eccentricity-constrained-cnn-training

name	eccentricity-constrained-cnn-training
title	Eccentricity-Constrained CNN Training Reveals Adaptive Information Coding Around the Visual Field
version	1.0.0
description	Methodology for training CNNs with eccentricity-constrained egocentric video data to reveal adaptive information coding that mirrors primate visual system organization, showing differential task-relevance between foveal and peripheral vision.
trigger_words	["eccentricity-constrained cnn","fovea-periphery vision coding","egocentric visual experience","adaptive information coding visual field"]
domain	neuroscience/computational-neuroscience
authors	["Dylan M. Diaz","Margaret M. Henderson"]
paper_id	arXiv:2607.19316
date	2026-07-21T00:00:00.000Z

Eccentricity-Constrained CNN Training Reveals Adaptive Information Coding Around the Visual Field

Overview

This methodology investigates how eccentricity-dependent visual coding can emerge from natural egocentric experience by training CNNs on gaze-contingent video data that isolates different regions of the visual field. The research demonstrates that models trained on fovea-only vs. periphery-only data develop systematic differences in their representations that align with known properties of the primate visual system.

Key Findings

Biological Alignment

• 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

Model Performance Differences

• Fovea-only models stronger on both face recognition (VGGFace2) and scene categorization (Places365)
• VEDB-pretrained models generalized better to scene categorization than face recognition overall
• Periphery-only models held small but consistent advantage in scene-selective cortex (PPA, RSC)

Neural Predictivity

• VEDB-pretrained models matched neural predictivity of ImageNet-100 models across visual cortex
• Egocentric data supports emergence of cortically-aligned representations
• Scene-selective cortex shows alignment with peripheral statistics

Implementation Steps

1. Data Preparation

• Use egocentric video and eye-tracking data from Visual Experience Dataset (VEDB)
• Create gaze-contingent crops to isolate different eccentricities:
  • Fovea-only crops (central vision)
  • Periphery-only crops (surrounding vision)
  • Periphery-only crops with NeuroFovea transform applied

2. Model Training

• Train ResNet-18 models using contrastive learning (SimCLR)
• Apply the same training protocol across all eccentricity conditions
• Ensure consistent preprocessing and augmentation strategies

3. Downstream Evaluation

• Evaluate on in-domain VEDB frame classification to assess differential informativeness
• Test downstream transfer performance on standard benchmarks: