Segmentation as Proactive Control: Gaze Patterns and Event Boundaries

Research Question

Overview

This project investigates whether patterns in gaze behavior—specifically, increases in gaze entropy—can predict when viewers will perceive an upcoming event boundary. By analyzing eye-tracking data collected during naturalistic movie viewing, I test the hypothesis that gaze becomes more dispersed (higher entropy) before major scene transitions, reflecting proactive shifts in cognitive control.

Background

Event segmentation is the process by which people parse continuous experience into discrete events. Traditional theories suggest that event boundaries are detected reactively—after a change has occurred. However, there is growing evidence that viewers may anticipate upcoming boundaries through:

• Predictive gaze patterns
• Changes in attentional allocation
• Proactive model updating

Gaze entropy quantifies the dispersion of fixations over time. Higher entropy indicates more exploratory gaze patterns, while lower entropy reflects focused attention on specific regions.

Methodology

Participants and Stimuli

• Participants watched naturalistic movie clips while their eye movements were recorded at high temporal resolution
• Movies included everyday activities with clear event boundaries (e.g., cooking, social interactions)

Eye-Tracking Data Collection

• High-frequency eye-tracking (500 Hz or higher)
• Calibration and validation procedures
• Gaze mapped to screen coordinates

Gaze Entropy Calculation

Gaze entropy was computed using the following pipeline:

1. Spatial Binning: Screen space divided into grid cells
2. Fixation Distribution: Proportion of fixations per cell over sliding time windows

3. Entropy Calculation: Shannon entropy computed for each time window:

   H(t) = -Σ p(i,t) × log₂(p(i,t))
   

   where p(i,t) is the proportion of fixations in cell i at time t

4. Event Boundary Alignment: Entropy timecourses aligned to manually segmented event boundaries

Analysis Pipeline

The analysis pipeline is available in the GazeEntropyEB repository and includes:

• Preprocessing scripts (Python): Clean raw eye-tracking data, remove blinks, interpolate missing data
• Entropy calculation (HTML/JavaScript): Compute spatial entropy over sliding windows
• Statistical analysis (R): Mixed-effects models testing entropy changes around boundaries

Key Findings

• Gaze entropy increases predictively before event boundaries, with peak increases occurring 1-2 seconds prior to transitions
• Effect is specific to event boundaries: Entropy does not increase randomly throughout the movie
• Individual differences: Some viewers show stronger predictive entropy increases than others
• Relation to segmentation ability: Viewers with higher boundary agreement show more pronounced entropy increases

Interpretation

These findings suggest that:

1. Event segmentation involves proactive cognitive control: Viewers anticipate upcoming boundaries through exploratory gaze patterns
2. Gaze entropy as a window into cognitive state: Changes in gaze dispersion reflect shifts in attentional strategy
3. Top-down guidance: Predictive entropy increases likely reflect top-down model-based predictions rather than bottom-up visual features

Analysis Code

All preprocessing, analysis, and visualization code is available in the GitHub repository:

• preprocessing/: Eye-tracking data cleaning and preparation
• entropy_calculation/: Scripts for computing spatial entropy
• statistical_analysis/: R scripts for mixed-effects models and visualization
• outputs/: Example figures and plots

Related Publications

This work builds on and extends findings from:

• Predictive looking and event segmentation research
• Mental model inference from gaze patterns
• Event cognition and cognitive control

Future Directions

• fMRI integration: Relate gaze entropy changes to neural signatures of event boundaries
• Computational modeling: Implement predictive models that generate entropy patterns
• Individual differences: Investigate what drives variation in predictive gaze control

Keywords: Eye-tracking, Event Segmentation, Gaze Entropy, Predictive Processing, Cognitive Control, Naturalistic Perception

Code & Resources