A Behavioral and Computational Investigation of one Correlational Mechanism for Learning Causal Relations

• In Event: Unifying Neurobiological, Computational, Environmental, and Clinical Studies to Advance Developmental Research: A Multimethodological Approach

Abstract

This study examined 2- and 3-year-old children’s ability to use second-order correlation learning to make causal judgements. Second-order correlation learning refers to the ability to make inferences about features that are only indirectly correlated. For example, if children learn that things with hands are goal-directed and that things with hands tend also to have eyes, can they infer that things with eyes are goal-directed based on a second-order, and hence indirect, correlation between both features? This kind of mechanism is likely used by children in a variety of contexts—including transitive inference (e.g., Mou, Province, & Luo, 2014), and language learning (e.g., Sandoval & Gomez, 2013)—and thus represents a plausible mechanism that may support children’s developing causal-learning capacities. Previous research has shown that the ability to use this mechanism undergoes a development transition between 7 and 11 months of age (e.g., Rakison & Yermolayeva, 2014). The present behavioral study and computational model extends this and related findings (e.g., Rakison & Benton, 2019) to show that 2- and 3-year-old children can use second-order correlation learning to detect the indirect relation between an object’s surface feature and its capacity to activate a novel machine but only if the children had encoded the first-order correlations on which the second-order correlation was based (Figure 1). Specifically, although there was no main effect of age (p > .05)—indicating that the 2- and 3-year-olds were equally likely to engage in second-order correlation learning—only those children who had encoded both sets of correlations that made up the second-order correlation subsequently detected the second-order correlation (p = .01, BF = 5.37). For example, children who had encoded the relation between object A and feature X and then the relation between object A and the machine’s activation, subsequently inferred that feature X and the machine’s activation were related by placing that object, but not another object, on the machine to make it activate. However, children who failed to encode both correlations did not detect the second-order correlation (p = .63, BF = 0.62). The computational (connectionist) model provided an account for why some children showed evidence of second-order correlation learning whereas others did not (Figure 2). The model suggested that the ability to detect second-order correlations depended on having sufficient information-processing capacities. In particular, compared to those models with more impoverished information-processing capacities, networks with increased processing speed (e.g., higher learning rate), improved memory retention (e.g., lower weight decay and momentum), and larger memory storage (e.g., more hidden units) engaged in successful second-order correlation learning. Ongoing work in my lab—which I will also present at the proposed symposium—is investigating whether older infants and young children can use second-order correlation learning to make causal inferences in a category context. For example, if infants observe that red things, as a category, have a particular internal feature (e.g., a cross) and that those same red things, without the internal feature, are causally efficacious, will they infer that a new thing with a cross is causal efficacious?

Author

• Deon Terrel Benton, Vanderbilt University
  Presenting Author