Suayb Arslan^1,2,4 (suayb.arslan@bogazici.edu.tr), Fazilet Yildirim-Keleş³, Hannan Toprak²; ¹Department of Computer Engineering, Boğaziçi University, ²Institute for Data Science and Artificial Intelligence, Boğaziçi University, ³Department of Psychology, Boğaziçi University, ⁴Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology

Visual acuity and crowding develop along different timelines in childhood. Acuity reaches adult-like levels by approximately 5–6 years, while crowding –the deleterious influence of nearby elements on object recognition– continues to improve beyond this age, suggesting that these functions rely on different stages of visual cortical development. To investigate the mechanisms underlying this dissociation, we implemented a biomimetic deep neural network that mimics hierarchical visual processing, from the retina through primary (V1) and intermediate (V2–V4) visual areas. The network incorporated foveated retinal sampling, V1-like oriented filters, and higher-level convolutional layers whose pooling, lateral interactions, and cortical magnification could be selectively adjusted to simulate developmental progression. We evaluated network performance at three simulated developmental stages. Acuity was measured by the model’s ability to identify isolated targets, while crowding was measured using targets flanked by surrounding stimuli, quantifying critical spacing. At the early stage, the network’s retinal and V1-like layers were fully mature, whereas higher-level layers remained immature. In this stage, the network achieved adult-like acuity but exhibited large critical spacing, consistent with prolonged crowding deficits in children. At intermediate stages, partial maturation of higher layers improved crowding performance, and at the final stage, fully matured higher layers yielded both adult-like acuity and crowding. Systematic manipulations revealed that increasing pooling size, strengthening lateral inhibition, or refining cortical magnification in higher layers reduced critical spacing without affecting acuity. Networks lacking hierarchical constraints failed to reproduce the developmental dissociation. These results indicate that early maturation of retinal and V1-like processing explains the rapid development of acuity, whereas delayed refinement of spatial integration in higher visual areas accounts for prolonged crowding. Biomimetic neural networks provide a precise framework to link neural maturation to behavioral outcomes and generate testable predictions for human visual development.

Acknowledgements: This work is supported by TUBITAK with grant number 125E068.