How does the similarity structure of memory influence credit assignment in reinforcement learning? Memory spaces vary in how integrated versus separable their constituent dimensions are, and how clustered versus distributed items are across dimensions. Greater integration may cause people to overattribute value to multiple dimensions, potentially leading to generalization errors. Greater clustering may bias people to attribute value to discrete category-centers (category-based learning) rather than map value continuously across the space (function learning). In this study, subjects complete a value-learning task in which stimuli are sampled from low-dimensional perceptual spaces, and reward is mapped to one dimension. Each space is intended to engender a different degree of integration and clustering in memory, such that effects of memory organization on learning can be probed. Additionally, we investigate how credit assignment on each of these artificial perceptual spaces differs to credit assignment on more complex spaces that define real-world semantic concepts.

Values drive our behavioral choices. Ample research has explored the cognitive and neural underpinnings of value- based computations related to decision-making. However, behaviorally relevant values that we associate with real-world objects are often not monetary. For instance, social values associated with specific people are crucial for social behaviors and interactions. Moreover, understanding and attributing social values allows for proper evaluations of potential interactions with others, and can lead to more beneficial social behaviors and relationships. Learning social values has been shown to recruit the same systems as reward values, however how they become associated with specific people remains to be established. The present study examined social value learning of other people using naturalistic face images. We found that before learning, distances between the faces in conceptual similarity spaces were organized corresponding to their perceptual similarity. However, after learning, faces were shifted in a manner that reflected similarity of their associated social values (generosity). Furthermore, distances were positively correlated with a post- learning index of preference to interact with a person in a future cooperative game. In other words, learned social values of the faces seemed to influence their representations in conceptual space, and such representational changes were related to propensities in future behavior.

Semantic knowledge can facilitate or distort new memories,depending on their alignment. We aimed to quantifydistortions in memory by examining how categorymembership biases new encoding. Across two experiments,participants encoded and retrieved image-locationassociations on a 2D grid. The locations of images weremanipulated so that most members of a category (e.g. birds)were clustered near each other, but some were in randomlocations. Memory for an item’s location was more precisewhen it was near members of the same category.Furthermore, typical category members’ retrieved locationswere more biased towards their semantic neighbors, relativeto atypical members. This demonstrates that the organizationof semantic knowledge can explain bias in new memories.

The mind adeptly registers statistical regularities inexperience, often incidentally. We use a visual statisticallearning paradigm to study incidental learning of predictiverelations among animated events. We ask what kinds ofstatistics participants automatically compute, even whentracking such statistics is task-irrelevant and largely implicit.We find that participants are sensitive to a quantity governingassociative learning, DP, independently of conditionalprobabilities and chunk frequencies, as previously considered.DP specifically reflects the uniqueness, as well as strength, ofconditional probabilities; we find that uniqueness is equallyaffected by a single strong alternative predictor as by severalweak predictors. Performance is well captured with anadapted version of the Rescorla-Wagner delta learning rule(Rescorla & Wagner, 1972). We conclude that incidentalpredictive learning is governed by considerations ofuniqueness, and that this is computed by normalizingconditional probabilities by events’ base-rates. This opens thepossibility of common mechanisms between statisticallearning, associative learning, and causal inference.

How do we learn non-physical properties of physical objects? We explored how the statistical structure of eventscan be a source of object property learning. Twenty-five participants saw sequences of visual events surrounding two distinctobjects. Object identity determined 1) the direction of a high transition probability between two events, and 2) the frequency oftwo other events. Learning was unsupervised and unguided. However, participants spontaneously noticed these regularities. Inan explicit, verbal judgment task, they discriminated between frequent vs. rare events (t(24) = 10.7, p< 0.00001) and betweenpredictive vs. non-predictive event pairs (t(24) = 3.04, p<0.01), as appropriate to the object. These statistics gave rise to distinctconceptual interpretations: participants ascribed a causal interpretation to the predictive statistics (t(24) = 1.91, p<0.05) morethan to events frequently co-occurring with the objects (t(24) = 3.00, p<0.01). Such learning may underlie concept acquisition,particularly of functional kinds like artifacts.

The generative capacity of language entails an ability to flexibly combine concepts with each other. Conceptual combination can occur either by using an attribute of one concept to describe another (attributive combination) or by forming some relation between two concepts to create a new one (relational combination). Prior research has addressed whether common or distinct processes support these two putatively different types of combinations. We turn the question around and ask whether the consequences of these combination types on our conceptual system might differ, by comparing semantic memory networks before and after participants perform either attributive or relational conceptual combinations. We find a general effect on the semantic networks: the structure of network decreases after participants conceptually combine some of the concepts in the network. However, the relational combination manipulation has a greater effect. Furthermore, only the relational combination manipulation leads to an increase in the network’s connectivity.

Observers rapidly extract summary statistics from sets ofvisually presented items, like the mean size of a set of circles,or the mean expression of a set of faces. Their excellentability to report summary statistics stands in contrast to near-chance representation of any of the individuals. Here weasked to what extent this ‘ensemble perception’ signatureextends to a more abstract property: relations amongelements. Participants watched ten unique animations ofvisually patterned objects (hereafter, ‘shapes’) colliding witheach other and producing a new shape. Collisions conformedto ABA patterns, such that the result shape always matchedone of the collider shapes. Recognition tests showed thatparticipants accurately recalled the collisions they saw, butalso falsely accepted foils which conformed to the ABApattern but which were not in fact specifically seen (wererearrangements of the original shapes across collisions). Onthe other hand, they were much less likely to accept foilswhich did not conform to the pattern, but were equallydistinct rearrangements (e.g., AAB). This suggests thatparticipants represented the overall, common pattern betterthan the specifics of what they saw; the superior encoding ofthe summary relative to the individuals thus applies tosummaries of relations. However, in contrast to prior findingswith visual features, we did not find that recall of individualpatterns was entirely at chance. Our paradigm offers a way topursue future questions such as the pressures and motivationswhich might govern the trade-off between summarizingevidence vs. retaining individual experiences.

A single concept can manifest in many varied forms,depending on the context in which it is activated. That is,concepts appear to be flexible rather than static. Here weimplement a compositional model of conceptual knowledge inwhich basic-level concepts are represented as graphtheoretical networks, with the specific goal of quantifyingconceptual flexibility. We collect within-concept statisticsusing online participants, construct network models, andvalidate these models in a classification analysis. We thenextract network measures and find that network diversity andcore-periphery structure correspond to conceptual flexibilityand stability, respectively. These results suggest that acompositional network model can be used to extract formalmeasures that are interpretable and useful in the study ofconceptual knowledge.

[This corrects the article DOI: 10.1371/journal.pcbi.1006234.].