SUSTAIN (Supervised and Unsupervised STratified Adaptive IncrementalNetwork) is a model of how humans learn categories from examples. SUS-TAIN initially assumes a simple category structure. If simple solutions prove inadequate and SUSTAIN is confronted with a surprising event (e.g., it is told that a bat is a mammal instead of a bird), SUSTAIN recruits an additional cluster to represent the surprising event. Newly recruited clusters are available to explain future events and can themselves evolve into prototypes/attractors/rules. Importantly, SUSTAIN's discovery of category substructure is affected not only by the structure of the world, but by the nature of the learning task and the learner's goals. SUSTAIN successfully extends category learning models to studies of inference learning, unsupervised learning, category construction, and contexts where identification learning is faster than classification learning.

No abstract

The prominence of Bayesian modeling of cognition has increased recently largely because of mathematical advances in specifying and deriving predictions from complex probabilistic models. Much of this research aims to demonstrate that cognitive behavior can be explained from rational principles alone, without recourse to psychological or neurological processes and representations. We note commonalities between this rational approach and other movements in psychology - namely, Behaviorism and evolutionary psychology - that set aside mechanistic explanations or make use of optimality assumptions. Through these comparisons, we identify a number of challenges that limit the rational program's potential contribution to psychological theory. Specifically, rational Bayesian models are significantly unconstrained, both because they are uninformed by a wide range of process-level data and because their assumptions about the environment are generally not grounded in empirical measurement. The psychological implications of most Bayesian models are also unclear. Bayesian inference itself is conceptually trivial, but strong assumptions are often embedded in the hypothesis sets and the approximation algorithms used to derive model predictions, without a clear delineation between psychological commitments and implementational details. Comparing multiple Bayesian models of the same task is rare, as is the realization that many Bayesian models recapitulate existing (mechanistic level) theories. Despite the expressive power of current Bayesian models, we argue they must be developed in conjunction with mechanistic considerations to offer substantive explanations of cognition. We lay out several means for such an integration, which take into account the representations on which Bayesian inference operates, as well as the algorithms and heuristics that carry it out. We argue this unification will better facilitate lasting contributions to psychological theory, avoiding the pitfalls that have plagued previous theoretical movements.

Conceptual features differ in how mentally tranformable they are. A robin that does not eat is harder to imagine than a robin that does not chirp. We argue that features are immutable to the extent that they are central in a network of dependency relations. The immutability of a feature reflects how much the internal structure of a concept depends on that feature; i .e., how much the feature contributes to the concept's coherence. Complementarily, mutability reflects the aspects in which a concept is flexible. We show that features can be reliably ordered according to their mutability using tasks that require people to conceive of objects missing a feature, and that mutability (conceptual centrality) can be distinguished from category centrality and from diagnosticity and salience. We test a model of mutability based on asymmetric, unlabeled, pairwise dependency relations. With no free parameters, the model provides reasonable fits to data. Qualitative tests of the model show that mutability judgments are unaffected by the type of dependency relation and that dependency structure influences iudgments of variability..

Concepts organize the relationship among individual stimuli or events by highlighting shared features. Often, new goals require updating conceptual knowledge to reflect relationships based on different goal-relevant features. Here, our aim is to determine how hippocampal (HPC) object representations are organized and updated to reflect changing conceptual knowledge. Participants learned two classification tasks in which successful learning required attention to different stimulus features, thus providing a means to index how representations of individual stimuli are reorganized according to changing task goals. We used a computational learning model to capture how people attended to goal-relevant features and organized object representations based on those features during learning. Using representational similarity analyses of functional magnetic resonance imaging data, we demonstrate that neural representations in left anterior HPC correspond with model predictions of concept organization. Moreover, we show that during early learning, when concept updating is most consequential, HPC is functionally coupled with prefrontal regions. Based on these findings, we propose that when task goals change, object representations in HPC can be organized in new ways, resulting in updated concepts that highlight the features most critical to the new goal.category learning | attention | computational modeling | hippocampus | fMRI C oncepts are organizing principles that define how items or events are similar to one another. Goals are critical to shaping concepts, by emphasizing some shared features over others. When goals change, previously experienced events may be organized in new ways, resulting in an updated concept that highlights the features most critical to the new goal. For instance, consider purchasing a home. One must learn which features make for the most desirable home. A young couple seeking a cosmopolitan lifestyle may organize potential houses based on trendy features like exposed brick walls, a wet bar, and room for vintage record collections. However, with the news of a baby on the way, the couple's goals are likely to shift. After pouring through parenting books and web forums to learn what makes for a child-friendly home, they may look at those previously seen potential homes in a different light. Instead, family-oriented features such as whether or not a home has a bathtub, is within walking distance to a park, and is in a well-respected school district may matter more resulting in a reorganization of which homes are a good buy. At the core of this example are the fundamental challenges we face in flexible goaldirected learning. When learning new concepts (e.g., child-friendly instead of a trendy house), attention changes focus to different information and items that were conceptually dissimilar (e.g., two houses with and without a wet bar) may become more similar (e.g., they both are close to a park) and vice versa (1). Understanding how conceptual knowledge is created and updated during learning is a cen...

