The Effect of Low Acuity in Infancy on Developing Ventral Visual Stream Representations

The seemingly deficient blurry beginnings of infant visual experience have been hypothesised to have adaptive value. Initial low acuity could serve as an inductive bias strengthening more holistic representations, contributing to the robustness of the developed visual system. Computational models of visual recognition using deep neural networks (DNNs) have supported this in principle, showing that training DNNs initially with blurred images broadens their receptive fields and improves face recognition performance. However, whether low acuity actually affects developing representations in humans has not been tested. We therefore characterised representational geometry in the ventral visual stream of infants using functional MRI (fMRI), and compared it to DNNs trained with various levels of smoothing. fMRI was acquired in awake 2- and 9-month-old infants (N=134) and adults (N=24) as they viewed 36 looming pictures, comprising 3 examples of 12 categories. To model the effect of initial low-acuity, we trained DNNs (ResNet-50) on 1000 categories of objects (ImageNet), with six DNNs receiving images smoothed to different degrees (gaussian blur with sigma ranging from 0-6 pixels). Representational similarity analysis was used to compare the representational geometry of each layer of the DNNs with that of the early and late regions of the ventral visual stream. For both regions, the infant groups and adults shared a considerable portion of representational geometry with the DNNs (Spearman r~0.2-0.5). The degree of image smoothing during DNN training substantially modulated the correspondence while the degree of smoothing during DNN inference had little effect. At 2-months, representational geometry was more similar to DNNs trained with greater smoothing compared to older age groups, supporting the hypothesis that developing ventral visual representations are shaped by initial low acuity. More generally, we demonstrate that awake infant fMRI and computational modelling synergise to provide unique insights into the developmental origins of the visual system.