Abstract
The hierarchical organization of the visual system has been supported by studies using convolutional neural networks (CNNs) to decode fMRI data from human adults. Along visual processing streams there is a transition from low-level features such as edges and luminance, represented in earlier model layers, to mid-level invariant features such as shapes and parts, represented in middle model layers, to high-level object identity and category, represented in later model layers. How and when the human visual system comes to be organized this way over development is unclear, partly because of the difficulty of conducting comparable fMRI studies early in development. Here we report our approach for early developmental fMRI based on adult-grade acquisition and analysis methods. Two cohorts of infants/toddlers (N=5 range: 3–8 months; N=8 range: 8–36 months), and an adult comparison group (N=8), watched a short but engaging cartoon during fMRI. Overall, we observed surprising similarity between the brain responses of infants and adults, particularly in visual cortex. To decode the specific contents of the infant brain, we used a CNN to represent the movie in terms of low-, mid-, and high-level content. Converging model-based analyses using univariate regression and representational similarity showed that all levels of content explained similar variance in early visual cortex of infants and adults. However, a developmental difference emerged in later visual areas: in adults, high-level content was better represented than low- and mid-level content and than high-level content in early visual cortex; these differences were not observed in infants, despite the fact that they exhibited reliable visual activity relative to a control auditory region. Additional steps were taken to control for the impact of differences in anatomical alignment and data quality. These findings provide an initial foray into early developmental changes in the large-scale functionality and selectivity of human visual cortex.