Summary Acts of cognition can be described at different levels of analysis: what behavior should characterize the act, what algorithms and representations underlie the behavior, and how the algorithms are physically realized in neural activity [1]. Theories that bridge levels of analysis offer more complete explanations by leveraging the constraints present at each level [2–4]. Despite the great potential for theoretical advances, few studies of cognition bridge levels of analysis. For example, formal cognitive models of category decisions accurately predict human decision making [5, 6], but whether model algorithms and representations supporting category decisions are consistent with underlying neural implementation remains unknown. This uncertainty is largely due to the hurdle of forging links between theory and brain [7–9]. Here, we tackle this critical problem by using brain response to characterize the nature of mental computations that support category decisions to evaluate two dominant, and opposing, models of categorization. We found that brain states during category decisions were significantly more consistent with latent model representations from exemplar [5] rather than prototype theory [10, 11]. Representations of individual experiences, not the abstraction of experiences, are critical for category decision making. Holding models accountable for behavior and neural implementation provides a means for advancing more complete descriptions of the algorithms of cognition.

SummaryData analysis workflows in many scientific domains have become increasingly complex and flexible. To assess the impact of this flexibility on functional magnetic resonance imaging (fMRI) results, the same dataset was independently analyzed by 70 teams, testing nine ex-ante hypotheses. The flexibility of analytic approaches is exemplified by the fact that no two teams chose identical workflows to analyze the data. This flexibility resulted in sizeable variation in hypothesis test results, even for teams whose statistical maps were highly correlated at intermediate stages of their analysis pipeline. Variation in reported results was related to several aspects of analysis methodology. Importantly, meta-analytic approaches that aggregated information across teams yielded significant consensus in activated regions across teams. Furthermore, prediction markets of researchers in the field revealed an overestimation of the likelihood of significant findings, even by researchers with direct knowledge of the dataset. Our findings show that analytic flexibility can have substantial effects on scientific conclusions, and demonstrate factors related to variability in fMRI. The results emphasize the importance of validating and sharing complex analysis workflows, and demonstrate the need for multiple analyses of the same data. Potential approaches to mitigate issues related to analytical variability are discussed.

Category knowledge can be explicit, yet not conform to a perfect rule. For example, a child may acquire the rule "If it has wings, then it is a bird," but then must account for exceptions to this rule, such as bats. The current study explored the neurobiological basis of rule-plus-exception learning by using quantitative predictions from a category learning model, SUSTAIN, to analyze behavioral and functional magnetic resonance imaging (fMRI) data. SUSTAIN predicts that exceptions require formation of specialized representations to distinguish exceptions from rule-following items in memory. By incorporating quantitative trial-by-trial predictions from SUSTAIN directly into fMRI analyses, we observed medial temporal lobe (MTL) activation consistent with 2 predicted psychological processes that enable exception learning: item recognition and error correction. SUSTAIN explains how these processes vary in the MTL across learning trials as category knowledge is acquired. Importantly, MTL engagement during exception learning was not captured by an alternate exemplar-based model of category learning or by standard contrasts comparing exception and rule-following items. The current findings thus provide a well-specified theory for the role of the MTL in category learning, where the MTL plays an important role in forming specialized category representations appropriate for the learning context.