RESEARCH DOSSIER
Bayesian Models of Cognition and Structured Induction
Abstract. Bayesian models of cognition explain inductive judgments by relating evidence to structured hypotheses and prior knowledge. This fundamental critical narrative monograph analyzes the historical emergence, mathematical foundations, and empirical scope of Bayesian structured induction across category generalization, language acquisition, compositional concepts, causal reasoning, continuous function learning, few-shot program induction, and pedagogy. Formal posterior prediction is evaluated alongside bounded sampling, local model revision, representational learning, and resource-constrained computation. The synthesis contrasts original experiments, computational demonstrations, and methodological criticisms; it distinguishes normative benchmarks from descriptive cognitive models and implemented mechanisms, and examines identifiability, transfer, independent prediction, developmental acquisition, and the human–artificial research horizon. The evidence supports domain-specific successes of structured probabilistic models while leaving open how priors arise, how hypotheses are constructed, and when human learners approximate the Bayesian calculation. This is a scholarly narrative synthesis, not a preregistered systematic review or new experiment.
Keywords: Bayesian cognition; structured induction; inductive reasoning; hierarchical Bayesian inference; structured priors; hypothesis spaces; category learning; causal induction; probabilistic programs; concept learning; few-shot learning; pedagogical sampling; Gaussian processes; bounded rationality; Monte Carlo approximation; model identifiability; cognitive science; Artificial Era
The Scientific Object of Bayesian Cognitive Induction
Bayesian models of cognition specify how a learner should revise graded commitments among hypotheses when evidence arrives, given a model of how the evidence was generated. Structured induction extends that account to hypotheses with internal organization: categories arranged in hierarchies, concepts expressed as rules or programs, causal networks, spatial relations, and explanatory theories. Its psychological object is the relation between prior knowledge, representation, observed examples, and generalization to new cases. The guiding question is how people generate conclusions far beyond a small sample without treating every conceivable generalization as equally likely. The foundational synthesis by Tenenbaum et al. (2006) makes this combination of structured domain knowledge and probabilistic evidence central to a theory of learning, and Kemp & Tenenbaum (2009) formalizes how alternative background structures generate different inferential patterns.
This monograph examines the inferential consequences of selecting a hypothesis space, defining priors, specifying the sample-generating process, integrating hierarchical knowledge, and approximating computations with limited cognitive resources. Observable outcomes include generalization judgments, confidence, predictive choices, concept extensions, interventions on causal systems, responses to teaching, and transfer to novel problems. The science requires a distinction among an ideal-observer benchmark, a descriptive model of individual behavior, an algorithmic process model, and a claim about neural implementation. Bayesian updating is a mathematical operation; a cognitive explanation must additionally specify what hypotheses an individual considers, which evidence is available, what assumptions are learned, and which responses are generated. The methodological criticism of Jones & Love (2011) is therefore incorporated into the research design rather than relegated to an afterthought.
The synthesis proceeds from identifiable original experimental reports and formal models through methodological appraisal and interdisciplinary interpretation. It is a critical narrative review, not a preregistered systematic review, meta-analysis, or new human-participant experiment. Direct experimental evidence is given greater weight for task-specific behavioral predictions than the elegance of a model or the prominence of a theoretical program. Reviews establish historical and conceptual context, while published experiments establish the conditions in which particular effects have been observed. Because computational models can reproduce similar mean judgments through different latent assumptions, the discussion distinguishes quantitative fit, independent prediction, experimental discrimination, and evidence about psychological mechanisms Griffiths et al. (2010) Bowers & Davis (2012).
The subject is situated within Mathematical and Computational Psychology and intersects Cognitive Psychology, Psychology of Language, and Learning and Behavioral Psychology. Its boundary is general-purpose learning from uncertain observations in structured hypothesis spaces. Sensory fusion and precision-weighted predictive perception belong principally to the neighboring Bayesian Perceptual Inference and Predictive Processing dossier. The complete domain of deduction, argumentative reasoning, and informal inference belongs to Reasoning and Inference. Causal induction appears here insofar as graph structure, intervention and likelihood constrain structured generalization; comprehensive causal discovery methodology remains a distinct field. Such boundaries permit cumulative explanation without treating a shared Bayesian equation as evidence that all psychological tasks recruit the same representational machinery.
Historical Development from Generalization to Generative Models
A historical route to modern structured induction begins with the experimental study of similarity and transfer. Roger Shepard's Shepard (1987) proposed that generalization gradients could be interpreted in terms of distances in psychological stimulus spaces and an approximately exponential relation between distance and transfer. This proposal supplied a powerful quantitative target but left open how psychological spaces are acquired, how classes of objects acquire conceptual meaning, and why different domains support different generalization functions. Tenenbaum & Griffiths (2001) extended generalization analyses through a Bayesian account of hypotheses that can represent meaningful groupings rather than isolated similarity neighborhoods. The shift broadened the explanatory target from interpolation among stimuli to extrapolation governed by the structure of a learner's possible concepts.
In the 1990s and 2000s, probabilistic graphical models, hierarchical statistical inference, and cognitive theories of intuitive domains converged. A decisive methodological step was to treat a category, causal relation, or lexical meaning as a candidate hypothesis whose prior plausibility depends on substantive background assumptions. Griffiths & Tenenbaum (2005) distinguished whether a causal link exists from how strong it is, showing that different probabilistic targets explain different human judgments. Tenenbaum et al. (2006) argued that inductive learning involves the interaction of statistical evidence with theories and schemas. The program thereby recognized that rational inference is conditional on the structure of the learner's knowledge; it never follows from probability theory alone which structures are psychologically available.
A second development made the structure itself a possible object of discovery. Kemp & Tenenbaum (2008) explored how learners might infer an abstract form such as a tree, chain, ring, or spatial arrangement from observed relations among entities. Kemp & Tenenbaum (2009) then articulated taxonomic, spatial, threshold and causal models within a common probabilistic framework, making domain-specific generalization patterns a central empirical constraint. Goodman et al. (2008) offered grammatical hypothesis spaces for rule-based concepts, while Xu & Tenenbaum (2007) tested category word learning from sparse positive examples. These works did not replace cognitive representations with undifferentiated probabilities; they made representations explicit mathematical objects whose prior probabilities and predictions could be tested against judgment data.
Subsequent research extended this approach into compositional programs and resource-constrained computations. Tenenbaum et al. (2011) organized the agenda around statistics, structure, and abstraction. Lake et al. (2015) demonstrated a probabilistic program model of handwritten character concepts, using structure to support one-shot classification and generative tasks within a constrained domain. At the same time, Sanborn et al. (2010), Vul et al. (2014), and Bramley et al. (2017) investigated approximate algorithms that make Bayesian objectives psychologically plausible under limited time and memory. Critical exchanges, particularly Jones & Love (2011), Bowers & Davis (2012), and Marcus & Davis (2013), shifted attention toward unconstrained fitting, unfalsifiable priors, process predictions, and normative claims. The contemporary field thus contains complementary research programs: expressive structured models, experimentally grounded model comparisons, and algorithmic explanations of bounded human inference.
The Mathematics of Belief Revision and Prediction
For a hypothesis h and observations D, Bayes' theorem states p(h|D) = p(D|h)p(h)/p(D). The prior p(h) expresses a distribution over hypotheses before the particular observations; the likelihood p(D|h) specifies the probabilities of those observations under the hypothesized generative process; the posterior p(h|D) represents the result of combining both. The evidence p(D) normalizes the posterior across possible hypotheses. In finite spaces, p(D) equals the sum of p(D|h)p(h) across h; in continuous spaces the analogous operation is integration. The formula itself is logically precise yet underspecified as a psychology: every substantive prediction depends on the representation of h, the choice of prior, the form of the likelihood, the data recognized as relevant, and the rule mapping posterior beliefs to observable responses Tenenbaum et al. (2006) Griffiths et al. (2010).
