The developing visual system biases the statistics of the infant visual environment

• In Event: How do the statistics of the environment shape the infant mind across domains?

Abstract

Continuous visual input in the first months of life – and the low-level statistics that emerge – are necessary contributors to developing a typical visual system (Wiesel & Hubel, 1963). The visual statistics of early input are based on both the external input to the retina (e.g., what views are in front of infants’ faces and eyes) and how this input is processed by an immature visual system before it reaches the primary visual area of the cortex (V1). Recent research shows how immature visual processing could produce better training data for basic visual tasks (Vogelsang et al., 2018), with the low contrast sensitivity of the newborn fovea “filtering out” fine edges to make input simpler. Although in-lab studies (e.g., Atkinson et al., 1977) show that infants under 3 months also prefer to look at stimuli that reflect these internal constraints (simple edges with lower spatial frequencies), an underlying assumption of everyday vision is that, because it is massive and spans all waking moments, the overall statistics of external input will be more or less similar across ages. The implication of this is that retinal constraints alone must do the work of filtering out irrelevant information to train the developing visual system.
In this talk, we show that in fact, external input also differs by age in line with the constraints of the internal system. Using lightweight head-mounted cameras, we recorded 28,000+ views of daily life from 1-to-3-month-old infants, sampling images at every 5 seconds (7.8 hours total across 10 infants). Because the cameras capture what is in front of infants’ faces, these images approximate the external input into the retina. For comparison, we recorded a similar number of scenes from a group of 6-to-8-month-olds, who have more mature visual systems and greater contrast sensitivity. We analyzed these images for low-level properties that correspond to early changes in visual sensitivity (amount of contrast and low spatial frequencies) and in terms of edges (amount and their orientations), which are the basic unit of representation in V1. We then used machine learning to classify images into typical clusters of co-occurring properties, allowing us to examine the structure of the input in time.
The external input generated by 1-to-3-month-old infants is typically simpler (i.e. has fewer edges and competing orientations) and more dominated by low spatial frequencies than the input generated by 6-to-8-month-olds, and 1-to-3-month-olds attend to these simple images for longer continuous periods than older infants do (Figure 1). Additionally, using image processing to approximate the typical 2-month-old visual system, we show that low contrast sensitivity would not meaningfully alter the visual structure (amount of edges and their orientation) or the temporal structure of 1-to-3-month-olds’ input (Figure 2). In other words, young infants seem to select their external input in line with their internal constraints. By characterizing properties of real-world input in early infancy, this work is the first to show that training a visual system is unlikely to depend solely on internal constraints but also on the views that infants themselves generate.

Authors

• Erin Morgan Anderson, Indiana University - Bloomington
  Presenting Author

• Linda B. Smith, Indiana University - Bloomington