The logical primitives of thought: Empirical foundations for compositional cognitive models

Area of Science:

• Cognitive Science
• Computational Linguistics
• Psychology

Background:

• The compositional language of thought (LOT) is a cornerstone of computational cognitive theories.
• Current models often presuppose primitive LOT components rather than deriving them empirically.
• Statistical inference over structured hypothesis spaces explains learning but lacks empirical grounding for primitives.

Purpose of the Study:

• To empirically determine the primitive components of the language of thought.
• To test how different sets of LOT primitives predict learning curves in concept learning experiments.
• To distinguish between specific language of thought theories using empirical data.

Main Methods:

• Developed a psychologically realistic approximate Bayesian inference framework for LOT models.
• Designed large-scale concept learning experiments focusing on Boolean connectives and quantification.
• Analyzed subject inferences to identify the most consistent LOT primitives.

Main Results:

• Different sets of LOT primitives systematically predict distinct learning curves.
• Subject inferences align with a rich set of Boolean operations.
• Evidence supports first-order, but not second-order, quantification in psychological concepts.

Conclusions:

• Specific language of thought theories can be empirically distinguished.
• The study provides empirical evidence for the nature of LOT primitives in human cognition.
• Findings advance our understanding of learning and concept formation through compositional representations.