Computational basis of hierarchical and counterfactual information processing

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Decision Making

Background:

• Humans employ complex hierarchical and counterfactual strategies for multistage decision problems.
• Understanding the computational constraints underlying these strategies is crucial for explaining cognitive flexibility.

Purpose of the Study:

• To design a task that reliably engages hierarchical and counterfactual decision-making strategies.
• To identify the specific computational constraints that give rise to these human cognitive strategies.
• To test the computational rationality of these strategies under identified constraints.

Main Methods:

• Designed a novel task to elicit hierarchical and counterfactual decision-making.
• Conducted hypothesis-driven experiments to identify key computational constraints.
• Trained recurrent neural networks (RNNs) with systematically varied constraints to model human behavior.

Main Results:

• Identified three key computational constraints: parallel processing bottleneck, limited counterfactual processing capacity, and working memory noise.
• RNNs trained with all three constraints successfully reproduced human-like decision-making strategies.
• Hierarchical, counterfactual, and postdictive strategies were shown to exist on a continuum of rational adaptations.

Conclusions:

• Human decision-making strategies are computationally rational adaptations to specific cognitive limitations.
• A shared set of computational constraints may underlie the flexibility and efficiency of human cognition.
• Provides a unifying framework for understanding diverse human decision-making processes.