Abstract

Humans are born with very low contrast sensitivity, meaning that developing infants experience the world "in a blur". Is this solely a byproduct of maturational processes or is there some functional advantage for beginning life with poor vision? We explore whether reduced visual acuity as a consequence of low contrast sensitivity facilitates the acquisition of basic-level visual categories and, if so, whether this advantage also enhances subordinate-level category learning as visual acuity improves. Using convolutional neural networks (CNNs) and the ecoset dataset to simulate basic-level category learning, we manipulated model training curricula along three dimensions: presence of blurred inputs early in training, rate of blur removal over time, and grayscale versus color inputs. We found that a training regimen where blur starts high and is gradually reduced over time - as in human development - improves basic-level categorization performance relative to a regimen in which non-blurred inputs are used throughout. However, this pattern was observed only when grayscale images were used (analogous to the low sensitivity to color infants experience during early development). Importantly, the observed improvements in basic-level performance generalized to subordinate-level categorization as well: when models were fine-tuned on a dataset including subordinate-level categories (ImageNet), we found that models initially trained with blurred inputs showed a greater performance benefit than models trained solely on non-blurred inputs. Consistent with several other recent studies, we conclude that poor visual acuity in human newborns confers multiple advantages, including, as demonstrated here, more rapid and accurate acquisition of visual object categories at multiple hierarchical levels. Author SummaryWhy do humans start life with extremely poor vision? The common evolutionary story is that head size is small to accommodate the development of human-level intelligence. However, there is growing evidence that beginning life in a premature state confers short-term advantages. The "starting small" principle states that learning can be facilitated by restricted or impoverished inputs that reduce the learning complexity. We suggest that blurred vision during early development biases learning toward shape features, which organize objects naturally into "basic-level" categories that are the foundation of human cognition (e.g., birds, cars, etc.). Such categories are defined by their visual similarity in global object shape. Since blurring restricts visual inputs to emphasize shape information, it should enhance the learning of basic-level categories. We trained artificial neural-network models on basic-level categorization using either blurred or non-blurred color or grayscale images and found that models trained with blurred images outperformed models trained with non-blurred images, but only for grayscale. These same models performed better in subsequent "subordinate-level" tasks that required discriminating between objects within a basic-level category. Our simulations provide evidence that initially poor vision in infants has an important functional role in organizing knowledge about complex environments.