Modeling Childhood Visual Development: The Benefits of Low Resolution Training Data on Category Learning
Human visual acuity sharpens significantly during the first several months of life. Visual object learning develops in tandem with visual acuity. There is evidence that the progression of increasing visual acuity in infants benefits visual tasks throughout life. Recent studies have reported benefits in facial recognition using convolutional neural networks (CNNs) when training on blurry images followed by clear images, intended to simulate the early childhood developmental process. We explore the effects of including different levels of blurred images in training data on object class discrimination accuracy. We test the effects of blurry training images on both highly similar and highly distinct sets of object classes. We train CNN models on two datasets, one with visually distinct object classes and one with visually similar classes. We find that for both datasets, training with lower-resolution images allows relatively robust recognition for clearer images, while learning from clear images does not equivalently benefit lower-resolution recognition. The benefits of blur training are more pronounced in when applied to recognition of more distinct objects. Our findings support the utility of learning from sequences of blurred and unblurred images for more robust large scale object recognition.
Charles, William, "Modeling Childhood Visual Development: The Benefits of Low Resolution Training Data on Category Learning" (2020). ETD Collection for Fordham University. AAI28000285.