On a neuronal level, these may reflect (1) content-selective attentional weighting or surprise signals (see Roesch et al., 2012 for a discussion of such signals in reinforcement learning); BMS-354825 price (2) within- and/or between-subject variation in the direction of signed aPEs; or (3) spatial intermixing of signed and unsigned aPE neurons at a spatial scale that cannot be resolved with fMRI. We also emphasize that the objective of this study is not to make a strong claim about whether or not computations about expertise necessarily involve a Bayesian updating mechanism. Rather, the Bayesian algorithms used here provide

a tractable framework through which we have been able to implicate specific neural structures in mediating computations important for tracking expertise. Although

it is unlikely that subjects uncovered the full structure of the process underlying the agents’ predictions, it is nonetheless the case that the agents in our task did not learn to track the asset behavior (because their performance stayed constant throughout the study). We therefore use the term “expertise” loosely to refer to the participants’ beliefs about the performance level of an agent within a specified domain. This is most likely to be an oversimplification in the real world, where an agent’s expertise is likely to depend on context. For example, someone might be good at picking winning stocks in bull markets, but not in bear markets; or might be good at forecasting stocks, but not bonds. Furthermore, the difficulty of the setting will modulate Galunisertib molecular weight real-world agent performance Sclareol and likely expertise judgments. Determining the role of these contextual factors in evaluating others will provide a richer characterization of social learning in naturalistic settings. A total of 31 human subjects participated

in the experiment. Two subjects were removed from further analysis due to excessive head motion, one because of experimenter error during data collection, and three because they showed no behavioral evidence of learning, resulting in 25 subjects (eight females/17 males, mean age 25 years, age range 18–30). We excluded volunteers who were not fluent English speakers and who had any history of a psychiatric or neurological disorder. All subjects provided informed consent prior to their participation following the rules of Caltech’s IRB. Subjects performed a task in which they had to learn about the performance of a financial asset, as well as about the ability of human and computerized agents who would predict the performance of the asset. Every trial, the asset went up with probability pTRUEt and down with probability 1-pTRUEt. These probabilities evolved over the course of the trial according to the time series shown in Figure 2B (dashed line). Each element of pTRUEt was drawn independently from a beta distribution with a fixed variance (SD, 0.07) and a mean that was determined by the true reward probability on the preceding trial.