We propose that when humans learn sets of relationships theyare able to learn the abstract structure or type of a family of re-lationships, and exploit that knowledge to improve their abilityto learn and generalize in the future, especially in the face ofsparse or ambiguous data. In two experiments we found thatparticipants choose patterns and extrapolate in ways consistentwith sets of previously learned relations, as measured by ex-trapolation judgments and forced-choice tasks. We take theseresults to suggest that humans can detect shared abstract re-lations and apply this learned regularity to perform rapid andflexible generalization.

Models of visual saliency normally belong to one of twocamps: models such as Experience Guided Search (E-GS),which emphasize top-down guidance based on task features,and models such as Attention as Information Maximisation(AIM), which emphasize the role of bottom-up saliency. Inthis paper, we show that E-GS and AIM are structurally simi-lar and can be unified to create a general model of visual searchwhich includes a generic prior over potential non-task relatedobjects. We demonstrate that this model displays inattentionalblindness, and that blindness can be modulated by adjustingthe relative precisions of several terms within the model. Atthe same time, our model correctly accounts for a series ofclassical visual search results.

Cognitive science is often concerned with questions about ourrepresentations of concepts and the underlying psychologicalspaces in which these concepts are embedded. One methodto reveal concepts and conceptual spaces experimentally isMarkov chain Monte Carlo with people (MCMCP), whereparticipants produce samples from their implicit categories.While MCMCP has allowed for the experimental study of psy-chological representations of complex categories, experimentsare typically long and repetitive. Here, we contrasted the clas-sical MCMCP design with a linked variant, in which each par-ticipant completed just a short run of MCMCP trials, whichwere then combined to produce a single sample set. We foundthat linking produced results that were nearly indistinguishablefrom classical MCMCP, and often converged to the desired dis-tribution faster. Our results support linking as an approach forperforming MCMCP experiments within broader populations,such as in developmental settings where large numbers of trialsper participant are impractical.

We explore whether children’s strategies on a causal learningtask show a bias observed in adults towards “exploitative” hy-pothesis revision. Adults and children (ages 4–6) were pre-sented with evidence which initially seemed to conform to asimple, salient rule (e.g. blue blocks activate a machine), butthen encountered evidence that violated this rule. The truerule in the “near” condition was more complex, but could bereached through iterative revision of the salient rule, while inthe “distant” condition, the true rule was comparatively sim-ple, but incremental revision could not yield the true rule. Par-ticipants then predicted the behaviour of a set of new blocks.Adults performed better in the near condition, while in the dis-tant condition adults did not appear to revise their initial hy-pothesis significantly. Unlike adults, children’s overall perfor-mance in both conditions was similar, while condition differ-ences may reflect a broader search for alternative solutions.

Past research indicates that humans can infer hidden causesfrom covariational evidence, and readily use temporal informa-tion to infer relationships among events. Here we explore a set-ting in which people can attribute events to a common hiddencause or causal relationships among observed events, includingcausal cycles, purely on the basis of timing information. Wepresent data from three behavioral experiments and extend pre-viously proposed Bayesian models that makes use of order anddelay information for causal structure learning. Our findingssupport the idea that people rely on the delays between eventsrather than order information alone. Meanwhile, deviationsfrom our model predictions suggest that people have an induc-tive bias against common hidden causes and rely on heuristicsto distinguish between causal structures, such as event over-laps, at least with the cover story considered in these experi-ments. Further, our data suggest that people have particularlyflexible representations of cyclic relationships.

In function learning experiments, where participants learnrelationships from sequentially-presented examples, peopleshow a strong tacit expectation that most relationships are lin-ear, and struggle to learn and extrapolate from non-linear rela-tionships. In contrast, experiments with similar tasks wheredata are presented simultaneously – typically using scatterplots – have shown that human learners can discover and ex-trapolate from complex non-linear trends. Do people have dif-ferent expectations in these task types, or can the results beattributed to effects of memory and data availability? In a di-rect comparison of both paradigms, we found that differencesbetween task types can be attributed to data availability. Weshow that a simple memory-limited Bayesian model is consis-tent with human extrapolations for linear data for both highand low data availability. However, our model underestimatesthe participants’ ability to infer non-monotonic functions, es-pecially when data is sparse. This suggest that people trackhigher-order properties of functions when learning and gen-eralizing.

