided the relation itself is identified by the observer. This raises empirical questions about learning the relevant relation, cognitive resource demands, and instruction sensitivity. A similarity study confined to surface ratings cannot by itself determine whether a participant encoded a relation or a collection of visual fragments. Cross-task tests with novel material and constrained transformation models are required. Hahn et al. (2003) Lake et al. (2015)

Prototype, Exemplar, and Decision-Bound Accounts

A prototype model relates a novel item to an abstraction, such as a category center or central tendency. An exemplar model retains multiple training cases and accumulates their individual similarity contributions. A decision-bound model represents how observations are partitioned into response regions. These three descriptions answer partly different questions about memory representation, similarity integration, and response selection, so direct comparisons must align their observable predictions. Ashby and Maddox demonstrated formal equivalences among selected models under restrictive assumptions, showing that different representational stories can yield the same categorization boundaries. Distinguishing them requires evidence beyond one set of response proportions. Ashby & Maddox (1993)

Experimental category structures offer targeted contrasts. A coherent cluster supports a compact prototype, while a category with isolated exceptions or irregular density may reward exemplar retrieval. Decision-bound theories can closely approximate either pattern when boundaries are sufficiently flexible. The discriminating test therefore samples items near disagreement regions, including novel stimuli whose relation to the centroid differs from their relation to individual training cases. Learning order, category prevalence, delay, and memory-load manipulations also affect model predictions. Comparison must consider how many free parameters each model uses, whether noise is modeled consistently, and whether parameter estimation recovers stable solutions. Medin & Schaffer (1978) Maddox & Ashby (1993)

Maddox and Ashby's 1993 comparison of decision-bound and exemplar accounts illustrates the value of embedding behavioral classification in explicit perceptual and decisional processes. The models can differ in assumptions about perceptual noise and the mapping from evidence to choices. This matters because a lower error rate in one categorization model does not establish that the brain stores individual exemplars in the exact form specified by its equations. Model adequacy involves both predictive accuracy and the identification of mechanism. Cross-validation can adjudicate prediction, whereas experimental perturbation and independent representation measures are needed for stronger mechanistic conclusions. Maddox & Ashby (1993) Ashby & Maddox (1993)

The relation between rules and similarity has been a recurrent source of conceptual confusion. Hahn and Chater argued that it is necessary to distinguish rule-based and similarity-based procedures by the way stored and novel representations are matched. Their analysis shows why experimental contrasts relying on intuitive labels can be inconclusive: a rule can be implemented through graded evidence, and a similarity system can approximate regular boundaries. Strong research identifies the representation, matching procedure, and response mechanism in each competitor. The outcome may be a constrained hybrid rather than a verdict assigning all observed performance to one exclusive faculty. Hahn & Chater (1998) Wu et al. (2025)

Bayesian Generalization and Structured Hypothesis Spaces

Bayesian accounts treat generalization as inference about the class of possibilities capable of producing the observed examples. Tenenbaum and Griffiths revisited Shepard's generalization law in 2001, explaining how assumptions about natural kinds and hypotheses can yield graded extensions beyond a learned case. Their framework is important because it shows that an exponential or related distance-based gradient need not prove a dedicated, purely metric similarity mechanism. A gradient can emerge from uncertainty about the extension of a latent kind. The probability assigned to a novel item depends on prior assumptions, observations, and the structure of the candidate hypotheses. Tenenbaum & Griffiths (2001)

The shape of a generalization curve can therefore be diagnostic of the learner's assumed generative structure. Under one hypothesis space, nearby examples may be likely to share a category; under another, categorical boundaries or higher-order relations may dominate. When several training items are provided, their pattern and dispersion can update beliefs about which hypotheses remain plausible. This predicts effects of sample size and example distribution that may not be described by a simple sum of fixed similarities to individual exemplars. The research challenge is to constrain priors and hypotheses independently enough to support prediction rather than post hoc rationalization. Tenenbaum & Griffiths (2001) Wu et al. (2025)

Bayesian and similarity models can be compared on withheld generalization trials after manipulating whether examples appear to be sampled deliberately or opportunistically. Inference about selection processes may affect which regions are considered representative and how strongly a small set of observations constrains a concept. The key empirical distinction is between the proximity of one example and the inferred extension of the rule or class from which examples arise. If two test stimuli have equal measured distance from the nearest exemplar, but differ in probability under a learned structural hypothesis, a Bayesian model can make a distinct forecast. Any reported advantage must be evaluated against model complexity and the possibility that an enriched similarity model captures the same pattern. Tenenbaum & Griffiths (2001) Ashby & Maddox (1993)

