Exploring a Specialized Eccentricity-Based Deep Neural Network Model to Simulate Visual Attention

TL;DR

This study uses a unique patient population and a specialized neural network model to investigate visual attention development, revealing insights into model-human differences and guiding architectural improvements.

Contribution

It introduces a novel approach combining neurodevelopmental data with a specialized deep neural network to simulate and analyze visual attention development.

Findings

Model and human reaction times show both similarities and differences.

Human-model divergences highlight architectural areas for improvement.

Longitudinal data reveal developmental trajectories in visual attention.

Abstract

While visual attention theories abound, neurodevelopmental research remains constrained by infants' unreliable responses and limited attention spans. Through collaboration with Project Prakash, we accessed a unique population: patients gaining vision later in life. This cohort enables investigation of visual process development in cognitively mature, cooperative participants rather than infants. We collected data from pre-operation patients, post-operation patients tracked longitudinally (1, 3, 6, and 12 months), and neurotypical controls wearing blurred goggles matched to patients' post-surgical acuity. All participants performed a modified pre-attentive pop-out visual search task. We implemented the eccNET CNN model (Gupta et al., 2021) to simulate visual search asymmetry, subjecting it to identical tasks as human participants. Reaction time comparisons revealed both convergent and divergent patterns between human and model performance. These findings enabled systematic model ablations informed by human physiological constraints and developmental trajectories observed across patient groups. Critically, human-model divergences proved most informative, directing our focus toward specific architectural modifications needed to better approximate human visual development patterns.