When gathering information, different sources typically have distinct levels of informativeness. Therefore, it is optimalto actively select the source of information to learn from (i.e., perform active learning). It has been debated whetherhumans optimize task performance in active learning or use a simple heuristic of seeking information that confirms theirbeliefs. Critically, depending on ones subjective beliefs, confirmation bias can in fact be optimal. Thus, without measuringsubjective beliefs, previous approaches were unable to distinguish between these alternatives. Using a perceptual decision-making task, we measured participants subjective beliefs before and after a new piece of information was presented.We then characterized confirmation-based and performance optimizing strategies with respect to these subjective beliefs.We found that participants strategy was dominated by confirmation bias, modulated only weakly by the performanceoptimization. We discuss potential reasons that may limit performance optimization in active learning.

There has been little cross-fertilization between research on ac-tive learning and teaching, despite extensive conceptual simi-larities. The current study aims to bridge the gap by show-ing that engaging in active learning can influence subsequentteaching performance. In a one-dimensional boundary teach-ing task, participants who first took the role of an active learnerwent on to become better teachers than participants who didnot. In order to disentangle the effect of active selection ofsamples from their information content, the performance ofactive learners was compared to that of yoked passive learn-ers. While prior passive learning also significantly boostedteaching performance, it did so to a lesser extent. However, inpaired comparisons, teachers with active learning experiencedid not differ significantly from their yoked-passive learningcounterparts. Based on the current results we cannot arguefor a teaching benefit specific to active learning as opposed toa more general improvement caused by experiencing the taskfrom the learner’s perspective. However, we suggest that thisis a promising line of inquiry using more complex learning andteaching tasks.

Prediction Error [PE] is a core component of some of the most influential theories of how animals use experiencesto update their knowledge (e.g, Rescorla & Wagner, 1972). The classic demonstration of PE is the single-cell recording doneby Schultz and colleagues (1997). However, there is no evidence that this signal plays any role in learning.Only two studies have related a neural correlate of PE to learning performance so far (Gl ̈ascher, Daw, Dayan, & O’Doherty,2010; McGuire, Nassar, Gold, & Kable, 2014). We provide a formal analysis demonstrating that non-PE learning can alsoexplain the results of these studies if the imaging signal they identify relates to the size of weight updates instead of PE.We conclude that the case for PE driving many forms of animal learning is not yet sufficiently proven, and identify approacheswhich can potentially resolve this question in future.

The optimal perceptual decision making strategy for weighting serially presented information depends on the degreeof sample dependence. Uniform weighting produces optimal estimates from independent samples, but increases in autocorre-lation should be matched by increasing and symmetric overweighting of early and late samples in order to maintain optimalperformance.In the current experiment, participants (N = 30) observed briefly presented sequences of eight dots and were asked to estimatetheir center of mass by dragging the cursor. The autocorrelation of the series was manipulated in two distinct blocks (either 0or .7). Preliminary results show that the weight assignment to uncorrelated inputs did not differ significantly from the optimaluniform allocation. In contrast, in the high-dependence block participants used different weighting profiles - overweightingthe first and/or last samples of the sequence. This suggests that humans flexibly adapt to changes in statistical structure in thepredicted direction of optimality.