Broader work on generalization increasingly integrates conceptual rules, latent structures, and graded prior experience. Wu, Meder, and Schulz's 2025 review organizes developments from concept learning through function learning and reinforcement learning, emphasizing hybrid accounts that combine mechanisms previously regarded as alternatives. The implication for psychological space research is methodological: distance can be a component of inference while the agent also tracks structure, goals, and sampling assumptions. Models should specify which predictions are produced by the metric component and which depend on learned hypotheses. This decomposition improves explanation even when several computational architectures generate similar behavioral functions. Wu et al. (2025) Roads & Love (2024)

ALCOVE and Error-Driven Attention Learning

Kruschke's ALCOVE model joined exemplar-based similarity with associative learning and learned selective attention. Its hidden exemplar units respond to test stimuli according to distance in a multidimensional psychological space, while connections to category responses and dimension weights change under prediction error. The 1992 paper presented the model as a means of accounting for category-learning phenomena that challenged simpler independent-cue approaches. The proposal is mechanistically explicit in an important sense: it specifies how repeated corrective feedback can alter which features count as similar, rather than treating the weights inferred at the end of learning as unexplained constants. Kruschke (1992)

Within ALCOVE, a change in attention reshapes the effective distance metric, and associative weights regulate how activation from stored exemplars supports particular responses. These operations interact. A learner can improve without acquiring an entirely different perceptual representation if task-relevant dimensions acquire more weight and connections to category labels become more accurate. The model therefore predicts learning trajectories in addition to terminal classification patterns. Time-resolved evaluation asks whether predicted errors decline at the right pace, which items remain difficult, and when attention shifts after category contingencies change. Learning curves can help discriminate models even when their final accuracy estimates resemble one another. Kruschke (1992) Nosofsky et al. (1992)

Nosofsky, Kruschke, and McKinley later examined how exemplar category representations could be combined with connectionist learning rules. Such work demonstrates the productive distinction between representational assumptions and algorithms that modify them. An exemplar model can be paired with different learning processes, while one error-driven procedure can operate over alternative feature systems. A model comparison should therefore specify which part is being tested: the storage of individual instances, the similarity function, the feedback-learning equation, or the transformation from evidence to overt response. Otherwise success attributed to exemplars may arise from flexible learning dynamics or vice versa. Nosofsky et al. (1992) Kruschke (1992)

To evaluate the attention mechanism causally, one can train two groups on different diagnostic dimensions while keeping physical stimulus distances constant, then obtain common generalization probes before additional feedback. ALCOVE predicts systematic changes in the effective influence of each feature. A stronger study also measures attentional allocation independently with controlled discrimination or eye movements, compares parameter-recovery performance, and reports individual differences rather than only group means. Such a design can ask whether learned weights predict transfer to wholly new objects, a more demanding test than classifying the training examples. The predicted contrasts arise from the formal learning mechanism and should be registered before observing the held-out responses. Kruschke (1992) Goldstone (1994)

Experimental Tasks and Behavioral Measurements

Similarity studies draw on several observation channels, each answering a different question. Direct ratings and pairwise judgments record explicit comparative decisions; identification confusion matrices reveal how often one stimulus is mistaken for another; categorization tasks measure responses after learning category–stimulus contingencies; generalization tests observe whether a learned response transfers to untrained cases. These measurements are related because each may depend on an underlying representation, but they differ in decision demands, memory requirements, and instructions. A complete experiment should report precisely which channel defines its similarity evidence and whether other channels provide independent validation. Attneave (1950) Nosofsky (1986)

A pairwise comparison design must decide which pairs are shown, in which order, how many times, and whether the displayed reference can influence response direction. If the investigator obtains all possible pairs from many stimuli, the number of comparisons grows rapidly, creating fatigue and dependence across responses. Adaptive sampling can reduce that cost but must be documented because it changes which regions of stimulus space are estimated precisely. Repeat judgments and reversed presentation order make it possible to separate reliable asymmetry from random noise or order effects. Scaling results derived from pooled ratings should be compared with individual-level configurations where heterogeneity is a scientific concern. Tversky (1977) Kruskal (1964)