A learner's prediction about an unseen item x can average over uncertain hypotheses: p(y|x,D) = sum_h p(y|x,h)p(h|D). This posterior predictive distribution distinguishes the uncertainty over what the world is like from the variability expected even if a particular hypothesis were true. A maximum-a-posteriori response instead commits to the highest-probability hypothesis, and posterior sampling produces a response by selecting one or more hypotheses from their posterior distribution. These procedures can lead to different trial-level responses even when they share the same probabilistic model of the environment. For cognitive psychology, the response rule is a testable component of the model: a probability assigned by a normative observer need not be directly readable as a participant's reported confidence or as a neural firing rate Vul et al. (2014) Sanborn et al. (2010).
The likelihood encodes how examples were sampled. If observations are independently selected from all objects in a category without regard to communicative intent, a sequence of identical exemplars may carry one evidential meaning. If a knowledgeable teacher intentionally chooses representative or diagnostic examples, that same sequence may imply a different intended concept. A researcher cannot infer a participant's prior from responses without confronting this sampling assumption. The evidence in Xu & Tenenbaum (2007) makes sampling sensitivity important for interpreting early word generalization, while Shafto et al. (2014) models how teacher and learner reason about example selection. These traditions show that experimental instructions, sampling protocols, and assumptions about another agent's knowledge enter the probabilistic structure rather than merely surrounding a formally self-contained calculation.
Hierarchical Bayes introduces shared parameters or hyperparameters governing families of hypotheses. Rather than treating beliefs about each item or domain as independent, it permits experience across related tasks to influence generalization in a new one. This mechanism provides a formal language for learning-to-learn and for rational transfer, but its empirical legitimacy depends on whether participants have reason to infer common structure across contexts. A higher-level prior can itself be uncertain, updated, and nested within a broader model. The resulting flexibility is scientifically valuable when bounded by out-of-sample predictions and independently measured structure. Without such bounds, adjusting prior families after inspecting results can hide the difference between learning a psychological principle and fitting the latest dataset Tenenbaum et al. (2011) Jones & Love (2011).
Hypothesis Spaces, Priors, and Inductive Bias
Induction is underdetermined by finite observations because many hypotheses can accommodate the same examples. A few red circles might indicate a category of red objects, circular objects, colorful geometric tokens, or objects introduced by a trusted demonstrator. The prior is therefore a representation of possible generalizations and their relative plausibility within a specific learning context. Structured priors give preference to candidates with taxonomic coherence, simpler compositional rules, causal regularities, or spatial continuity when the domain warrants those structures. The prior is not a license to posit any desired psychological knowledge: it is an explicit scientific hypothesis about a learner's expectations and their developmental or experiential origins. The contrast between a generic similarity account and a structural hypothesis account appears directly in Tenenbaum & Griffiths (2001) and Kemp & Tenenbaum (2009).
Simplicity can enter probabilistic induction through shorter descriptions, grammar production probabilities, or lower complexity of a program. Under a specified grammar, compact rules may be assigned more prior mass, while highly elaborate exceptions receive less. The formal outcome is sensitive to the grammar's primitives and coding conventions; there is no representation-free measure of conceptual simplicity. In Goodman et al. (2008), grammar-based rules supplied a compositional concept language whose generalizations were compared with human judgments. An important consequence is that multiple kinds of conceptual flexibility become possible: a learner can entertain disjunctions and conjunctions, but the distribution over rules must predict which forms people find natural and which they neglect.
Inductive biases are task-dependent. A taxonomy privileges categories nested by inclusion; a spatial map privileges proximity; a causal graph privileges directed dependence and interventions; a threshold structure privileges one-sided monotonic ranges. The models examined by Kemp & Tenenbaum (2009) clarify why the same empirical demonstration might produce different generalizations in a biological taxonomy, geographic context, or causal mechanism. Successful structured induction is therefore a coordination of probabilistic calculation with domain organization. A model's predictive accuracy depends upon how well the chosen structure resembles the psychological structure of the participant's knowledge and not simply on whether the researcher labels the inference Bayesian.
An unresolved question concerns how a learner constructs the space of hypotheses before posterior updating. It is tempting to depict the learner as searching over a fixed menu of complete theories; human cognition also invents distinctions, reuses parts of previous concepts, and changes its representational vocabulary. Kemp & Tenenbaum (2008) investigates learning structural forms, while Tenenbaum et al. (2011) emphasizes abstraction as a central feature of human cognitive development. Formal models of program induction make compositional hypothesis construction possible but often presuppose a language of programs. A complete psychological account requires empirical tests of the acquisition of representational primitives, transfer of learned abstractions, and limitations that keep the hypothesis search computationally feasible.
Category Generalization and the Size Principle
Category induction reveals why positive examples can carry more information than their mere presence suggests. A learner hearing a new label applied to three small spotted dogs must choose among an animal kind, a narrower dog breed, and a one-off description of the named instances. A Bayesian model formalizes candidate extensions of the label and the probability of observing those examples under each extension. Under a strong-sampling assumption, examples are sampled approximately uniformly from the true extension; repeated observations within a small category can then favor that narrow category over a broad one. This result is commonly called the size principle. Its precise form depends on the sampling process and the definition of possible categories. The central experimental question is whether participants use such information in the circumstances modeled by Xu & Tenenbaum (2007).
Xu and Tenenbaum tested word learning in adults and children with unfamiliar labels for object categories at more than one level of a taxonomy. Their experiments reported systematic changes in generalization based on the number and distribution of positive examples, consistent with the proposed Bayesian treatment Xu & Tenenbaum (2007). These findings demonstrate sensitivity to evidential structure within the tasks; they do not establish that a child explicitly enumerates every hypothesis or computes exact normalizing constants. Developmental inferences require care because linguistic comprehension, attention, pragmatic assumptions, and familiarity with categories also influence choices. The appropriate scientific interpretation is that structured prior knowledge and observations constrain word extension, while implementation and developmental acquisition demand additional evidence.
The strong-sampling assumption is consequential. If an experimenter deliberately selects typical exemplars, an observer may infer category size and communicative intention differently than if objects were drawn randomly from a large bag. In laboratory tasks, participants might infer stronger sampling intentions from demonstrations than experimenters anticipate. Generalization therefore can appear context-sensitive even when prior concepts remain stable. A decisive experiment manipulates the source of examples, the learner's knowledge of the sampling process, and the availability of alternatives while keeping category membership fixed. Such interventions permit a test of whether an observed narrowing of generalization is driven by size-based likelihood, communicative pragmatics, or a generic preference for similarity Shafto et al. (2014).
Similarity accounts and hypothesis-based accounts can overlap in their predictions. Both can produce sharper generalization near observed examples, and both can fit gradients after an appropriate choice of psychological metric or prior. A discriminating test targets cases where similarity and category structure diverge, such as taxonomically related exemplars with low superficial resemblance. Tenenbaum & Griffiths (2001) provides a probabilistic framework for interpreting generalization beyond a single exemplar, and Shepard (1987) identifies the historical significance of gradient-based analyses. The psychological claim grows stronger when a structural model forecasts transfer to a new category organization and when participants' judgments reveal the specific generalizations implied by that structure.
Rule Grammars and Compositional Concepts
Human concepts may be organized by explicit conditions, exceptions, hierarchical combinations, prototypes, and relations among features. A structured rule model chooses hypotheses from a grammar in which expressions can be combined to build more complex candidate concepts. The learner computes evidence for rules consistent with examples and predicts responses to items never encountered during training. Goodman et al. (2008) developed such a model using logical descriptions of feature-based categories and compared its generalizations with established category-learning data and a broad set of seven-feature concepts. The results supported the viability of structured probabilistic representations in these tasks, including patterns of individual variation. They did not make the grammar unique; rival grammars and non-symbolic learners can be evaluated on new diagnostic conditions.
A grammar of concepts specifies a representational vocabulary. Primitive predicates might refer to color, shape, orientation, or relational features; compositional operations combine predicates through logical relations. The grammar assigns probability to expressions, often favoring a short expression over a long one. When concepts must be learned from sparse data, compositionality permits generalization from known parts to a new whole. But a psychological grammar must be grounded in actual attention to features, vocabulary access, memory constraints, and task instructions. If a model permits features unavailable to a participant, it can fit behavior while mischaracterizing the psychological computation. A theory of rule induction is consequently a theory of what can be represented as well as how candidate representations receive posterior weight Goodman et al. (2008) Tenenbaum et al. (2011).
