"How—or by what cognitive mechanism—do children reason about causal events?  An experimental and computational case for domain-general associative learning"

Abstract: Causal reasoning—or the capacity to distinguish causes from effects, to make effects happen, and to make explicit causal judgements—is a key cognitive ability. To date, research has established that this ability emerges by about 18 months of age and undergoes considerable development between 18 months and 5 years of age. Yet, much less is known about how—or by what cognitive mechanisms—children make such causal inferences. For example, it is unresolved whether causal reasoning in children reflects rational processes that can be described by Bayesian inference or instead reflects less sophisticated, albeit equally powerful, processes such as domain-general associative learning. Those who subscribe to the first perspective have argued that associative learning cannot explain how children reason about causes for two reasons. The first reason concerns the fact that infants, children, and adults have been shown to engage in various forms of retrospective reevaluation. This refers to the capacity to update one's causal beliefs at time t based on information presented at time t+1. The finding that certain rudimentary associative models fail to capture people's retrospective reevaluations has been taken to mean that Bayesian inference, but not associative learning, underlies people's causal judgements. The second reason is that because the world contains innumerable correlations, it is impossible to use basic associative learning (like that above) to discover the relevant correlations and use those to make causal inferences.

In the first half of the talk, I will show through behavioral experiments and computational modeling (i.e., Bayesian models and artificial neural networks) that in contexts that matter, such as those that stretch children's information processing, children’s retrospective reevaluations are better captured by simple associative learning than by Bayesian inference. In the second half of the talk, I will show, again through experiments and computational modeling (i.e., artificial neural networks), that children also have access to a more sophisticated form of associative learning and can use it to make causal inferences. I will also describe in the second half of the talk precisely how the model carries out this more sophisticated form of associative learning. As I will suggest, elucidating how such sophisticated associative learning is carried out in the model can shed light on how, mechanistically, it might unfold in the human mind. I will conclude the talk by arguing that although mechanisms such as Bayesian inference can explain causal reasoning in children in some contexts, associative learning, but not Bayesian inference, ultimately provides a better account of children's causal inferences in other contexts.