Categorization paradigms require explicit separation of training and test data. During training, repeated exemplars may be reinforced or labeled; during test, novel exemplars can be presented without feedback so that extension beyond acquired instances is observed. A balanced probe set should sample distances close to and far from trained cases, boundary regions where models disagree, and contrasts along both relevant and irrelevant dimensions. Response accuracy alone may conceal the geometry of generalization, so the distribution of category probabilities and reaction times is often informative. Counterbalancing stimulus–label mappings helps exclude a fixed preference for one response key or label. Medin & Schaffer (1978) Maddox & Ashby (1993)

Reliability and validity require separate scrutiny. A rating can be internally repeatable but reflect verbal association rather than discrimination; a perceptual confusion score can be sensitive to exposure duration or visual acuity; a learned-response gradient can mix generalization with response inhibition. Investigators should preserve trial-level data, describe exclusion criteria, quantify uncertainty, and test whether findings persist when the observation channel changes. Especially for neuroimaging or machine-learning comparisons, apparent predictive performance can depend on arbitrary normalization choices. The most compelling representation claims survive independent tasks with clearly specified measurement models and an explicitly controlled stimulus-generating procedure. Roads & Love (2024) Kriegeskorte et al. (2008)

Model Recovery, Cross-Validation, and Robustness

A model's capacity to match a dataset and its capacity to identify a psychological process are different achievements. A sufficiently elaborate exemplar scheme may fit a collection of classification responses, but a prototype or decision-bound model can sometimes generate almost identical response distributions. Ashby and Maddox's mathematical comparisons show why competing mechanisms can be observationally equivalent under certain parameterizations. Before interpreting parameters as attention or memory operations, researchers should simulate data from the proposed alternatives and test whether the planned estimation procedure recovers the generating model. This procedure diagnoses identifiability rather than merely optimizing the fit of a favored theory. Ashby & Maddox (1993) Maddox & Ashby (1993)

Parameter recovery tests whether a fitting method can retrieve known values from simulated observations generated with those values. Model recovery asks whether a comparison rule can identify which competing model generated a synthetic dataset under plausible trial numbers and noise. If recovery succeeds only with unrealistically extensive observations, the design's eventual causal interpretation will remain weak. Recovery can also reveal parameter trade-offs: category sensitivity, stimulus scale, attention weights, and response-noise assumptions may compensate for one another. The analyst should document such trade-offs and use experimental manipulations to separate them where possible. Nosofsky (1984) Kruschke (1992)

Cross-validation evaluates whether a model generalizes to unobserved trials, participants, or stimulus families, but these validation targets are not interchangeable. Randomly withholding repetitions of familiar exemplars estimates a narrower form of prediction than withholding entire novel stimulus types. For similarity theories, out-of-stimulus and out-of-task validation can be especially consequential, since fitted embeddings may be adapted tightly to training comparisons. Reporting predictive uncertainty and scoring the full probability distribution, rather than only the most probable category, makes it possible to assess calibration and anomalous stimulus regions. A strong result should remain interpretable when reasonable alternative distance metrics and preprocessing choices are considered. Roads & Love (2024) Marjieh et al. (2024)

Robustness also concerns replicability across laboratories and participant populations. A stable aggregate gradient may conceal subgroups with different learned metrics, and a model can appear successful because a particular set of stimuli happen to favor its assumptions. Independent replication should vary stimuli, task instructions, available dimensionality, and sampling contexts while preserving the theoretical prediction to be tested. Model complexity, optimization restarts, data exclusions, software version, and analytic choices belong in the audit trail. The aim is to establish which relation between distance and generalization persists when incidental details change and which components legitimately depend on them. Shepard (1987) Wu et al. (2025)

Generalization Across Dimensional and Category Structures

The geometry used in a generalization model can be varied experimentally by constructing stimuli with independent dimensions or correlated attribute structures. Shepard, Hovland, and Jenkins's classification-learning program made formal structure an experimental variable: categories could be assigned to a controlled multidimensional set in different ways. Modern studies can extend this logic by testing whether a model learned under one category arrangement correctly predicts transfer after a reversal, a new boundary, or a change in feature reliability. The classic Type II advantage was subjected to a major reassessment by Kurtz and colleagues in eight experiments published in 2013; they found that the advantage depended on particular experimental conditions, with outcomes also showing reversals and nondifferences. This result narrows the validity of treating one learning-difficulty ordering as a universal benchmark for a generalization theory. Such designs identify the boundaries of the representation and learning rule while controlling exposure and stimulus frequencies. Shepard et al. (1961) Kurtz et al. (2013) Kruschke (1992)

Integral and separable dimensions are a particularly instructive case. Under separability, weighting features independently can be psychologically plausible; under integrality, the perceptual experience may resist such decomposition, or attention to one attribute may change discrimination of another. Garner and Felfoldy's experiments connected the interference pattern to the way dimensions combine in processing. A model that treats independent weights as freely adjustable in an integral domain may fit observations but overlook the limits of selective attention. Experimental tests can compare combinations of independent and correlated cues and inspect whether predictions extend beyond the sampled training range. Garner & Felfoldy (1970) Nosofsky (1987)

Sparse exceptions, dispersed clusters, and multi-modal category distributions create contrasting demands on prototypes and exemplars. A nearest-neighbor account may excel on irregular local cases, while a category-center account may provide economical predictions in compact distributions. Yet the distinction weakens when prototypes are extended to mixtures or when exemplar kernels become broad enough to approximate a smooth central tendency. A meaningful comparison must predefine the rival model classes and reserve probes where they make distinct numerical predictions. Interpretation then depends on the predicted pattern of errors and confidence, not a loose preference for a named family of representations. Medin & Schaffer (1978) Ashby & Maddox (1993)

Generalization can also cross changes in task goals. A learned perceptual distance may transfer from identification to classification if both tasks rely on a sufficiently shared representation, but decision criteria and attention weights will differ. Nosofsky's work used identification–categorization relations to investigate this problem. The extension to recognition memory, semantic judgments, and motor responses calls for similar cross-task tests. Every transfer experiment should identify what is held fixed—stimulus coding, category labels, feedback, or response mapping—and which change would constitute a meaningful failure of the proposed invariance. Nosofsky (1986) Nosofsky (1987)

Naturalistic Stimuli and the Generalization Law

Natural stimuli offer a more demanding test of geometric generalization because objects carry correlated visual, auditory, functional, and semantic information. The controlled color patches and schematic faces used in many early studies simplified distance estimation; a photograph of a bird or an audio recording contains many possible dimensions, only some of which govern a particular response. Marjieh and colleagues' 2024 study revisited Shepard's proposed law using naturalistic materials and behavioral similarity evidence. Its significance lies in testing whether an exponential relation can remain useful when psychological space must be inferred from far richer observations. Marjieh et al. (2024) Shepard (1987)

To interpret naturalistic results, investigators should specify the process by which distance is estimated. Features generated by a pretrained image network, ratings obtained from human participants, and multidimensional coordinates inferred from identification errors may disagree. When a curve is fitted using the same ratings that established the space, an apparent match to the generalization law does not demonstrate transfer to independent behavior. A stronger analysis obtains a candidate representation from one task and then predicts a withheld generalization decision under conditions that preserve a clear operational criterion for learning. This separation tests the relevance of the inferred space rather than its capacity to summarize a single matrix. Roads & Love (2024) Marjieh et al. (2024)

Naturalistic examples also make category knowledge salient. Two birds may be perceptually close but belong to categories that an expert distinguishes immediately; two tools may look different while sharing a functional identity that guides generalization. The source of similarity can therefore move with expertise, instructions, or prior learning. A hierarchical design that recruits both novices and experts, measures familiarity, and samples perceptual and semantic attributes independently can test where representational structures diverge. The theory should identify in advance whether it predicts a common geometry across domains or task-specific changes in the dimension weights and comparison rules. Tversky (1977) Wu et al. (2025)

Sampling the natural world creates another source of apparent universality. Stimulus collections often reflect convenience, commercial image databases, or cultural availability, which can make their category boundaries and attribute frequencies highly nonrepresentative. The estimated gradient may describe that collection well without extending to unfamiliar habitats, languages, or visual styles. Researchers should document provenance and rights for the stimuli, split training and test sets by meaningful families where possible, and evaluate whether results hold across independently assembled collections. This produces a more informative scientific statement than treating diverse photographs in one database as a sufficient sample of natural experience. Roads & Love (2024) Marjieh et al. (2024)

Representational Similarity in Neuroscience

Representational similarity analysis relates neural responses to psychological hypotheses by comparing dissimilarity patterns across a set of conditions. Kriegeskorte, Mur, and Bandettini demonstrated how this method makes it possible to compare human and animal neurophysiology, functional imaging, and computational model representations without assuming that corresponding recording channels occupy identical coordinates. For a similarity and generalization study, the central promise is to ask whether the dissimilarity structure recovered from behavior resembles that recovered from neural activity, and whether both predict transfer to new cases. Agreement is meaningful when the measures and analysis pipeline are specified independently enough to avoid circular comparisons. Kriegeskorte et al. (2008)

Neural dissimilarity can be calculated using correlations, classification-based measures, or cross-validated estimators, each with different sensitivity to noise and signal scale. A neural pattern may reflect attention, familiarity, or decision preparation as well as stimulus encoding. Consequently, a high correlation between a behavioral and neural matrix does not identify the direction of causation or prove that the same distance metric drives choices. Time-resolved measurements, task manipulations, and independent estimates of decision activity can separate some alternatives. The strongest evidence would show that a neural geometry predicted from one task constrains behavioral generalization in another, with uncertainty propagated through both estimated spaces. Kriegeskorte et al. (2008) Roads & Love (2024)

A neural embedding also creates a dimensionality problem. High-dimensional patterns may contain multiple representational axes, some related to experimental stimulus categories and others to contextual state or measurement artifacts. Reducing the pattern to a two-dimensional visualization can aid inspection while removing information important for generalization. Interpretations of neural axes should therefore be based on explicit model predictions, perturbations, or independent covariates rather than visual proximity in a projection. Cross-validation and stimulus-level generalization tests can reveal whether an apparently coherent space predicts responses outside the observations used to select a dimensionality or analysis metric. Kriegeskorte et al. (2008)

Representational analysis is most informative when treated as one element of convergent evidence. A behavioral classification probability, an explicit similarity rating, and a neural pattern each contain different response and noise components. If all three exhibit a related structure after training, researchers can ask which transformations are common and which are unique. Such observations constrain mechanisms but leave multiple computational implementations possible, a point made vivid by the formal equivalences among categorization models. Causal manipulation or sufficiently discriminating tests are needed before declaring that a particular neural population instantiates the exact exemplar or prototype algorithm favored by one behavioral fit. Ashby & Maddox (1993) Kriegeskorte et al. (2008)

Computational Embeddings and Artificial Cognition

Contemporary computational models provide numerical representations of images, sounds, language, and actions, often enabling distances to be computed before a psychological experiment begins. These embeddings can serve as hypotheses about human generalization, as tools for selecting discriminating stimulus pairs, or as objects of comparison in their own right. Their scientific status depends on the target claim: predictive agreement between an embedding and a human judgment matrix is evidence of some shared relational structure under that benchmark, whereas equality of mechanisms, meanings, or experience would require different kinds of evidence. The distinction is especially important in the Artificial Era, where artificial systems become active participants in learning and comparison tasks. Roads & Love (2024) Lake et al. (2015)

One rigorous experimental design obtains machine-model distances without fitting them to the behavioral test subjects, estimates a relation to human similarity in one stimulus subset, and evaluates predictive accuracy on new stimulus families. A second design uses model disagreement to select informative tests: choose stimuli judged close by the machine and far by people, then ask which account predicts subsequent categorization or recognition errors. Such comparisons should include simpler pixel-, feature-, and category-based baselines so that any machine-model advantage can be interpreted. An opaque representation's high predictive accuracy is scientifically useful but can conceal what perceptual or conceptual dimensions produce its responses. Roads & Love (2024) Kriegeskorte et al. (2008)

Generative program-induction models demonstrate another approach to generalization. Lake, Salakhutdinov, and Tenenbaum's 2015 study learned novel handwritten character concepts from sparse examples by inferring structured programs for their production. The model's relevance is theoretical: an observer may generalize because a new item is consistent with an inferred generative procedure, even when raw appearance-based distances are an incomplete description. Such models supply a way to examine structural knowledge alongside graded similarity. Their reported success on a specific character-recognition domain should not be generalized automatically to all perception or to human-level understanding across unrelated tasks. Lake et al. (2015)