Posterior inference over rules offers a principled interpretation of graded membership and ambiguity without requiring that concepts themselves be fuzzy objects. Several crisp rules may receive nonzero posterior probability, so the posterior predictive chance of including a particular object can be intermediate. This distinction matters when an investigator asks whether people represent prototypes, exemplars, rules, or mixtures. A probabilistic mixture of crisp hypotheses can mimic the surface effect of a graded concept; careful trial-level tests are needed to distinguish uncertainty over rules from a single graded representation. Response times, confidence, transfer after feedback, and sensitivity to exceptions can impose constraints unavailable from aggregate category accuracy alone Sanborn et al. (2010) Jones & Love (2011).
Compositional concepts invite a broader research program in analogy and systematic generalization. A learner who understands a relational rule can apply it to new content, but this ability depends on identifying roles and mapping them to the correct objects. Bayesian program induction represents one proposed explanation of such portability. It remains necessary to separate an expressive hypothesis language from evidence that a person implements program search. Designs that manipulate rule depth, working-memory demand, and transfer distance can discriminate among exact Bayesian grammar inference, heuristic rule generation, exemplar memorization, and network-based generalization. The value of a generative account lies in forecasts for unseen combinations rather than a retrospective description of training examples Goodman et al. (2008) Lake et al. (2015).
Causal Structures and Inductive Explanation
Causal induction requires estimating how observations would differ under alternative possible relations among events. If a drug and symptom frequently co-occur, the learner must consider background symptom rates, alternative causes, and the possibility that the apparent association is incidental. Griffiths & Tenenbaum (2005) separated inference about whether a causal link exists from estimation of the link's strength. Their causal-support model assessed evidence for a link relative to a no-link alternative and accounted for several phenomena of elemental causal learning. This distinction clarifies a general property of structured induction: learning a parameter within an accepted graph and learning which graph is plausible are scientifically different problems with different uncertainty distributions.
Representing causation through a directed graph creates a compact description of conditional dependence, but observational patterns often fail to uniquely identify direction. Controlled interventions alter the data-generating process and can distinguish structures that would otherwise predict the same passive correlations. Psychological experiments can exploit this distinction by allowing participants to activate candidate causes or block alternative pathways. In Bramley et al. (2017), causal-structure learning was analyzed under the difficulty of maintaining and searching over complex graph possibilities, motivating a local revision algorithm. Behavioral evidence from three experiments was reported as support for a model that updates a working causal hypothesis while encountering successive observations. The work connects inference objectives with plausible representational limits.
A common interpretive error is to equate the observer's inferred causal model with verified causal reality. Participants' judgments can be systematically influenced by beliefs about mechanisms, intervention costs, and whether unobserved common causes are likely. A model can describe their predictions accurately while the assumed underlying causal structure remains experimentally underdetermined. Strong tests compare alternative causal graphs through interventions chosen for their divergent predictions, independently estimate participants' background beliefs, and report uncertainty about recovered mechanisms. The separation among causal-support evidence, estimated causal power, and actual causal effects is essential in avoiding confusion between computational representations and interventionally established relations Griffiths & Tenenbaum (2005) Bramley et al. (2017).
Structured causal induction also links explanations to category generalization. When a property belongs to a causally coherent mechanism, learners may expect it to transfer along causal connections rather than superficial similarity. The models examined by Kemp & Tenenbaum (2009) include a causal domain alongside taxonomic and spatial examples, making these differing patterns explicit. An explanatory theory can constrain which properties are stable under environmental change, whereas a simple associative description may be more brittle when a causal system is manipulated. Whether any given participant reasons in this way is an empirical question to be settled through counterfactual predictions, intervention choices, and generalization following structural changes, rather than by asserting that people always use causal theories.
Taxonomic, Spatial, and Threshold Domains
The same set of exemplars can license different generalizations depending on the relational structure of the domain. In a taxonomic domain, properties often extend through nested biological or conceptual classes; in a spatial domain, proximity may constrain an unknown feature; in a threshold domain, the relevant relation can be monotonic, as when a substance becomes active beyond a critical dose. Kemp & Tenenbaum (2009) implemented separate statistical accounts that express these domains through different structural priors while maintaining a shared Bayesian rationale. This approach emphasizes explanatory discipline: predictions are not derived from Bayes' theorem alone but from a psychologically substantive assumption about which domain structure the learner brings to the task.
Taxonomic induction is naturally sensitive to the breadth and diversity of premises. A novel property found in two closely related species may support a different projection than the same property observed in more phylogenetically distinct examples. Yet the relevant psychological taxonomy may be shaped by culture, ecological experience, and conventional labels, so formal distance in a researcher-supplied tree is not automatically cognitive distance. A strong test elicits or measures a participant's taxonomy independently of the induction task and asks whether the resulting model predicts judgments on new properties and species. This prevents circular inference in which a taxonomic model is chosen because it best reflects the responses it is meant to explain Kemp & Tenenbaum (2009).
Spatial induction confronts an analogous problem with maps and relations. A property observed near a landmark may generalize by Euclidean distance, travel networks, barriers, or similarity of environmental conditions. A probabilistic prior over spatial functions can encode these alternatives, but interpretation requires the researcher to specify the participant's presumed knowledge and the spatial transformations preserved by the hypothesis family. Threshold models provide still another family: examples near a boundary may be unusually informative, and positive and negative evidence can constrain a cutoff asymmetrically. These examples show why a domain-informed likelihood and prior jointly determine induction, and why a superficially successful generic kernel may conceal important psychological differences Kemp & Tenenbaum (2009) Lucas et al. (2015).
The scientific lesson is transportability with explicit constraints. If a model fits one taxonomic set, its usefulness grows when it predicts generalization after reorganization of categories, movement of spatial exemplars, or reversal of threshold ordering. Such experiments expose the stability of structural assumptions and whether participants switch structures with a task. Hierarchical learning may help explain these switches, but evidence must distinguish selecting among previously learned structures from inventing a new one. The discovery-of-form model in Kemp & Tenenbaum (2008) provides a route toward structural learning; the inductive comparisons in Kemp & Tenenbaum (2009) make the behavioral importance of selecting the right representation concrete.
Few-Shot Learning and Probabilistic Programs
Learning a reusable concept from one example is a demanding test of induction because the data leave a very large space of plausible interpretations. A learner copying an unfamiliar handwritten character must infer a generative process with strokes, ordering, spatial proportions, and style; successful use also requires recognizing new examples and producing variants. Lake et al. (2015) approached handwritten characters as programs specifying stroke-based procedures with probabilistic components. Their Bayesian program learning model performed at approximately human-level on a one-shot classification task and generated characters evaluated in visual discrimination tests. The principal result is about a specified class of visually structured concepts and carefully defined tasks, not about every human concept or a general equivalence between program induction and cognition.
Program representations offer a hypothesis about why a small observation may support extensive extrapolation. A program encodes reusable parts and their relations, so a learned concept can generate novel instances by varying parameters while preserving structure. The posterior combines priors over plausible drawing processes with the visual evidence provided by the exemplar. For a human learner, comparable competence could arise from prior experience with motor strokes, visual character systems, and graphic conventions. A computational model whose priors incorporate these constraints may display few-shot learning because much relevant structure was acquired or engineered before the test. Consequently, one-shot performance must be described relative to the prior training and representations available to each system Lake et al. (2015) Tenenbaum et al. (2011).
The strongest interpretation of generative success requires multiple kinds of assessment. Classification accuracy measures recognition under a particular set of alternatives; reconstruction assesses memory for the stimulus; generation assesses whether the model can produce recognizable variations; and transfer asks whether the learned program contributes to new tasks. These measures place different constraints on the inferred representation. Visual human-evaluation tests are informative but can conflate structural quality with stylistic familiarity. New work should contrast program-induction predictions with neural generative models, exemplar transformations, and hybrid models under equivalent exposure, comparable task interfaces, and independent evaluation items. A high score on one benchmark licenses a bounded claim about that benchmark, while a broader theory must explain the accompanying distribution of responses and errors Lake et al. (2015).
