Incremental vs. Global Updating of Event Representations

Research Question

Overview

This dissertation project investigates the dynamics of event model updating: when people perceive a change during an ongoing event (e.g., a character’s goal shifts, a new object appears), do they update their mental representation incrementally (changing only the affected features) or globally (restructuring the entire event model)?

This question has profound implications for understanding:

• How cognitive resources are allocated during comprehension
• Whether event models are compositional or holistic
• The computational principles underlying event cognition

Theoretical Background

Event Segmentation Theory

Event Segmentation Theory (EST) proposes that:

1. People maintain working models of “what is happening now” (event models)
2. When prediction error increases, an event boundary is perceived
3. The event model is updated with new information

However, EST does not specify the nature of updating:

• Incremental updating: Only changed features are revised (efficient, local)
• Global updating: Entire model is restructured (resource-intensive, holistic)

Computational Predictions

Incremental Updating predicts:

• Gradual changes in neural representations
• Feature-specific memory updates
• Linear relationship between change magnitude and update cost

Global Updating predicts:

• Sudden shifts in neural representations (phase transitions)
• Whole-model reactivation
• All-or-none update dynamics

Hybrid Models

A hierarchical hybrid model suggests:

• Low-level features: Incrementally updated (e.g., object locations)
• High-level structure: Globally updated (e.g., goals, causal relations)

Dissertation Studies

This dissertation includes three complementary studies:

Study 1: Behavioral Evidence for Global Updating

Method: Controlled narrative reading task

Design:

• Participants read short narratives describing everyday events
• Critical manipulation: Mid-narrative change in protagonist’s goal or location
• Measure: Reading times, memory probes, prediction accuracy

Predictions:

• Incremental: Gradual increase in reading time post-change
• Global: Sudden spike in reading time, followed by stabilization
• Hybrid: Depends on feature type (location → incremental; goal → global)

Preliminary Results:

• Goal changes trigger sudden reading time increases
• Location changes show more gradual patterns
• Suggests hierarchical updating (features vs. goals)

Study 2: fMRI Evidence for Neural Reorganization

Method: fMRI during movie watching with embedded event changes

Design:

• Participants watch movie clips with controlled event changes
• Representational Similarity Analysis (RSA): Compare neural patterns before vs. after changes
• Regions of interest: Medial prefrontal cortex (mPFC), posterior medial cortex (PMC), hippocampus

Predictions:

• Incremental: Smooth trajectory in representational space
• Global: Discontinuous jump in representational space
• Hybrid: PMC shows global shifts; early visual cortex shows incremental changes

Current Status: Data collection completed, analysis in progress

Analysis Approach:

# Pseudocode for RSA analysis
def compute_representational_trajectory(neural_data, event_boundaries):
    """
    Compute trajectory of neural representations across event boundaries

    Returns:
    - trajectory: Timecourse of representational change
    - discontinuity_score: Measure of sudden vs. gradual change
    """
    # 1. Extract neural patterns per timepoint
    # 2. Compute pairwise similarity matrix
    # 3. Track trajectory across event boundary
    # 4. Quantify discontinuity (derivative analysis)

Study 3: Computational Modeling of Updating Dynamics

Method: Implement and compare computational models

Models:

1. Incremental Model: Feature-by-feature updating

   Update(t) = Model(t-1) + α × Error(feature_i)
   

2. Global Model: Whole-model resampling

   if Error > threshold:
       Model(t) = Resample(prior, evidence)
   

3. Hierarchical Hybrid: Multi-level updating

   Low-level: Incremental
   High-level: Global (if error exceeds threshold)
   

Evaluation:

• Fit models to behavioral data (reading times, memory)
• Test neural predictions (fMRI representational dynamics)
• Compare model evidence (Bayesian model comparison)

Preliminary Results:

• Hierarchical hybrid model best explains behavioral and neural data
• Threshold for global updating varies by individual
• Prediction error accumulates before global reset

Methodology

Behavioral Paradigms

Narrative Reading:

• Sentence-by-sentence self-paced reading
• Memory probes after each narrative
• Prediction questions (what happens next?)

Movie Watching:

• Naturalistic movie clips (2-3 minutes)
• Event segmentation task (button press at boundaries)
• Post-viewing memory test

fMRI Protocol

Acquisition:

• 3T scanner, TR = 2s
• Whole-brain coverage
• High-resolution structural scan for registration

Preprocessing:

• Motion correction, slice-timing correction
• Spatial smoothing (6mm FWHM)
• High-pass filtering (128s)

Analysis:

• GLM: Model event boundaries and changes
• RSA: Compare neural patterns across time
• Searchlight analysis: Identify regions showing global vs. incremental updating

Computational Modeling

Implementation:

• Python + PyMC3 for Bayesian modeling
• Simulate updating dynamics under different models
• Parameter estimation via MCMC

Key Questions

1. Is updating incremental or global?
   • Hypothesis: Depends on hierarchical level (features → incremental; goals → global)
2. What triggers global updating?
   • Hypothesis: Cumulative prediction error exceeds threshold
3. Which brain regions implement global updating?
   • Hypothesis: mPFC and PMC show global dynamics; sensory cortex shows incremental
4. Are there individual differences?
   • Hypothesis: Working memory capacity predicts global updating threshold

Implications

If global updating is confirmed:

• Cognitive Architecture: Event models are not compositional but holistic
• Resource Allocation: Updating is costly but discrete (not continuous monitoring)
• Neural Implementation: Requires mechanisms for sudden state transitions (attractor dynamics?)

If hierarchical hybrid is confirmed:

• Structured Representations: Event models have hierarchical organization
• Efficient Updating: Balance between local efficiency and global coherence
• Flexibility: Adapt updating strategy to task demands

Timeline

• 2023 Fall: Behavioral studies completed
• 2024 Spring: fMRI data collection completed
• 2024 Fall: Computational modeling (current phase)
• 2025 Spring: Manuscript preparation
• 2025 Summer: Dissertation defense (target)

Related Projects

• Predictive Looking: Prediction errors trigger event boundaries
• Gaze Entropy: Gaze patterns before boundaries suggest proactive control
• Mental Representations (CLIP): Semantic content of event models

Publications & Presentations

In Preparation:

• “Global vs. Incremental Updating of Event Models: Behavioral and Neural Evidence” (target: Psychological Science)
• “Hierarchical Event Model Updating: A Computational Framework” (target: Cognitive Science)

Conference Presentations:

• Subject matter exam: “Incremental vs. Global Updating in Event Cognition” (June 2024)
• Departmental talks: Fall 2023, Spring 2024

Keywords: Event Segmentation, Event Models, Incremental Updating, Global Updating, fMRI, Computational Modeling, Prediction Error, Hierarchical Representations, Dissertation