Per
Babies’ visual experiences are distinctly characterized by simple, high-contrast patterns, as revealed by research using head-mounted cameras to document visual information from young babies’ daily lives. This early visual “diet” is crucial for the development of human vision and has implications for resolving visual impairments and training AI visual systems.
Research shows that babies primarily see simple, high-contrast patterns, forming an essential visual foundation for later development and influencing both human vision and AI training.
What do babies see? What do they look at? The answers to these questions are very different for younger babies than for older babies, children and adults. Characterized by some high-contrast edges in simple patterns, these early scenes also contain the materials necessary to build a solid foundation for human vision.
This is the conclusion of a new study recently published in 
To see what babies see and look at, Smith’s Lab placed cameras on babies’ heads to wear at home during activities of daily living. Credit: Photo courtesy of Indiana University Department of Psychological and Brain Sciences
Previous laboratory and clinical studies have shown that babies prefer to look at simple, high-contrast scenes with large black stripes and checkerboards. The present study is the first to ask the extent to which these preferences constitute their contribution to everyday life. “To see what babies see and look at,” says Anderson, a former postdoctoral researcher in Smith’s Cognitive Development Lab, she and her colleagues put cameras on babies’ heads to use at home during play activities. daily life.
“You can buy newborn ‘flash cards’ that show these simple, high-contrast images,” she explains. “What the front camera videos show, what this work shows, is that babies encounter these types of images all around them in their daily lives, just by looking at things like lights and ceiling corners.”
Linda Smith is a distinguished professor in the Department of Psychological and Brain Sciences at Indiana University. Credit: Photo courtesy of Indiana University Department of Psychological and Brain Sciences
“What we discovered was a very special early ‘diet’ for visual development,” adds Smith. “As with nutrition, young babies don’t start out with rich, complex meals or pizza, but rather simple, developmentally specific foods.”
Previous work has recognized the critical nature of this early period for the future development of human vision. For example, it has been shown that babies who are born with visual abnormalities such as cataracts, or those who live in orphanages with limited visual experiences, experience lifelong visual impairments. The present study offers some preliminary data to address these shortcomings. It also has important implications for creating visual AI systems, which also acquire stronger visual skills when training begins with the same simple, high-contrast visual content.
“The enormous scale of information in daily life”
To identify the properties of visual input in infants approximately three to 13 months of age, researchers placed head-mounted video cameras on 10 infants and 10 of their adult caregivers, collecting and analyzing 70 hours of visual documentation of everyday life at home. . Clear differences emerge between the content of infants’ and adults’ images, with a greater concentration of simple patterns and high-contrast edges in infants’ views than in adults’ views.
Smith infers that the reason for these views is not just that children turn their heads to look at features of the world they can see, but also because parents or caregivers are likely to place them in places where they like to look at things. “You have to think about why they are where they are. There is probably some implicit natural knowledge on the part of parents to leave babies where they like to look at things. Mom won’t bother you if you’re not fussy,” she notes.
However, is this small group of participants from Bloomington, Indiana representative of children more broadly around the world? To answer this question, Smith’s lab conducted the same experiment with a collaborator in a small, populated fishing village in Chennai, India, where electricity is minimal and much of daily life takes place outdoors. And although the front-facing camera images of the 6- and 12-month-olds look very different from their Bloomington counterparts, the youngest babies share a common “diet” of high-contrast outlines and simple patterns in both Chennai and Bloomington.
Bigger images, past and future
Smith and his collaborators also showed that the same sequence of images improves the training of visual AI systems. In a follow-up to the current study, published in the 2023 Neural Information Processing Systems Conference Proceedings, they found that if you train an AI system by first feeding it images characteristic of early childhood, it will have greater success learning to identify visual images than if you feed it images in a random developmental order or simply provide images typical of an adult’s daily life. The most precise development sequence produced the best results.
His work opens new avenues for evolutionary speculation. As Smith explains: “One of the things I always asked when I was a graduate student – and maybe we’re getting the opportunity to answer – is why human babies have such slow motor development. They spend about three months just listening and looking and another six months with some posture and head control. Why are they so slow? The horses come out and run.”
This research suggests that “over evolutionary time, these slow, incremental, optimized tendencies work to build a very intelligent visual and auditory system,” she says. “That’s a story that could be told.”
Meanwhile, his work raises new questions about the visual content of early childhood and its role in the development of the visual system, whether human or AI.
Reference: “An Extreme Simplicity Bias in Visual Input for Infants” by Erin M. Anderson, T. Rowan Candy, Jason M. Gold, and Linda B. Smith, May 10, 2024, Science Advances.
DOI: 10.1126/sciadv.adj8571
The study was funded by
” data-gt-translate-attributes=”[{“attribute”:”data-cmtooltip”, “format”:”html”}]” tabindex=”0″ role=”link”>National Institutes of Health and the US National Science Foundation.