Program induction also faces the challenge of hypothesis construction. Searching all candidate programs becomes infeasible as expressions grow, especially when many programs yield similar visual or behavioral outcomes. Practical systems therefore use recognition heuristics, restricted grammars, amortized proposals, or search strategies. These details are central to a process-level theory: different proposals predict different order effects, error clusters, latencies, and responses to time pressure. In the psychology of induction, the critical discovery is that structured programs can be a productive representational language. Whether the mind searches such programs, compiles them into reusable procedures, or learns related structures through distributed representations is an open empirical question. The algorithmic issue connects Lake et al. (2015) with the broader resource-rational program of Gershman et al. (2015).
Learning Continuous Relations and Functions
Structured induction extends from discrete categories to continuous functions. When people observe several paired values, they may infer linear trends, smooth curves, periodic relations, abrupt transitions, or more complex patterns. The evidence often supports several possibilities, and predictions outside the sampled range depend strongly on assumptions about functional form. Lucas et al. (2015) developed a rational account of human function learning through Bayesian regression and Gaussian processes, connecting explicit rule learning with similarity-based predictions. This program clarifies why continuity, extrapolation, and changes in slope are principled questions about prior distributions over functions rather than superficial details of a graph-drawing task.
A Gaussian process defines a distribution over possible functions through a mean and covariance kernel. The kernel encodes assumptions about relations among function values: nearby inputs may be expected to yield similar outputs, or patterns may recur periodically. Posterior predictions then provide both an expected function value and uncertainty under the assumed model. This is attractive for psychology because participants may alter extrapolation with prior knowledge, task framing, and visible variation in the observed data. Yet the kernel is a substantive psychological commitment. A flexible researcher-selected mixture can fit many datasets; independent evidence is needed to establish which inductive tendencies participants actually possess and how they are acquired Lucas et al. (2015) Jones & Love (2011).
The experiments synthesized by Lucas et al. (2015) show that rule-based and similarity-based approaches can be reconciled within a probabilistic framework across a variety of function-learning phenomena. Such reconciliation should not erase empirical differences between explicit symbolic extrapolation, interpolation by resemblance, and trial-and-error associative learning. Carefully chosen extrapolation trials, discontinuities, and systematic shifts of function families can create model divergence. Time pressure and concurrent working-memory load may reveal whether participants use the same strategy with different approximation accuracy or switch among representational procedures. Independent task transfer offers an especially demanding criterion: a learner who acquires a periodic pattern should generalize it to a novel response scale if the hypothesized structure was learned rather than memorized.
The acquisition of continuous patterns also raises a distinction between uncertainty from sparse observations and uncertainty about the correct class of models. A learner may confidently estimate a slope within a linear model while remaining unsure whether the true relationship is linear, nonlinear, or periodic. Reporting only the conditional uncertainty of the chosen model understates the uncertainty of the full induction problem. Hierarchical model averaging offers one representation of both levels. Experimental reports should therefore distinguish predictions conditioned on a specified functional family from predictions averaged over multiple families. Failure to do so risks treating apparently precise forecasts as evidence of uniquely identified cognitive representations Lucas et al. (2015) Griffiths et al. (2010).
Pedagogical Sampling and Social Induction
Learning frequently occurs through examples intentionally chosen by knowledgeable people. A child observing a caregiver point to three objects must infer both what the objects share and why those particular objects were selected. A pedagogical model represents a teacher's choice of examples as influenced by the information the teacher wishes the learner to acquire, while the learner interprets the selection with an assumption about that purpose. Shafto et al. (2014) formulated reciprocal inference about teaching and learning for rule-based, prototype, and causal concepts. The study reported three experiments examining model predictions for those domains. Its significance is the explicit incorporation of another agent's evidence-selection policy into the likelihood of the learner's observations.
Pedagogical induction differs from learning from a neutral random sample because examples communicate by being selected. An unusual example may be highly informative when the observer knows it was deliberately chosen to disambiguate a concept; the same example may be treated as an outlier under random sampling. This creates a scientific route from basic Bayesian inference to pragmatic reasoning in language and instruction. The mechanism also depends on trust, inferred expertise, and the learner's understanding of whether the setting is didactic or accidental. A model that assumes a cooperative, knowledgeable teacher must be tested with manipulations of source reliability and intentionality rather than applied universally to all social information Shafto et al. (2014) Xu & Tenenbaum (2007).
A distinctive methodological difficulty appears because teacher and learner beliefs are mutually dependent. The teacher may select examples based on anticipated learner assumptions; the learner may infer the teacher's intention based on anticipated selection rules. Higher-order recursions can improve predictions but also increase flexibility and computational expense. Behavioral tests should independently estimate what teachers choose and what learners infer, ideally across novel concepts and with conditions in which teaching incentives are varied. This helps distinguish genuine pedagogical reasoning from a model that implicitly encodes successful exemplars as especially probable without testing the social process that generates them. Model comparison must evaluate both sides of the interaction Shafto et al. (2014).
In the Artificial Era, the same research questions extend to human learning from adaptive software tutors and artificial conversational systems. The appropriate scientific problem is whether students adjust their hypothesis spaces according to their understanding of an artificial instructor's knowledge, intention, and reliability. The pedagogical model supplies testable predictions about example selection and belief revision, but research cannot simply transfer assumptions about human communicative goals to artificial systems without identifying the system's actual policy. Studies should manipulate the selection algorithm, disclose or conceal its provenance, measure learner beliefs, and compare outcomes across human and artificial instructors. This establishes a tractable bridge between social induction and human–artificial cognitive partnerships Shafto et al. (2014) Gershman et al. (2015).
Learning Structure Across Domains and Development
A central developmental problem is explaining where structured priors originate. Young children display nontrivial expectations about object categories and word reference before they can articulate explicit statistical rules, yet early performance cannot itself determine whether these expectations are innate, acquired through experience, or reconstructed during the task. The experiments of Xu & Tenenbaum (2007) provide evidence about generalization after limited exposure, not a complete developmental genealogy of hierarchical hypothesis spaces. Longitudinal designs can track how category knowledge, linguistic practices, attention to diagnostic features, and assumptions about sampling co-develop. Cross-cultural comparisons can ask whether apparently shared induction rules rest on similar ecological experiences or different paths to convergent performance.
Structural learning may involve inference over families of representations. Kemp & Tenenbaum (2008) presented a computational account of discovering abstract organizational forms from observed relational data, while Tenenbaum et al. (2011) argued for integration across statistics, structure, and abstraction. A learner may infer that a domain is hierarchically organized before identifying particular branch relations, or may discover that a causal network better describes the observations than a spatial gradient. The mechanism of selecting structural form becomes an additional level of induction. Empirically, designs should present novel domains with competing organizing relations and examine whether learners choose the structure predicted by diagnostic examples, then transfer that relation to fresh entities.
Developmental evidence is especially sensitive to the presentation format. Children may have an accurate expectation about a domain yet fail to express it when questions are linguistically demanding, when response options are unfamiliar, or when memory load obscures the relevant comparison. Conversely, apparent early competence may partly reflect pragmatic information from the adult demonstrator. Reliable inference therefore benefits from multiple response modalities, counterbalanced demonstrations, preregistered analyses, and explicit tests of whether children interpret examples as random samples, helpful demonstrations, or accidental encounters. Developmental models should connect assumptions about priors to measurable experience rather than explaining each age difference by adjusting prior concentration after observing responses Xu & Tenenbaum (2007) Shafto et al. (2014).
Across cultures and domains, researchers should separate the mathematical availability of a prior from its empirical distribution in a population. A shared Bayesian formalism may organize results from different groups while their learned representations, task semantics, educational conventions, and sampling expectations differ substantially. Group-average fits should not be treated as proof of a single cognitive algorithm. Individual-response distributions, nested models, and out-of-context transfer can identify where a proposed shared computational architecture is supported. The unresolved scientific question is how agents acquire, revise, and transmit the background structures that make successful induction from scarce evidence possible Griffiths et al. (2010) Tenenbaum et al. (2011).
