The Developing Infant Creates a Curriculum for Statistical Learning

Smith et al., 2018

Source: Smith et al., 2018


  • What is the nature of the environment that supports learning (in infants)?
  • Data collected from head-cameras and eye-trackers worn by infants in everyday environments help address this question
  • Analyses of these egocentric views has shown systematic changes throughout development, effectively creating a curriculum for learning
  • Links: [ website ]


  • In the first two years of life, infants make significant progress in many domains, e.g. language, vision, and social behaviors
  • Statistical learning methods rely on both the learning machinery and the data on which it operates
    • There’s an assumption that the learning environment is rich but noisy, and therefore the focus has been on learning machinery that can sort through this messy data
  • The infant’s view of the world changes as its sensorimotor abilities develop
    • Start out with limited acuity and locomotion
    • Learning to crawl enables moving towards distant objects
    • Manually playing with objects creates a dataset of multiple views of a single object


  • Analyses of head-camera images revealed distinct changes in visual experience
    • Under three months of age, consists primarily of close, frontal views of faces which dimishes as they reach their first birthday
    • After one year, hands become more prominent, with hands of others generally acting on objects
  • Individuals with congenital cataracts removed as early as four months show permanent deficits in face processing
    • Might be due to missing out on the dataset of close frontal views of faces infants acquire in their first three months
  • Everday learning environments are characterized by a few types of objects appearing very frequently
    • These skewed distributions may help facilitate learning: consistency, bootstrapping, or desirable difficulty


  • Infants create a curriculum for learning
    • After their first year, egocentric views of toddlers differ from 8-10 months as well as adults
    • Toddler visual experiences are shaped by what they can manually do
    • The changing abilities of infants open and close environments for learning
  • Many approaches to statistical learning assume the computational problem, dataset, and learning machinery remain constant, which clearly does not reflect the reality of an infant


  • Data for learning and the learning machinery cannnot be studied separately
    • The changing structure of the egocentric view is key to the robustness of infant learning
Elias Z. Wang
Elias Z. Wang
AI Researcher | PhD Candidate