eccentricity-constrained-cnn-visual-field

name	eccentricity-constrained-cnn-visual-field
description	Eccentricity-Constrained CNN Training methodology for adaptive visual information coding around the visual field using egocentric data
trigger_words	["eccentricity constrained cnn","visual field coding","egocentric video training","fovea periphery models","gaze contingent crops"]
categories	["neuroscience","computational neuroscience","computer vision","deep learning"]
paper	{"title":"Eccentricity-Constrained CNN Training Reveals Adaptive Information Coding Around the Visual Field","authors":["Dylan M. Diaz","Margaret M. Henderson"],"arxiv_id":"2607.19316v1","published":"2026-07-21","conference":"Proceedings of the Conference on Cognitive Computational Neuroscience 2026"}

Eccentricity-Constrained CNN Training for Visual Field Coding

This skill implements the methodology from the paper "Eccentricity-Constrained CNN Training Reveals Adaptive Information Coding Around the Visual Field" (arXiv:2607.19316v1) which demonstrates how visual processing adapts to different parts of the visual field using egocentric experience data.

Key Insights

The research shows that:

• Center-preferring cortical populations have higher spatial resolution and overlap face/word-selective regions
• Periphery-preferring populations have lower spatial resolution and overlap scene-selective regions
• This "eccentricity bias" reflects differential task-relevance across the visual field
• Egocentric experience with eye-tracking data can adaptively constrain cortical information processing

Implementation Steps

1. Data Preparation

Use egocentric video and eye-tracking data from the Visual Experience Dataset (VEDB):

• Extract frames with gaze-contingent modifications
• Create three types of crops:
  • Fovea-only crops: Central region around gaze point
  • Periphery-only crops: Outer regions excluding central area
  • NeuroFovea-transformed periphery: Periphery crops with neural-inspired transformation

2. Model Training

Train ResNet-18 models using contrastive learning (SimCLR):

# Pseudo-code for eccentricity-constrained training
def create_eccentricity_crops(frame, gaze_point, crop_type='fovea'):
    if crop_type == 'fovea':
        return extract_foveal_region(frame, gaze_point, radius=64)
    elif crop_type == 'periphery':
        return extract_peripheral_region(frame, gaze_point, inner_radius=64, outer_radius=256)
    elif crop_type == 'neurofovea':
        periphery = extract_peripheral_region(frame, gaze_point, inner_radius=64, outer_radius=256)
        return apply_neurofovea_transform(periphery)

3. Evaluation Protocol

Evaluate using downstream tasks and neural alignment:

• In-domain classification: VEDB frame categorization across eccentricities
• Downstream classification:
  • Scene categorization (Places365)
  • Face recognition (VGGFace2)
• Neural alignment: Compare with human fMRI data (Natural Scenes Dataset)

4. Analysis Framework

Analyze model performance across visual cortex regions:

• Scene-selective cortex (PPA, RSC): Expect periphery-only model advantage
• Face/word-selective regions: Expect fovea-only model advantage
• General visual cortex: Compare with ImageNet-100 trained models

Expected Outcomes

• Fovea-only models show stronger performance on fine-grained tasks (face recognition, reading)
• Periphery-only models show advantage in scene understanding tasks
• VEDB-pretrained models achieve neural predictivity comparable to ImageNet-100 models
• Scene-selective cortex shows consistent advantage for periphery-only models

Usage Scenarios

Use this methodology when:

• Developing vision systems that need to handle both central and peripheral visual processing
• Creating brain-aligned computer vision models
• Studying how egocentric experience shapes visual representations
• Building adaptive visual systems for AR/VR applications

References

• Diaz, D. M., & Henderson, M. M. (2026). Eccentricity-Constrained CNN Training Reveals Adaptive Information Coding Around the Visual Field. arXiv:2607.19316v1
• Visual Experience Dataset (VEDB): https://vedb.io/
• Natural Scenes Dataset: https://natural-scenes-dataset.org/