Bounded Inference and Approximate Cognitive Algorithms
Exact Bayesian inference is often computationally difficult in the spaces that make structured models explanatory. The number of possible causal graphs increases rapidly with the number of variables; the number of compositional programs increases with grammar depth; and hierarchical models can require integration over uncertain parameters and structures. A psychologically plausible account must identify how an agent allocates limited time, attention, and memory to these problems. Sanborn et al. (2010) developed rational approximations for category learning using computational sampling ideas and examined whether approximation mechanisms might account for behavior. The major advance was a proposal for algorithm-level connections rather than the assumption that an ideal posterior is literally computed by the brain.
A sampling interpretation predicts trial-to-trial variability even when the underlying evidence and prior remain fixed. If a learner draws only one candidate from a posterior distribution, a highly probable hypothesis will be selected often but alternative plausible hypotheses will still appear. With more samples, responses may become more stable while requiring additional computation. Vul et al. (2014) analyzed a decision-theoretic tradeoff between the value of taking more samples and the time cost of gathering them. Under assumptions explored in that work, decisions from few samples could be efficient when considered over an extended sequence of choices. The result is a conditional account of bounded rationality; it should not be generalized into the claim that any noisy decision is optimal.
Local revision offers a different family of approximate processes. In online causal learning, a participant might maintain a single working graph and modify an edge when new evidence challenges it, rather than evaluate every possible graph afresh. Bramley et al. (2017) developed and evaluated this idea through a Neurath's-ship-inspired model of causal structure change, relating gradual hypothesis revision to participants' judgments and information-seeking choices. This approach predicts path dependence: two people exposed to the same total evidence in different orders may retain different intermediate hypotheses and pursue different interventions. Such order-sensitive predictions are valuable because a fully exchangeable Bayesian posterior based on the same observations could be invariant to presentation order under its assumptions.
Approximation models should specify their own parameters and potential failures. Sampling temperature, proposal quality, search depth, memory persistence, and stopping thresholds can explain behavior only if they are constrained independently or recovered reliably. Behavioral response times, fixation patterns, confidence, strategy reports, and sequential intervention choices may distinguish algorithms with identical final predictions. Gershman et al. (2015) frames cognitive computation in terms of rationality under constraints, while Vul et al. (2014) and Bramley et al. (2017) supply specific models. The core lesson is that bounded computation is a positive theory requiring measurable process predictions, rather than an unspecified correction applied after an ideal model misses an observation.
Normative, Descriptive, and Mechanistic Explanations
A normative Bayesian model evaluates inference relative to a specified hypothesis space, prior, likelihood and loss function. A descriptive model attempts to reproduce human response distributions. A mechanistic model proposes representational and algorithmic states that generate those responses. These targets can inform one another, but success at one level does not automatically establish another. Jones & Love (2011) argued that unconstrained rational analyses could lose explanatory force when they ignored process data and environmental measurement. Bowers & Davis (2012) criticized flexible accounts of near-optimal cognition that risked accommodating evidence after the fact. Their arguments require cognitive modelers to state what an explanation predicts before a result is known.
The distinction is evident in simple estimation. Suppose a participant's response closely matches a posterior mean computed from a fitted prior. This establishes correspondence under the specified model; it leaves open whether the participant represents that prior, learned it through exposure, retrieves a memorized estimate, or employs a heuristic with similar outputs. If the same model predicts distinctive responses when examples are intentionally selected, priors are independently elicited, or category structure changes, the explanatory claim becomes stronger. A genuinely mechanistic account additionally predicts latency distributions, serial dependencies, errors, and changes under resource constraints. Sanborn et al. (2010) demonstrates why algorithmic approximations can help build such a bridge.
Marcus & Davis (2013) questioned the robustness and post hoc practices of probabilistic models of higher cognition and challenged the inference from probabilistic fit to optimality. The reply to criticism is best supplied by research designs rather than by definitional dispute. A Bayesian model can be treated as a substantive descriptive theory without declaring the agent optimal; competing hypotheses should be articulated in advance and evaluated on data not used to choose their priors or grammars. The existence of alternative probability models within the same domain is an advantage when experiments elicit divergent predictions. Explanatory claims should match what observations actually distinguish.
The most productive scientific stance is pluralistic about model roles while demanding transparency about each claim. Bayesian ideal observers help diagnose information constraints, compositional grammars describe a possible representational organization, causal graphs clarify interventions, and sampling models generate process-level predictions. Some research combines these roles in a coherent hierarchical architecture, but each additional assumption creates an opportunity for refutation. It is unnecessary to treat the Bayesian framework as a complete theory of mind to appreciate its exact contributions: it forces investigators to reveal how background knowledge and evidence jointly determine generalization Tenenbaum et al. (2006) Griffiths et al. (2010) Jones & Love (2011).
Model Comparison, Identifiability, and Reproducibility
The flexibility of Bayesian cognition models makes methodological discipline indispensable. A model with a rich prior can sometimes fit data generated by a much simpler heuristic, and a rule-based model may mimic a similarity model over a restricted stimulus set. Identifiability concerns whether distinct parameter values or competing structures imply distinguishable distributions of observations. If two accounts are observationally equivalent on the available trials, their relative fit cannot establish which mental representation generated behavior. Strong comparisons use recovery simulations, predictive evaluation on held-out tasks, and experiments chosen to maximize the divergence of competing hypotheses. The critical arguments in Jones & Love (2011) identify the scientific cost of fitting unconstrained models without process data.
In-sample likelihood is useful but can reward flexibility. Cross-validation, predictive checks, complexity-aware comparison and hierarchical estimation offer ways to reduce overfitting, provided they are implemented consistently across rivals. A psychological paper should report the available response variables, the fitting procedure, the role of participant-level variability, the priors on parameters, and the data used for model selection. The researcher should then show where a model fails, including systematic deviations under time pressure, novel exemplars, cultural variation and changes in task instructions. A fitted posterior predictive distribution is an observable model output, whereas the actual latent posterior attributed to the participant remains a theoretical inference Griffiths et al. (2010) Bowers & Davis (2012).
A robust evaluation distinguishes three questions. First, does the model reproduce data from the experiment in which it was fitted? Second, does it predict new participants and new materials within the same task family? Third, does it generalize to a different task in which the same representational principle should operate? The third test is particularly important for claims about common structured cognition. A taxonomic hypothesis space that succeeds only with one researcher-designed tree is less compelling than a model that transfers across categories with independently measured structures. Likewise, a program-induction model requires tests beyond the characters and exposure regime used to motivate its design Kemp & Tenenbaum (2009) Lake et al. (2015).
Reproducibility requires access to model specifications, stimuli, code, parameter constraints, and sufficiently detailed analysis procedures. Researchers should distinguish original experimental reports from modeling reanalyses of previously collected judgments and should report failures of independent replication when established. The present monograph does not claim a numerical replication rate for this field because no comprehensive, verified denominator or meta-analytic synthesis has been established here. Its evidentiary conclusions are consequently scoped to inspectable findings of individual studies and comparative methodological arguments. This precision helps prevent an authoritative-sounding narrative from being mistaken for a quantitatively exhaustive evidence review Jones & Love (2011) Marcus & Davis (2013).
Interpreting Evidence: Strengths, Limitations, and Rival Accounts
The strongest positive evidence for structured Bayesian cognition comes from cases in which structured expectations predict a qualitative signature that a less constrained rival would miss. Examples include distinct generalization patterns for taxonomic and spatial domains Kemp & Tenenbaum (2009), word extension influenced by the distribution of positive exemplars Xu & Tenenbaum (2007), teacher-sensitive inference from intentionally selected examples Shafto et al. (2014), and the ability of a compositional program model to support classification and generation from sparse exposure Lake et al. (2015). These are different empirical achievements. Each strengthens its own local theoretical program without demonstrating that all cognition employs a universal probabilistic algorithm.
