Young children combine sensory cues with learned information in a statistically efficient manner: But task complexity matters

Abstract

Human adults are adept at mitigating the influence of sensory uncertainty on task performance by integrating sensory cues with learned prior information, in a Bayes-optimal fashion. Previous research has shown that young children and infants are sensitive to environmental regularities, and that the ability to learn and use such regularities is involved in the development of several cognitive abilities. However, it has also been reported that children younger than 8 do not combine simultaneously available sensory cues in a Bayes-optimal fashion. Thus, it remains unclear whether, and by what age, children can combine sensory cues with learned regularities in an adult manner. Here, we examine the performance of 6- to 7-year-old children when tasked with localizing a 'hidden' target by combining uncertain sensory information with prior information learned over repeated exposure to the task. We demonstrate that 6- to 7-year-olds learn task-relevant statistics at a rate on par with adults, and like adults, are capable of integrating learned regularities with sensory information in a statistically efficient manner. We also show that variables such as task complexity can influence young children's behavior to a greater extent than that of adults, leading their behavior to look sub-optimal. Our findings have important implications for how we should interpret failures in young children's ability to carry out sophisticated computations. These 'failures' need not be attributed to deficits in the fundamental computational capacity available to children early in development, but rather to ancillary immaturities in general cognitive abilities that mask the operation of these computations in specific situations.

Keywords: Bayesian modeling; inference; spatial localization; statistical learning; task complexity.

Figure 1:. Learning and inference in a spatial localization task.

(A) An illustration of a typical trial. Participants estimated the location of a ‘hidden’ target, randomly sampled on each trial from a mixture of two underlying Gaussian distributions. Upon touching a ‘GO’ button, they were presented with uncertain sensory information in the form of a dot cluster, centered on the target location, and subject to one of three levels of variability (low variability shown here; see inset for an illustration of the three levels). Feedback was provided post-touch. (B) An illustration of Bayes-optimal behavior. Considering the example of a ‘hidden’ target drawn from the more variable underlying distribution, the ideal observer would estimate the target location by learning the mean and variance of the prior (the task-relevant underlying distribution) and integrating this knowledge with the likelihood (the sensory information).

Figure 2:. Learning and inference in Experiment 1.

(A) Weights assigned to the centroid of sensory information. 6–7-year olds behaved in a manner dramatically different from that observed previously with adults (see Fig. S2), and inconsistent with the predictions of Bayes-optimal behavior. Each temporal bin includes 300 trials split between the two prior conditions and the four bins are depicted in temporal order. (B) Mean response locations in the absence of sensory information. Participants rapidly learned the true prior means for both the broad (left) and narrow (right) priors. For illustrative purposes, the y-axis represents deviation from the true mean, for each prior condition. Each temporal bin includes 100 trials split between the two prior conditions and the four bins are depicted in temporal order. Columns represent means and error bars represent SEM across participants.

Figure 3:. Learning and inference in Experiment 2.

(A) Weights assigned to the centroid of the sensory information. Like adults (see Fig. S3) and consistent with the predictions of Bayes-optimal behavior, 6–7-year olds relied less on the sensory information (i.e., assigned a smaller weight to the centroid of the sensory information) as sensory uncertainty increased. Furthermore, this drop was greater as participants gained more exposure to the task. Each temporal bin includes 150 trials and the eight bins are depicted in temporal order. (B) Mean response locations in the absence of sensory information. Participants rapidly learned the true prior mean. For illustrative purposes, the y-axis represents the mean deviation from the true mean. Each temporal bin includes 50 trials and the eight bins are depicted in temporal order. Columns represent means and error bars represent SEM across participants.

Figure 4:. Data from the final temporal bin, for a representative participant in Experiment 2.

Consistent with the predictions of Bayes-optimal behavior, as sensory uncertainty increased, participants shifted from selecting the centroid of the sensory information (the diagonal), towards selecting the mean of the underlying prior distribution (zero on the y-axis) as their estimate for the target location. For illustrative purposes, the mean of the underlying prior distribution was subtracted from both the response locations, and the centroid locations. Each dot represents a trial and solid lines represent the best-fit regression lines in each condition.