A key aspect of theory of mind is the ability to reason about other people's desires. As adults, we know that desires and preferences are subjective and specific to the individual. However, research in cognitive development suggests that a significant conceptual shift occurs in desire-based reasoning between 14 and 18 months of age, allowing 18- but not 14- month-olds to understand that different people can have different preferences (Lucas et al., 2014; Ma & Xu 2011; Repacholi & Gopnik, 1997). The present research investigates the kind of evidence that is relevant for inducing this shift and whether younger infants can be trained to learn about the diversity of preferences. In Experiment 1, infants younger than 18 months of age were shown demonstrations in which two experimenters either liked the same objects as each other (in one training condition) or different objects (in another training condition). Following training, all infants were asked to share one of two foods with one of the experimenters – they could either share a food that the experimenter showed disgust towards (and the infants themselves liked) or a food that the experimenter showed happiness towards (and the infants themselves did not like). We found that infants who observed two different experimenters liking different objects during training later provided the experimenter with the food she liked, even if it was something they disliked themselves. However, when infants observed two experimenters liking the same objects, they later incorrectly shared the food that they themselves liked with the experimenter. Experiment 2 controlled for an alternative interpretation of these findings. Our results suggest that training allows infants to overturn an initial theory in the domain of Theory of Mind for a more advanced one.

Function learning research has highlighted the importance ofhuman inductive biases that facilitate long-range extrapola-tions. However, most previous research is focused on aggre-gate errors or single-criterion extrapolations. Thus, little isknown about the underlying psychological space in which con-tinuous relationships are represented. We ask whether peoplecan learn the distributional properties of new classes of rela-tionships, using Markov Chain Monte Carlo with People, andfind that (1) people are able to track not just the expected pa-rameters of a linear function, but information about the vari-ability of functions in a specific context and (2) in many casesthese spaces over parameters exhibit multiple modes.

Recent work has shown that young children can learn about preferences by observing the choices and emotional reactions of other people, but there is no unified account of how this learning occurs. We show that a rational model, built on ideas from economics and computer science, explains the behavior of children in several experiments, and offers new predictions as well. First, we demonstrate that when children use statistical information to learn about preferences, their inferences match the predictions of a simple econometric model. Next, we show that this same model can explain children's ability to learn that other people have preferences similar to or different from their own and use that knowledge to reason about the desirability of hidden objects. Finally, we use the model to explain a developmental shift in preference understanding.

How was the evolution of our unique biological life history related to distinctive human developments in cognition and culture? We suggest that the extended human childhood and adolescence allows a balance between exploration and exploitation, between wider and narrower hypothesis search, and between innovation and imitation in cultural learning. In particular, different developmental periods may be associated with different learning strategies. This relation between biology and culture was probably coevolutionary and bidirectional: life-history changes allowed changes in learning, which in turn both allowed and rewarded extended life histories. In two studies, we test how easily people learn an unusual physical or social causal relation from a pattern of evidence. We track the development of this ability from early childhood through adolescence and adulthood. In the physical domain, preschoolers, counterintuitively, perform better than school-aged children, who in turn perform better than adolescents and adults. As they grow older learners are less flexible: they are less likely to adopt an initially unfamiliar hypothesis that is consistent with new evidence. Instead, learners prefer a familiar hypothesis that is less consistent with the evidence. In the social domain, both preschoolers and adolescents are actually the most flexible learners, adopting an unusual hypothesis more easily than either 6-y-olds or adults. There may be important developmental transitions in flexibility at the entry into middle childhood and in adolescence, which differ across domains.