The evidentiary weight of an experiment depends on its design. A controlled manipulation of the evidence-generating process supports a causal claim about that manipulation's influence on judgments when confounding is adequately addressed. A correlation between fitted prior strength and performance supports an association under that model. A simulation matching a human average shows predictive correspondence. A participant-level model recovered from independent behavioral measurements provides stronger support for a candidate mechanism, yet neural implementation requires additional biological evidence. Distinctions among these claims matter because Bayesian cognition literature spans behavioral experiments, theoretical analyses, statistical fitting and engineering demonstrations Griffiths & Tenenbaum (2005) Sanborn et al. (2010) Lake et al. (2015).
Rival theories deserve explicit consideration. Exemplar accounts can generalize through retrieved similarities; connectionist models can learn structured representations from distributed activity; rule learners can operate by heuristic search; and resource-limited learners can combine learned associations with symbolic constraints. The question is not whether a rival is labeled Bayesian but whether its specified representation and dynamics predict the diagnostic data. A network may implement an approximate probabilistic computation, while a nominally Bayesian model may incorporate a hand-designed grammar too flexible to be identified from participant behavior. Good scientific comparison therefore holds training exposure and testing conditions constant and measures errors, transfer, response times and uncertainty expression Goodman et al. (2008) Jones & Love (2011).
The current synthesis finds well-developed formal models and a diverse set of supportive local experiments, together with continuing uncertainty about the origin of structured priors, the psychological reality of some hypothesis languages, the selection of approximate algorithms, and cross-domain generality. The literature contains methodological criticism that is substantive rather than merely philosophical. These unresolved issues identify the next discriminating experiments: independently measure prior knowledge, manipulate how examples are sampled, vary cognitive resources, evaluate predictions prospectively, and require transfer across unfamiliar structures. They establish a scientific agenda grounded in the intersection of representation, evidence and computational limits Tenenbaum et al. (2011) Jones & Love (2011) Gershman et al. (2015).
Eliciting Priors and Distinguishing Them from Decision Policies
An inferred prior is evidence about cognition only if the experimental design can separate prior expectations from the likelihood, response rule and choice costs. In a category task, a participant may choose a narrow extension because narrow categories are expected, because the examples were understood as deliberately selected, or because the response scale rewards specificity. Fitting a concentrated prior to final choices can reproduce the observations without distinguishing these explanations. A stronger approach measures expectations before target evidence, uses a second task to elicit beliefs about sampling, and evaluates the fitted model on responses to novel evidence. The importance of independent representational constraints follows from the methodological critiques of Jones & Love (2011) and Bowers & Davis (2012), while sampling assumptions emerge directly in Xu & Tenenbaum (2007).
One practical strategy is to construct matched tasks that selectively target competing model components. In a prior-elicitation phase, learners rank candidate structures before any examples are shown. In a likelihood phase, they predict which examples a particular teacher or process is likely to present. In an inference phase, they draw a conclusion from the selected observations. Finally, a decision phase can vary costs and reward structure without changing the evidence itself. Agreement across these stages would support a specific generative account more strongly than an overall match to a single generalization curve. Because elicitation itself can influence subsequent reasoning, studies need randomized groups and counterbalanced order; the measured prior should not be assumed unaffected by the act of asking for it Shafto et al. (2014) Jones & Love (2011).
Hierarchical priors can be investigated using training histories. When one group repeatedly encounters taxonomies and another repeatedly encounters spatially organized features, subsequent transfer to ambiguous domains can reveal how prior knowledge is updated. Such a design should measure whether participants actually encoded the training structures rather than treating all exposure as equivalent. It also requires comparison with models based on memorized exemplars and on general task familiarity. A learner who transfers a relational organization to new entities supplies evidence of structural learning; a learner who improves only on repeated exemplars may primarily exhibit item-specific practice. The discovery-of-form account of Kemp & Tenenbaum (2008) and the cross-domain comparisons of Kemp & Tenenbaum (2009) articulate plausible model predictions for this distinction.
The broader measurement principle is to avoid defining a latent cognitive property solely through the parameter fitted to the behavior attributed to that property. Prior strength, sampling sensitivity, and algorithmic noise should be identifiable from more than one kind of observable consequence. Researchers can exploit repeated responses, confidence, information-seeking decisions and response times, but every added response measure requires its own observation model and reliability assessment. An independently constrained prior may reduce predictive fit relative to an unconstrained one while increasing scientific explanatory value, because fewer assumptions have been selected after observing the decisive outcome. These considerations make preregistered model contrasts and recovery analyses particularly important for claims about individual differences Sanborn et al. (2010) Jones & Love (2011).
Generalization under Distribution Shift and Novel Structures
A learner's success in a familiar category is not sufficient evidence that the same inferential process will generalize to a reorganized domain. Distribution shift arises when the relation between training examples and future observations changes: visual exemplars may come from a new style, words may be used by a different speaker, and interventions may alter a causal system. A probabilistic model can represent uncertainty about the changed process, but adaptation depends on which higher-level variables it allows to change. If the learner assumes one stable likelihood for every environment, a technically coherent posterior may nevertheless yield poor transfer. Hierarchical model selection and structural revision become essential components of understanding cognition across contexts Tenenbaum et al. (2011) Bramley et al. (2017).
Structured representations offer potential benefits when new problems preserve relationships while changing surface features. A taxonomic property might generalize to unfamiliar species through their location in a newly learned tree; a drawing program might adapt to a new character font through variation of strokes; a causal model might carry over when the same mechanism appears under a new measurement scale. These possibilities should be tested with deliberately separated training and transfer environments, with each model frozen before novel outcomes are scored. Evidence of a successful transfer is stronger when controls establish that the learner could not solve the problem using similarity to memorized instances alone. The structured induction models of Kemp & Tenenbaum (2009) and Lake et al. (2015) motivate such experiments.
A second source of difficulty is uncertainty about whether a novel situation belongs to an old domain at all. A participant who sees an unfamiliar biological sample might infer a taxonomic hierarchy, a spatial contagion mechanism, or an arbitrary classification rule. Cross-domain transfer can fail because the prior knowledge is absent, because the wrong structure is selected, or because an approximate algorithm never searches the right alternative. These failures imply different interventions: provide more structural evidence, encourage alternative hypotheses, or grant more processing time. Experimental designs that independently manipulate these factors can reveal where human generalization is constrained Kemp & Tenenbaum (2008) Vul et al. (2014).
Prediction under shift also highlights the difference between calibration and accuracy. A model may make accurate point predictions in familiar situations but remain unjustifiably confident when confronted with an unfamiliar generative process. Researchers should measure whether confidence changes with structural novelty and distinguish uncertainty within a selected hypothesis from uncertainty about the hypothesis family itself. A model that explicitly averages over structural alternatives has a formal advantage only when the alternatives are meaningful and the weights are learned from evidence. Out-of-distribution tests of this kind are particularly important for comparisons involving artificial learning systems, which may carry different training histories and representational restrictions despite similar responses on benchmark tasks Tenenbaum et al. (2011) Gershman et al. (2015).
Scientific Reasoning as an Application of Structured Induction
Research itself illustrates the fundamental problem of induction. A scientist proposes mechanisms, chooses experiments, evaluates observations and revises theory, often in the face of more possible explanations than any finite dataset can discriminate. Bayesian model comparison supplies one formal language for this activity, but substantive scientific progress depends on the investigator's representation of explanatory alternatives and assumptions about measurement. Psychological research can study these operations in constrained tasks without claiming that professional scientific reasoning is merely the execution of Bayes' theorem. In a laboratory model-discovery task, participants can be asked to select interventions among competing causal networks, giving an observable record of how explanatory hypotheses are maintained and tested Griffiths & Tenenbaum (2005) Bramley et al. (2017).
A strength of the structured approach is that causal, mathematical and conceptual assumptions become inspectable. An experiment may favor a theory because it predicts an observation unlikely under a competitor, or it may leave several candidates equally plausible. A rigorous investigator represents uncertainty about models and asks which future observation would separate them. This is also a methodological recommendation for evaluating cognitive theories: tests should target situations where competing accounts diverge, rather than accumulating demonstrations each designed to fit the favored explanation. The critique of accommodating models after observing data in Marcus & Davis (2013) is particularly relevant to scientific psychology's own use of probability.
