Poster #10 - Number perception and statistical learning: Investigating how children develop number sense

• In Event: Poster Session 19

Abstract

Human newborns, young children and even many non-human animals can perceive number, through what is often termed their Approximate Number System (ANS) suggesting that the ANS is innate and domain-specific (Feigenson et al., 2004; Howard et al., 2019; Xu & Spelke, 2000). Recent machine learning models have challenged this view, suggesting that number perception might instead emerge through gradual experience with natural scenes (i.e., that the ANS is not innate) and domain-general mechanisms like statistical learning (i.e., that the ANS is not domain-specific; Stoianov & Zorzi, 2012; Testolin et al., 2020). These conflicting views presents an important theoretical and empirical challenge about the nature of number perception.
To test the proposal that number perception emerges through gradual experience and statistical learning, we explored a prominent statistical feature of natural scenes: a low number prior (i.e., objects often occur in sets less than 10 and decrease at a power-function-like rate for higher numbers; Piantadosi, 2016). If human’s number perception emerges through experience with natural scenes and statistical learning, then we should expect most natural environments to contain few objects. Additionally, this expectation should become stronger with experience and age. To test this expectation, we embedded typical number stimuli with perceptual noise (e.g., reducing the contrast). Given efficient-coding theories (Barlow, 1961), the more uncertain humans are when estimating the number of objects, the more they should rely on past experiences and thus make an educated guess of lower numbers.
Younger (five and six-year-olds) and older (seven and eight-year-olds) children (N = 80) completed a number discrimination task in which participants were asked to verbally report which of two sides contained more items (Fig. 1). Perceptual noise was introduced by systematically reducing the contrast of one of the two side-by-side Gabor dot arrays. Additionally, the ratio between the two dot arrays varied (e.g., 3.0, 1.5, 1.25, and 1.06), making certain trials more difficult than others.
A two-way ANOVA revealed that there was a statistically significant effect of contrast on number perception (p < 0.001), consistent with the idea of a low-number prior. However, we found no statistically significant interaction between the effects of Contrast and Age (F(2, 78) = 0.25, p = 0.62; Fig 2). Our results suggest that human children possess a low-number prior by age five. However, the lack of differences among age groups suggests that this low number prior does not become stronger with more experience with natural scenes.
Despite recent research in machine learning, the current findings suggest that human learners may not utilize the same underlying mechanisms (e.g., the role of experience and statistical learning) for number perception that deep neural network models employ.

Authors

• Miranda N Long, University of British Columbia (UBC)
  Presenting Author

• Darko Odic, University of British Columbia (UBC)