ORCID
https://orcid.org/0000-0002-8433-4678
Date of Award
Winter 12-2022
Degree Name
Master of Arts (AM/MA)
Degree Type
Thesis
Abstract
To act effectively, humans store event schemas and use them to predict the near future. How are schemas learned and represented in memory, and used in online comprehension? One means to answer these questions is modeling event comprehension. What are, then, computational principles of event comprehension? We proposed three candidate properties: 1) abstract representation of visual features, 2) predictive mechanism and prediction error as feedback, and 3) contextual cues to guide prediction, and adapted a computational model embodying these properties. The model learned to predict activity dynamics from one pass through an 18-hour corpus of naturalistic human activity. Evaluated on another 3.5 hours of activities, it updated at times corresponding with human segmentation and formed human-like event categories—despite being given no feedback about segmentation or categorization. These results establish that a computational model embodying the three proposed properties can naturally reproduce two important features of human event comprehension.
Language
English (en)
Chair and Committee
Professor Jeffrey Zacks, Chair
Committee Members
Professor Todd Braver, Professor Wouter Kool
Recommended Citation
NGUYEN, TAN, "A Hybrid Model of Event Comprehension Predicts Human Activity at Human Scale" (2022). Arts & Sciences Electronic Theses and Dissertations. 2822.
https://openscholarship.wustl.edu/art_sci_etds/2822