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

Area of Science:

• Neuroscience
• Computer Vision
• Visual Perception

Background:

• The primate visual system exhibits an "eccentricity bias," where central vision has high spatial resolution for tasks like face recognition, and peripheral vision has lower resolution, suited for scene understanding.
• This bias suggests that different visual tasks are relevant to different parts of the visual field.

Purpose of the Study:

• To investigate if natural, egocentric visual experience can lead to the development of eccentricity-dependent coding in artificial neural networks.
• To determine if models trained on egocentric data align with human brain activity in visual cortex.

Main Methods:

• Trained ResNet-18 models using contrastive learning (SimCLR) on egocentric video frames from the Visual Experience Dataset (VEDB).
• Modified frames to isolate central (fovea-only) and peripheral (periphery-only) visual information.
• Evaluated model performance on downstream classification tasks (scene and face recognition) and alignment with human fMRI data.

Main Results:

• Models trained on egocentric data showed differential informativeness for various categories, indicating task-specific processing.
• Fovea-only models showed stronger performance on both scene and face categorization tasks.
• Models trained on egocentric data aligned with neural predictivity in visual cortex, with periphery-only models showing an advantage in scene-selective regions.

Conclusions:

• Egocentric visual experience can adaptively shape visual information processing in the brain.
• The findings suggest that the visual system's specialization for different tasks emerges from natural, everyday visual input.
• Peripheral vision's role in scene processing is supported by its alignment with peripheral visual statistics.