The comparison of human and artificial spaces should be anchored in observable tasks, model versions, training provenance, and reproducible evaluation. Artificial embeddings can change after fine-tuning, context updates, or architectural replacement; human behavioral similarities can shift with expertise or instruction. Measuring both before and after matched learning interventions can reveal whether they change in comparable ways. This is a research program about identifiable functional relations across biological and artificial systems. The broad philosophical horizon of Artificial Era motivates questions of comparative cognition while empirical claims remain defined by the observations and interventions actually performed. Lake et al. (2015) Roads & Love (2024)

Individual Differences, Development, and Cultural Generality

A population-average psychological space can conceal substantial differences in the features that individual observers attend to. Observers with different expertise may use alternative diagnostic dimensions and therefore show different generalization gradients on the same physical items. The mathematical models already reviewed permit participant-specific sensitivity and attention weights, but interpreting those parameters requires enough repeated observations per participant. A group mean may fit a simple smooth function even when individuals exhibit several distinct profiles. Hierarchical models or carefully stratified individual analyses can therefore improve both descriptive validity and future predictions. Nosofsky (1986) Roads & Love (2024)

Developmental generality is an empirical issue rather than an automatic extension of adult laboratory results. A child and an adult may differ in vocabulary, relevant prior exposure, attentional control, task comprehension, and category knowledge. A comparison based entirely on verbal ratings could therefore attribute to representation a difference actually arising from response demands. Developmental tests should use age-appropriate tasks, examine measurement equivalence, and identify what is held constant across groups. Without such evidence, a law fitted to adult decisions cannot establish identical psychological distances or the same learning mechanism throughout the lifespan. The same caution applies to clinical populations when sensory capacities or task demands differ. Shepard (1987) Wu et al. (2025)

Cultural and linguistic experiences can alter which features are salient in everyday objects or conceptual categories. Naturalistic stimulus studies must therefore report where the materials came from, who encountered them, and whether participants recognized their intended function or labels. Generalization may rely on common sensory constraints while higher-level spaces reflect learned categories and social practices; this is a hypothesis inviting matched, independent testing across settings. Adequate studies should translate instructions carefully, obtain culturally meaningful examples, and distinguish differences in familiarity from differences in representational organization. The absence of a cross-cultural test is a limit on inference rather than proof of either similarity or divergence. Tversky (1977) Marjieh et al. (2024)

These questions converge on transportability. A model's fitted attention weights, gradient sensitivity, and coordinate axes may predict well for participants drawn from the original setting, yet require re-estimation for other settings. To test transport, researchers should preregister a target population or stimulus domain, fix the transferable parts of the model, and quantify prediction degradation. The result can reveal whether a common structural principle survives after differences in content and experience are acknowledged. Such an analysis is a direct test of the scientific claim that psychologically measured similarity supports generalization beyond its original experiments. Roads & Love (2024) Wu et al. (2025)

Evidence Appraisal and Sources of Bias

The evidence base for similarity spaces includes laboratory experiments, quantitative model fits, methodological theory, explicit judgments, and neural or artificial representational comparisons. These sources support different conclusions. Controlled categorization learning establishes effects of specified examples and feedback on observed choices; an estimated attention parameter provides a model-dependent summary of those effects; a computational embedding expresses a representation chosen by an analyst or learned by a machine. Confusing these levels makes a theory seem better established than its underlying observations permit. The strongest syntheses evaluate original results alongside the constraints and alternatives of the methods used to produce them. Medin & Schaffer (1978) Nosofsky (1986) Roads & Love (2024)

Several classic demonstrations have unusually strong within-task control and limited participant sampling. Nosofsky's 1986 quantitative experiments on identification and categorization illustrate the scientific value of intensive individual measurements, while also requiring caution before making population-wide claims. Other studies draw on multiple experiments but narrow families of schematic materials. The relevant assessment asks what the design can establish: precise patterns of learning under controlled conditions do not automatically determine the similarity structures used across real-world settings, populations, and tasks. Broader naturalistic studies are essential, but those studies introduce more stimulus confounds and ambiguity in the measured construct. Nosofsky (1986) Medin & Schaffer (1978) Marjieh et al. (2024)

Quantitative evidence often arrives through goodness-of-fit comparisons rather than direct measurement of a latent internal geometry. Likelihood differences and predictive scores can inform choice among prespecified models, yet researcher flexibility over metric, dimensionality, feature weights, and response rule can generate spurious confidence. Because pairwise observations share stimuli and participants, standard errors also require methods suited to correlated data. Reporting held-out predictions, individual variability, uncertainty intervals, and parameter or model recovery offers a more defensible basis for inference. Where only abstracts have been examined, the manuscript refrains from claiming a detailed risk-of-bias audit of inaccessible original analyses. Ashby & Maddox (1993) Kruskal (1964) Roads & Love (2024)

Evidence of perceptual reorganization deserves particular scrutiny because a learning manipulation can alter stimulus coding, attention allocation, response criteria, and memory retrieval in combination. Goldstone's category-learning experiments provide a substantive demonstration of changed post-training discrimination, but separating the locus of that change calls for cross-task controls and temporal measures. The same principle extends to neural and computational comparisons: similarity between two distance matrices may originate from shared task structure while their generating mechanisms differ. Critical appraisal therefore weighs independent convergence more heavily than a single highly optimized fit and treats incompatible observations as opportunities for discriminating experiments. Goldstone (1994) Kriegeskorte et al. (2008)

Cross-Model Synthesis and Theoretical Limits

A systematic synthesis distinguishes three scientific questions: how a stimulus is represented, how representations are compared, and how comparison influences a response. Multidimensional scaling and representational similarity methods address the empirical reconstruction of relational structure. Exemplar, prototype, contrast, transformation, and Bayesian accounts define alternative operations over representations. Learning models such as ALCOVE specify how feedback modifies selected parameters and, potentially, the operative structure itself. These questions overlap but are not logically identical. An investigator can adopt the same spatial embedding while testing several decision rules, or keep one rule fixed while testing alternative representational geometries. Shepard (1962a) Kruschke (1992) Tenenbaum & Griffiths (2001)

The exponential generalization principle occupies a special position in this architecture. It links a distance assumption to a predicted graded extension of behavior and can be assessed across many domains. Shepard's 1987 account and Marjieh and colleagues' 2024 examination of naturalistic stimuli make this relation scientifically consequential. However, observing an approximate decay with distance does not by itself determine whether a learner is applying a similarity kernel, recovering stored exemplars, evaluating transformed structures, or inferring a category's extension from a structured prior. Such mechanisms can converge on similar predictions under restricted circumstances. The scientific objective is to identify when they diverge and test those divergences. Shepard (1987) Marjieh et al. (2024) Tenenbaum & Griffiths (2001)

Context sensitivity is likewise a central empirical property, not an inconvenient exception to be erased by extra parameters. Tversky's asymmetries challenge a fixed symmetric distance account, while Goldstone and colleagues' contextual comparisons suggest that dimensions selected during judgment can alter observed relations. Models that allow task-dependent weighting can explain some of these effects, but representation, comparison, and response must still be separately identified. The most parsimonious theory is one that specifies an explicit range of contexts and predicts how a common representational base will change or be deployed within that range. Tversky (1977) Goldstone et al. (1997) Nosofsky (1986)

A general model should ultimately account for learning, transfer, and uncertainty at multiple levels. It should forecast not only which category a participant chooses, but how error varies across novel stimuli, whether the prediction survives a change in goal, and which experimental manipulations alter effective similarity. Bayesian approaches explain why prior structure and sampling can influence extension, while exemplar and connectionist approaches model the effects of remembered instances and feedback. Contemporary theoretical synthesis increasingly treats these ingredients as potentially complementary components of a coherent behavioral model. The value of integration is measured by new constrained predictions, not by creating a vocabulary that can accommodate every outcome after the fact. Wu et al. (2025) Nosofsky et al. (1992) Hahn & Chater (1998)

An Empirical Research Program

The first priority is cross-task measurement of the same candidate space. Investigators can obtain identification confusion data, explicit similarity judgments, categorization learning trajectories, and delayed transfer decisions for an identical controlled set of stimuli. A preregistered modeling plan then determines which representation parameters are shared and which decision parameters may vary. Models should be fitted on specified training measures and evaluated on held-out tasks, with individual participant structure retained where feasible. This design directly tests whether a common geometry carries predictive information beyond the particular task that originally generated it. Nosofsky (1986) Roads & Love (2024)

