Infinite Ends From Finite Samples: Open-Ended Goal Inference as Top-Down Bayesian Filtering of Bottom-Up Proposals

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Psychology

Background:

• Humans possess a remarkable ability to infer the goals and motivations of others, even with vast possibilities.
• Understanding this capacity is crucial for explaining human social cognition and theory of mind.
• Existing models struggle to account for the speed, accuracy, and rationality of human goal inference.

Purpose of the Study:

• To introduce and validate a novel computational model for open-ended goal inference.
• To explain how humans rapidly and accurately infer plausible goals from observed actions.
• To investigate the integration of top-down and bottom-up processing in human goal understanding.

Main Methods:

• Developed open-ended sequential inverse plan search (SIPS), a sequential Monte Carlo model.
• SIPS combines Bayesian inverse planning with subgoal statistics for efficient goal hypothesis generation.
• Validated SIPS in the Block Words task, comparing its predictions to human performance and other models.

Main Results:

• SIPS accurately predicts human goal inferences, including mean, variance, efficiency, and resource rationality.
• The model achieves high accuracy comparable to exact Bayesian inference but with significantly lower cognitive cost.
• SIPS demonstrates rational pruning of irrational goal hypotheses, a capability lacking in purely bottom-up approaches.

Conclusions:

• Integrating top-down and bottom-up processing is essential for explaining human goal inference.
• SIPS provides a parsimonious and effective computational framework for understanding theory of mind.
• The model highlights the cognitive efficiency and rationality underlying human social cognition.