The contribution of high- and low-level image characteristics to infant categorization

• In Event: Poster Session 11
  In Poster Session: PS 11 Topic - Biological Processes: Psychophysiology

Abstract

Categorization is an important mechanism to structure experiences, but little is known regarding the contribution of high- and low-level image characteristics to visual categorization. Low-level cues can be diagnostic for category membership: For example, round and blurry shapes are typical of animate beings, whereas right angles and straight lines are associated with inanimate objects (Oliva & Torralba, 2003). Prior studies indicate that for familiar categories, low-level characteristics contribute to categorization but do not explain it: Categorization of stimuli containing only low-level cues is reduced compared to categorization of original stimuli in infants and adults (including e.g. faces, houses and cars; de Heering & Rossion, 2015; Peykarjou, Hoehl, Rossion, & Pauen, 2020). However, it remains unclear how low-level image characteristics contribute to categorization when first encountering a new category.
To investigate this question, we conducted a Fast Periodic Visual Stimulation (FPVS; Rossion, 2014) study with 7-month-old infants (N = 20). Fan-like stimuli differing in the shape of individual parts (angular vs. round) and colour (red vs. blue) stimuli were presented at a fixed rate of 6 Hz (6 images/second). Corresponding control stimuli were created by phase-scrambling the power-spectra of images (Sadr & Sinha, 2004). This procedure preserves low-level characteristics such as the presence of lines and edges, but disrupts stimulus shape. Angular/red stimuli were presented as standards, and at every 5th position, a round/blue oddball item was introduced, corresponding to a frequency of 1.2 Hz (1.2 times/second; Figure 1). Recent work indicates that infants are able to categorize original images even without prior experience (Pauen & Peykarjou, submitted to Developmental Science), but the role of high- and low-level characteristics to this categorization response remains obscure. Original and phase-scrambled trials were presented within-subjects. In addition, re-test reliability of categorization responses was assessed following a two-weeks interval.
Categorization responses were observed at the harmonics 2-16 of 1.2 Hs over the visual cortex at both measurement times (centred on electrodes O1, O2, Oz; Figure 2). For original images, strong categorization was observed (signal-to-noise ratio (SNRs) range 1.21-1.30, Z-scores > 1.64) and a smaller and non-significant response for phase-scrambled trial versions (SNRs range 1.09 – 1.16, Z-scores < 1.64). A Bayesian rmANOVA confirmed that responses to original trials including both high- and low-level visual cues were stronger than to phase-scrambled trials, BF = 2.94 The number of participants showing a significant categorization response was higher for original than phase-scrambled images, but half the sample also categorized the control images (16/20 vs. 10/20, Z = 2.45). Data collection to run re-test analyses is still ongoing but should be completed by January 2021. Cronbach’s alpha for individual electrodes will be calculated.
These results indicate that while some infants are able to categorize unfamiliar stimuli based only on low-level image characteristics, their performance is increased when a Gestalt is added to the visual input. Although the statistical regularities omnipresent in our natural environment may help infants to deduce the categorical structure of entities, they seem to be equipped to group objects based on their shape.

Authors

• Stefanie Peykarjou, Heidelberg University, Germany
  Presenting Author

• Sabina M Pauen, Ruprecht-Karls-Universität Heidelberg
  Non-Presenting Author