Starting Blurry: The Impact of Initial Sensitivity to Low Spatial Frequencies on Infant Categorization Skills

Humans are born with very low visual acuity, forcing infants to view the world “in a blur”. Gradually, the range of spatial frequencies realized in the retinal map to V1 increases, allowing infants to see their surroundings in more detail. Is there any functional advantage for infants’ initial low perceptual acuity? This study explores whether reduced visual acuity among infants facilitates both learning rate and asymptotic accuracy in the acquisition of basic-level visual categories and, if so, whether such a boost assists in learning subordinate-level categories once visual acuity improves. Using convolutional neural networks (CNNs) to simulate an infant’s visual development, we compared models that differed in terms of rate of decrease in blur, type of image (grayscale versus colored), and number of blurred inputs. Experiments revealed that gradual blur improves basic-level model performance, but only under conditions where the model environment is constrained, much like the restricted environments infants experience during early development. Furthermore, we found that all models initially trained on basic-level categories performed better during subordinate-level categorization training than models trained solely on subordinate-level categories. We conclude that poor visual acuity in human newborns is not an accident - rather it confers an important initial advantage in bootstrapping an infant’s acquisition of visual object categories.