Abstract
One challenge our brains face when making decisions is the inherently ambiguous sensory information they receive. Numerous studies have shown that adult observers overcome this challenge by combining their observations with their previous experience (priors) in way that can be close to statistical optimality (Berniker et al., 2010). However, children’s abilities to do this are still developing to at least 9-10 years of age (Chambers et al., 2018). It is possible that children are slower at learning to use new priors than adults. We measured learning rates for priors in an “octopus catching” task, where adults and 6-11-year-old children judged the position of a “hidden” octopus, drawn from a Gaussian distribution with a narrow variance. To accurately predict the location of the octopus, subjects could combine their prior expectations of where the octopus is likely to appear with a noisy sensory cue, a single dot from a Gaussian distribution. The prior variance increased halfway through the experiment to test how participants learned to adapt their behaviour in response to this change of prior. We determined the relative weight given to the prior as compared with the optimal weight. We found that adults showed greater reliance on the prior as the study progressed, approaching Bayesian predictions. Further, adults responded to a switch in the prior variance almost immediately, as shown by a rapid re-weighting of the prior information. In contrast, children were slower to approach the optimal prior weight, and to re-weight the prior during the second half of the experiment. We also found 6-8 year-olds to be slower than 9-11-year-olds, highlighting the developmental trajectory of the ability to integrate expectations in their perceptions. In conclusion, our results suggest that learning to use and weight novel statistical regularities is a major contributor to difficulties with performing Bayesian computations in childhood.