A data-driven investigation of human action representations.

Area of Science:

• Cognitive Science
• Neuroscience
• Psychology

Background:

• Human action understanding involves integrating complex information about agents, scenes, objects, and interactions.
• The cognitive architecture for organizing this action space remains incompletely understood.

Purpose of the Study:

• To identify the fundamental organizing dimensions underlying intuitive judgments of everyday action similarity.
• To determine the dimensionality and semantic content of these action representations.

Main Methods:

• Collected large-scale intuitive similarity judgments from naturalistic videos of everyday actions.
• Applied cross-validated sparse non-negative matrix factorization to uncover underlying structural dimensions.
• Validated dimension robustness through stimulus set perturbations and an odd-one-out experiment.

Main Results:

• A low-dimensional representation (9-10 dimensions) accurately reconstructed human similarity judgments.
• Identified semantic (food, work, home), social (people, emotions), and visual (scene setting) axes.
• Dimensions were robust, reproducible, and interpretable, though not directly aligned with prior hypotheses.

Conclusions:

• The human mind employs a concise set of robust dimensions to organize the complex space of everyday actions.
• Data-driven approaches are crucial for uncovering the structure of behavioral representations.
• Findings inform models of action perception and social cognition.