To glean information from social networks, people must be adept at distinguishing real evidence from hearsay. Here, we investigate human inferences within a simulated social network. We introduce a novel social learning task in which participants infer their surrounding social network structure while also using social communications to aid judgments about the shared environment. We find that the majority of adults correctly identified independent and acyclic network structures, though they struggled to identify cyclic communications. Comparison of judgments with several social inference models supports a naïve social learning account that integrates social evidence with the learner’s own observations in a structure insensitive way. This suggests that people are capable of using communication patterns to identify social influences, but they may still be misled by the distortions of evidence that these network dynamics can produce.

Investigating how people evaluate more or less complex causal theories has been a focal point of research. However, previous studies have either focused on token-level causation or restricted themselves to very small sets of explanatory variables. We provide a new approach for modeling theory selection that foregrounds the balance between observed and latent structure in the mechanism being explained. We combine a Bayesian framework with program induction, allowing an unbounded and partially observable model space through sampling, and reflecting how a preference for simplicity emerges naturally in this setting. Through simulation, we identify two rational principles: (1) Simpler explanations should be favored as latent uncertainty (the number of hidden variables) increases; (2) latent structure is attributed a larger role when the observable patterns become less compressible. We conducted a behavioral experiment and found that human judgments tended to reflect these principles, indicating that people are sensitive to latent uncertainty when selecting between explanations.

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.

The mechanisms of learning stimulus-stimulus relationships are a longstanding research subject in psychology and neuroscience. Although traditional computational models provide valuable insights into learning processes, they often focus on the average behavior of a population. Individual learning trajectories, however, exhibit a diverse range of behaviors not captured by these models. In this paper, we compare sampling-based process-level models (i.e., particle filters) to representative associative and causal models (i.e., augmented Rescorla-Wagner and PowerPC) in their ability to capture individual learning behavior. We use likelihood-free inference incorporating machine-learned summary statistics for model estimation. We conduct a simulation study to demonstrate high model identifiability and test the models on an existing dataset and a newly conducted experiment which replicates and extends previous studies. We find that most participants are best explained by a particle filtering account, but more targeted experimental designs are required to estimate the best-fitting sub-type of these particle filter models.

To what extent is symbolic processing required for intelligent behaviour? Advances in both sub-symbolic deep learning systems and explicitly symbolic probabilistic program induction approaches have recently reinvigorated this long standing question about cognition. While sub-symbolic approaches have shown impressive results, they still lag far behind human cognition, e.g., in the compositional re-use of learned concepts or generalizing to new contexts. Symbolic systems have successfully addressed some of these shortcomings, but face other unsolved issues relating to feature selection, thorny search spaces and scalability. This workshop intends to bring together established and newly emerging perspectives on the debate and explore the recently rekindled interest in hybrid architectures.

Bayesian experimental design (BED) is a methodology to identify designs that are expected to yield informative data. Recent work in cognitive science considered BED for cognitive models with tractable and known likelihood functions. However, as cognitive models have become more complex and richer, their likelihood functions are often intractable. In this work, we leverage recent advances in BED for intractable models and demonstrate their application on a set of multi-armed bandit tasks. We further propose a generalized latent state model that unifies two previously proposed models. Our experiments show that data gathered using optimal designs results in improved model discrimination and parameter estimation, as compared to naive designs. Furthermore, we find that increasing the number of bandit arms increases the expected information gain in experiments.