The second priority is causal manipulation of attention and learned representation. One can change which stimulus dimensions predict feedback, obtain repeated discrimination and eye-movement measures, and test whether transfer depends on inferred selective weighting or category boundary movement. Strong hypotheses predict distinct error patterns for a representation-change model and a response-rule model. Simulation-based recovery should show that the planned sample and stimulus range could distinguish these alternatives before the experiment is interpreted. The statistical report should make clear which conclusions are supported by direct behavioral manipulation and which depend on model-dependent estimates of internal processing. Goldstone (1994) Kruschke (1992) Maddox & Ashby (1993)

The third priority is transfer beyond curated low-dimensional materials. Independent natural-image, sound, or object collections can test whether distance measures estimated from one sample predict learning and generalization for unfamiliar cases. Human ratings, confusion matrices, neural pattern dissimilarities, and machine embeddings should be benchmarked against one another with out-of-stimulus evaluation and carefully documented preprocessing. The resulting contrasts would reveal whether measured similarities are domain-invariant, whether task-specific attention is required, or whether structured hypothesis inference accounts for systematic deviations. Cross-laboratory replication and participant diversity are integral parts of the test. Marjieh et al. (2024) Kriegeskorte et al. (2008) Tenenbaum & Griffiths (2001)

The fourth priority is theoretical discrimination and transparent reporting. Researchers should publish complete stimulus definitions or lawful access routes, trial-level data subject to participant privacy, model code, parameter recovery tests, fit criteria, and all predefined exclusions. They should test models on cases selected to maximize predicted disagreement, including asymmetric comparisons, sparse category exceptions, and transformations that preserve relational structure while changing surface features. For comparative Artificial Era work, the same standards apply to machine models and human participants: state exact versions, evaluation tasks, training exposure where available, and inferential limits. A cumulative scientific understanding emerges when new experiments make explanations more constrained and transferable. Tversky (1977) Hahn et al. (2003) Lake et al. (2015) Wu et al. (2025)

Conclusion

Psychological similarity spaces are empirically useful representations of relations among perceived, remembered, or conceptualized objects. They permit investigators to organize behavioral comparisons and generate explicit predictions about how learning extends beyond observed examples. Multidimensional scaling established methods for recovering relational configurations; Shepard's work connected distance to graded generalization; exemplar, prototype, decision-bound, feature-based, transformation, connectionist, and Bayesian theories developed different explanations of the relation between similarity and choice. Their collective achievement is a research language in which assumptions about representation and generalization can be stated quantitatively and compared experimentally. Shepard (1962a) Shepard (1987) Wu et al. (2025)

The empirical literature establishes that learning conditions and attention can materially affect the similarities that guide behavior. Controlled studies of identification, categorization, and perceptual discrimination reveal the importance of stimulus structure, category contingencies, and selective feature weighting. Contemporary naturalistic work supports continued investigation of generalization–distance relations beyond elementary laboratory patterns, while emphasizing the difficulty of recovering psychological distance from rich, familiar objects. Different formal accounts may nonetheless reproduce the same observed classification probabilities, so predictive adequacy and identification of a psychological mechanism remain distinct scientific demands. Nosofsky (1986) Goldstone (1994) Marjieh et al. (2024) Ashby & Maddox (1993)

The most secure scientific position is that a psychological space is a model-supported, task-tested structure whose explanatory scope must be demonstrated through independent behavior. Its axes and metrics are open to refinement as observers learn, attend, and compare under changing conditions. Direct judgments, cross-task transfers, causal interventions, neural representations, and artificial embeddings provide complementary constraints; none alone licenses an unrestricted interpretation of an underlying cognitive architecture. Generality should be evaluated at the level of explicitly tested invariances across stimuli, tasks, persons, and learning histories. Tversky (1977) Roads & Love (2024) Kriegeskorte et al. (2008)

The continuing research agenda integrates precise experimental design, explicit model comparison, and independent prediction. A successful account will explain why certain dimensions matter, how generalization gradients change with learning, when proximity loses predictive power to structure or context, and which portions of a representation transfer across domains. Within the broader Artificial Era, comparative work can ask how these questions operate in biological and artificial learning systems using common observable criteria. The enduring significance of similarity-space research lies in turning the transition from experience to new judgment into a tractable and revisable scientific problem. Tenenbaum & Griffiths (2001) Lake et al. (2015) Wu et al. (2025)

References