Computational basis of hierarchical and counterfactual information processing

Abstract

Cognitive theories attribute humans unparalleled capacity in solving complex multistage decision problems to distinctive hierarchical and counterfactual reasoning strategies. Here, we used a combination of human psychophysics and behaviorally-constrained neural network modeling to understand the computational basis of these cognitive strategies. We first developed a multi-stage decision-making task that humans solve using a combination of hierarchical and counterfactual processing. We then used a series of hypothesis-driven behavioral experiments to systematically dissect the potential computational constraints that underlie these strategies. One experiment revealed that humans have limited capacity for parallel processing. Another indicated that counterfactuals do not fully compensate for this limitation because of working memory limits. A third experiment revealed that the degree to which humans use counterfactuals depends on the fidelity of their working memory. Next, we asked whether the strategies humans adopt are computationally rational; i.e., optimal under these constraints. To do so, we analyzed the behavior of a battery of task-optimized recurrent neural networks (RNNs) that were subjected to one or more of these constraints. Remarkably, only RNNs that were subjected to all these constraints behaved similarly to humans. Further analysis of the RNNs revealed that what cognitive theories posit as distinctive strategies such as hierarchical and counterfactual are subdivisions in a continuum of computationally rational solutions that includes optimal, counterfactual, postdictive, and hierarchical.