Modeling human activity comprehension at human scale: prediction, segmentation, and categorization.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Human event comprehension involves segmenting continuous activity into discrete events.
• The cognitive mechanisms underlying event segmentation and model switching are not fully understood.
• Existing models often lack the ability to learn event schemas or emulate human-like segmentation.

Purpose of the Study:

• To develop a computational model that learns event schemas and predicts human activity.
• To investigate computational mechanisms for event model transitioning in human-like segmentation.
• To compare prediction uncertainty and prediction error as triggers for event segmentation.

Main Methods:

• Constructed a hybrid recurrent neural network and Bayesian inference architecture.
• Trained the model on naturalistic human activity data.
• Evaluated model performance against human segmentation and categorization data.

Main Results:

• The model learned to predict human activity with human-like segmentation and categorization performance.
• Both prediction uncertainty and prediction error variants learned to segment and categorize events.
• The prediction uncertainty variant showed a closer match to human segmentation and categorization.

Conclusions:

• Computational models can learn event schemas and emulate human event segmentation.
• Event model transitioning based on prediction uncertainty or error are plausible mechanisms for human event comprehension.
• The findings suggest prediction uncertainty is a key factor in human event segmentation.