Scientific research and everyday learning nevertheless differ in institutional practices, specialized tools, explicit models and collective criticism. A laboratory participant often receives a small finite hypothesis space, whereas a research community may revise the variables and concepts used to frame the question. Structured induction can study aspects of hypothesis generation, selective evidence gathering and explanatory revision, but representing historical scientific change requires more than a static set of candidate models. The abstract-form discovery framework Kemp & Tenenbaum (2008) and grammar-based concept models Goodman et al. (2008) suggest ways of making possible structural changes mathematically explicit, though these remain hypotheses whose psychological and historical reach must be tested.
The application yields a reflexive lesson about the quality of this research dossier. Published model comparisons and experimental findings support definite claims about identified tasks; they cannot by themselves justify a universal law governing all inferential practice. A research synthesis should display unresolved assumptions, identify where the evidence concerns simulations rather than humans, and avoid presenting an unregistered narrative literature review as a quantitative census. Such restraint is an evidentiary distinction within scientific reasoning: the scope of an explanatory claim must follow the discriminative power of the observations on which it rests Jones & Love (2011) Bowers & Davis (2012).
Levels of Neural and Computational Implementation
Bayesian formalism can be applied at several descriptive levels. At the computational level, it specifies how uncertainty might be integrated under environmental assumptions. At an algorithmic level, a cognitive theory requires feasible representations, operations and resource constraints. At a biological level, it must identify neural mechanisms and show how their measurable dynamics support the proposed algorithm. The levels are related through hypotheses and experiments, not through the mere use of probabilistic notation. The critical literature has repeatedly emphasized that a claim about near-optimal behavioral output must be distinguished from a claim that neural tissue explicitly represents posterior distributions Bowers & Davis (2012) Jones & Love (2011).
Behavioral evidence can help bridge the first two levels. An exact marginalization account predicts one pattern of responses, a posterior-sampling approximation may predict greater trial variability, and a local graph-revision model may predict sequential path dependence. Studies can discriminate these alternatives using repeated inference problems, controlled resource limitations, and observations of the information people select. Even these tests may leave multiple algorithms viable, especially when researchers observe only final choices. A richer process account benefits from latencies, error profiles and transfer across tasks, while acknowledging that the same algorithm can be implemented through more than one physical substrate Sanborn et al. (2010) Bramley et al. (2017).
Neural measurements require additional experimental discipline. A correlation between neural activity and fitted uncertainty could reflect attention, task difficulty, salience, decision conflict or an approximation to uncertainty, and a modeler must assess plausible alternative variables. Neural interventions, temporally resolved signals and independently estimated cognitive states can sharpen causal interpretation, but translating a mathematical posterior into a neural code is a substantive theory with testable assumptions. The framework of computational rationality in Gershman et al. (2015) helps situate bounded computation within this interdisciplinary problem; its explanatory value depends on specificity about which mechanism is attributed to the observed system.
Artificial implementations present a complementary opportunity because some algorithmic states and training histories can be inspected directly. A program-learning system can be examined for its actual search procedure, prior generation and update steps, while a learned neural system may require separate mechanistic analysis. Comparing an artificial model to human judgments then becomes a question of identifying corresponding functional operations and experimentally distinct predictions, rather than attributing the same internal architecture to systems that share an outcome. The achievement of probabilistic program induction in Lake et al. (2015) motivates these comparisons, and its bounded domain underscores the necessity of checking representation and mechanism independently in each system.
Experimental Methods and Measurements of Inductive Cognition
Experimental tests of structured induction should distinguish the information available to the learner from the information a modeler assumes the learner has. A laboratory protocol needs explicit descriptions of stimuli, prior experience, exemplar-selection rules, instructions, feedback and response mapping. Generalization tasks elicit predictions about unseen cases; confidence judgments measure graded commitment; active-learning paradigms observe which evidence participants request; and intervention tasks examine whether the learner selects a causal manipulation that discriminates hypotheses. Each task probes a different part of the inference process. For instance, selecting a diagnostic example demonstrates information-seeking behavior but does not uniquely establish the full posterior distribution that guided it. Classic category, word-learning and causal experiments illustrate the value of pairing judgments with carefully varied evidence Xu & Tenenbaum (2007) Griffiths & Tenenbaum (2005) Shafto et al. (2014).
Measurement design must make psychological assumptions legible. If responses are forced choices among four categories, accuracy depends on the composition of the alternatives. If participants report probabilities, the response scale and incentives may influence calibration. If they generate explanations, verbal accessibility becomes an additional requirement beyond inference itself. Repeated trials can reveal sequential dependencies and individual variability that disappear in a group mean. Model-based analysis should report the observation model connecting latent beliefs to such responses. A cognitive model without this component risks conflating internally represented probabilities with experimentally elicited statements and choices. Sampling-based process theories are especially suited to predictions about variability across repeated conditions Vul et al. (2014) Sanborn et al. (2010).
Designing a discriminating study begins with a model contrast. The investigator specifies two or more plausible models, simulates the responses each predicts over candidate stimuli, and selects conditions where their predictions differ sharply. In a word-learning experiment this may involve contrasting random and pedagogical sampling; in function learning it may involve extrapolation beyond the training interval; in causal learning it may involve interventions that separate competing graphs. After data collection, researchers evaluate preregistered primary outcomes, examine held-out predictions, and report the degree to which conclusions depend on plausible alternative priors. The purpose is to expose the commitments of each model rather than to optimize a fit after seeing every result Lucas et al. (2015) Bramley et al. (2017).
For individual differences, hierarchical models can partially pool estimates across participants while retaining person-level variation. Such models help prevent noisy parameter estimates from being confused with stable cognitive traits. They also require sensitivity analysis because shrinkage depends on the chosen population distribution. Researchers should report how many individuals and trials contribute to each condition, provide uncertainty intervals, and test whether model parameters can be recovered from the same design. A study using only summary accuracy data may be insufficient to distinguish a strong prior from a high decision threshold or a restricted hypothesis search. Combining choices, confidence, response times, and active evidence requests offers a richer identification strategy Jones & Love (2011) Sanborn et al. (2010).
Cross-Domain Integration and the Artificial Era
Bayesian structured induction contributes to an integrated account of learning by identifying three interacting elements: a representational language of possible explanations, a model of evidence generation, and an algorithm that transforms observations into judgments or actions. The same formal architecture appears in word learning, category generalization, causal induction and continuous-function learning, but each domain supplies distinctive structures and experimental tests. This is a meaningful unification at the computational level. It becomes a stronger psychological unification only when independently measured representational and process constraints explain behavior across domains. A shared mathematical notation alone cannot establish a domain-general mental mechanism Kemp & Tenenbaum (2009) Tenenbaum et al. (2011).
Artificial learning systems make these distinctions more consequential. A model may generalize from a single observation because its developers engineered a compositional grammar, because pretraining created a strong latent prior, because it performs online posterior inference, or because a retrieval procedure matches the example to familiar patterns. Behavioral success does not determine which explanation is correct. Comparative experiments should document training exposure and accessible tools, provide structurally novel tasks, and test interventions aimed at the proposed mechanism. Lake et al. (2015) demonstrates the potential of probabilistic programs for a precisely delimited class of visual concepts, and Gershman et al. (2015) situates computational rationality at the interface of brains, minds and machines. These foundations support testable cross-system comparisons without presuming identity of mechanisms.
Human–artificial collaboration adds a social layer to inductive learning. When an artificial tutor chooses examples, the human learner must estimate the system's reliability, the policy behind the examples and the relevance of the examples to an intended concept. A Bayesian pedagogical model offers testable predictions about how demonstrations influence human generalization. The same questions arise when humans rely on machine-generated explanations: selective examples can change an observer's apparent evidence base. Scientific assessment should distinguish true improvement in transfer from agreement with a model's displayed answer, and it should measure the effects of disclosed versus undisclosed example selection. The original account of teaching by examples supplies a rigorous starting point for these experiments Shafto et al. (2014).
