A Window into the Infant's Visual World: Through a NeuroAI Lens

Abstract

For centuries, scientists, philosophers, and caregivers have wondered what infants perceive of the world around them. While recent work has greatly advanced our understanding of infant perceptual development and early visual experience, much less is known about the visual features encoded by the developing brain when infants view objects. In adults, NeuroAI approaches, which combine neuroimaging with artificial intelligence (AI) models, have proven highly effective at characterising the features encoded by the ventral visual stream (VVS), the cortical pathway that supports object processing. Due to methodological challenges, comparable work has not previously been possible in infants. I address this gap using awake infant functional MRI (fMRI), that we acquired through the Foundations of Cognition project, and developmentally inspired deep neural networks (DNNs) to characterise the features encoded in infant visual cortex. I focus specifically on the spatial scale of these features, as objects can be discriminated using visual information from global shape to fine detail. Limited visual acuity in young infants has been proposed to bias visual processing towards coarse visual features, yet the cortical circuitry supporting mature spatial integration is still developing, and behavioural sensitivity to global object shape continues to develop beyond infancy. It therefore remains unclear how visual features at different spatial scales are encoded across the developing VVS. To address this question, I first demonstrate that awake infant fMRI can yield robust group-level visual responses suitable for computational modelling. I then compare developing VVS responses to DNNs trained on images with graded amounts of blur, systematically varying model sensitivity to coarser versus finer object features. Comparing model and brain responses reveals that, by 2 months, V1 already shows adult-like tuning to object features across spatial scales, whereas downstream visual regions show more protracted development, with tuning patterns that remain distinct from adults at 9 months. Finally, by isolating spatial frequencies with bandpass filtering and tracking model-brain correspondence across training, I reveal distinct inductive biases in minds and machines: DNNs rapidly favour fine visual features, or high spatial frequencies, whereas infant extrastriate cortex (V2-V4) shows a protracted preference for shape-based, low-spatial-frequency features. This persistence, beyond what acuity alone would predict, suggests that cortical constraints scaffold early visual learning over and above visual input. Together, these findings place new empirical constraints on theories of how coarse-to-fine object representations develop and demonstrate how NeuroAI can provide a new window into the infant's visual world.