Within the Artificial Era, psychology can therefore compare human and artificial systems in terms of observable inductive competence, underlying learned structure, resource allocation, and the stability of generalization under intervention. Such comparison belongs to an empirical program whose operational variables are specified in advance. Philosophical categories concerning forms of reason may guide questions of interpretation, but they do not substitute for data about particular algorithms or psychological processes. A scientifically useful comparison tests whether two systems represent comparable task hypotheses, whether they interpret sampled evidence in analogous ways, and whether their error patterns respond similarly to manipulations of structure and cognitive resources Tenenbaum et al. (2011) Gershman et al. (2015).
Open Problems and a Cumulative Research Program
The origin of priors is the first major unresolved issue. Every learner brings experience and expectations to a problem, and strong priors permit rapid induction; scientific explanation requires an account of their acquisition and plasticity. Longitudinal experiments should follow the emergence of particular category structures and sampling assumptions, while transfer studies should test whether learned abstractions are reused outside the domain that originally supported them. For modelers, the obligation is to describe the information previously available and distinguish hand-specified structural knowledge from knowledge acquired through experience. The representational agenda of Tenenbaum et al. (2011) and Kemp & Tenenbaum (2008) supplies a basis for such work.
The second problem concerns identifiability of representation and process. A distribution over rules and a distributed similarity system may generate nearly identical choices under common training arrangements. To separate them, researchers require diagnostic tasks with new compositional combinations, richer response measurements, process manipulations and independent replication. Posterior sampling, local graph revision and exact marginalization likewise can agree on aggregate decisions while differing in serial dependence or sensitivity to time pressure. The approximate-algorithm work of Sanborn et al. (2010), Vul et al. (2014) and Bramley et al. (2017) points toward such process-sensitive predictions.
A third problem is the ecological organization of evidence. Many experiments ask people to learn from highly controlled samples, but everyday examples arrive through observation, instruction, search and algorithmic selection. Their sampling mechanisms are rarely interchangeable. Future work should characterize how learners infer source policies, respond to changes in reliability, and distinguish accidental from communicative examples. In social settings, teaching can alter the evidence likelihood; in causal settings, active intervention changes what is observed. Treating the generative process as a measurable object creates a unified research target for pedagogy, causal learning and digital information environments Shafto et al. (2014) Griffiths & Tenenbaum (2005).
A fourth problem is cross-domain generalization of models. Research must establish which priors and approximation strategies are shared and which are task specific. A rigorous program should compare models on held-out domains and participants, document failures, report uncertainty about model choice, and maintain reproducible code and stimulus materials. Hierarchical models can represent both shared structure and individual differences, but transfer claims should survive tests outside the original fitting environment. Normative optimality should be evaluated relative to explicit environmental objectives and resource costs. These measures convert broad claims about rational cognition into falsifiable predictions grounded in experimental observations Jones & Love (2011) Gershman et al. (2015).
Conclusion
Bayesian models of cognition and structured induction provide a formal account of how uncertain evidence interacts with the organized knowledge that makes generalization possible. The central explanatory object is a learner choosing among structured possibilities under assumptions about what data are likely to be observed. Priors capture knowledge about candidate concepts, categories, relations and theories; likelihoods capture the evidence-generating process; posterior distributions quantify revised uncertainty; predictive and decision rules connect that uncertainty to behavior. This architecture has been fruitfully elaborated in taxonomic and spatial induction, lexical acquisition, rule-based categories, causal learning, continuous function prediction, pedagogy, and one-shot visual concept learning Tenenbaum et al. (2006) Kemp & Tenenbaum (2009) Lake et al. (2015).
Original empirical work supports specific instances of structured probabilistic explanation, while methodological criticism identifies the conditions required for stronger cognitive interpretation. A mathematical posterior may provide an illuminating standard and a predictive description without establishing the algorithm or representation used by an individual person. Approximation models introduce additional hypotheses about samples, search, memory and sequential revision that are open to behavioral testing. The evidence is most persuasive when models predict new observations, outperform independently specified rivals, and make their representational commitments explicit Sanborn et al. (2010) Jones & Love (2011) Bramley et al. (2017).
The resulting synthesis is a research framework with clearly bounded successes and unresolved scientific tasks. Future progress depends on tracing the acquisition of structured priors, independently identifying hypothesis languages, measuring inference processes, and testing the transportability of learned structures across contexts and systems. These inquiries illuminate a general fact of cognition: powerful induction from scarce evidence requires both constraints on what can be imagined and principled revision when the world provides new information. Bayesian structured models offer one precise and productive language for investigating that relationship, subject throughout to the standards of experimental discrimination and reproducible explanation.
References
- Bowers, J. S., & Davis, C. J. (2012). Bayesian just-so stories in psychology and neuroscience. Psychological Bulletin, 138(3), 389–414. https://doi.org/10.1037/a0026450
- Bramley, N. R., Dayan, P., Griffiths, T. L., & Lagnado, D. A. (2017). Formalizing Neurath’s ship: Approximate algorithms for online causal learning. Psychological Review, 124(3), 301–338. https://doi.org/10.1037/rev0000061
- Gershman, S. J., Horvitz, E. J., & Tenenbaum, J. B. (2015). Computational rationality: A converging paradigm for intelligence in brains, minds, and machines. Science, 349(6245), 273–278. https://doi.org/10.1126/science.aac6076
- Goodman, N. D., Tenenbaum, J. B., Feldman, J., & Griffiths, T. L. (2008). A rational analysis of rule-based concept learning. Cognitive Science, 32(1), 108–154. https://doi.org/10.1080/03640210701802071
- Griffiths, T. L., & Tenenbaum, J. B. (2005). Structure and strength in causal induction. Cognitive Psychology, 51(4), 334–384. https://doi.org/10.1016/j.cogpsych.2005.05.004
- Griffiths, T. L., Chater, N., Kemp, C., Perfors, A., & Tenenbaum, J. B. (2010). Probabilistic models of cognition: Exploring representations and inductive biases. Trends in Cognitive Sciences, 14(8), 357–364. https://doi.org/10.1016/j.tics.2010.05.004
- Jones, M., & Love, B. C. (2011). Bayesian fundamentalism or enlightenment? On the explanatory status and theoretical contributions of Bayesian models of cognition. Behavioral and Brain Sciences, 34(4), 169–188. https://doi.org/10.1017/S0140525X10003134
- 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
- Kemp, C., & Tenenbaum, J. B. (2009). Structured statistical models of inductive reasoning. Psychological Review, 116(1), 20–58. https://doi.org/10.1037/a0014282
- Lake, B. M., Salakhutdinov, R., & Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 350(6266), 1332–1338. https://doi.org/10.1126/science.aab3050
- Lucas, C. G., Griffiths, T. L., Williams, J. J., & Kalish, M. L. (2015). A rational model of function learning. Psychonomic Bulletin & Review, 22(5), 1193–1215. https://doi.org/10.3758/s13423-015-0808-5
- Marcus, G. F., & Davis, E. (2013). How robust are probabilistic models of higher-level cognition? Psychological Science, 24(12), 2351–2360. https://doi.org/10.1177/0956797613495418
- Sanborn, A. N., Griffiths, T. L., & Navarro, D. J. (2010). Rational approximations to rational models: Alternative algorithms for category learning. Psychological Review, 117(4), 1144–1167. https://doi.org/10.1037/a0020511
- 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. (1987). Toward a universal law of generalization for psychological science. Science, 237(4820), 1317–1323. https://doi.org/10.1126/science.3629243
- Tenenbaum, J. B., & Griffiths, T. L. (2001). Generalization, similarity, and Bayesian inference. Behavioral and Brain Sciences, 24(4), 629–640. https://doi.org/10.1017/S0140525X01000061
- 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
- Tenenbaum, J. B., Kemp, C., Griffiths, T. L., & Goodman, N. D. (2011). How to grow a mind: Statistics, structure, and abstraction. Science, 331(6022), 1279–1285. https://doi.org/10.1126/science.1192788
- Vul, E., Goodman, N., Griffiths, T. L., & Tenenbaum, J. B. (2014). One and done? Optimal decisions from very few samples. Cognitive Science, 38(4), 599–637. https://doi.org/10.1111/cogs.12101
- 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