Decoding Preverbal Infants’ Visual Processing from Eye Movement Patterns

• In Event: Poster Session 03
  In Poster Session: PS 03 Topic - Perceptual, Sensory, Motor

Abstract

Eye-tracking techniques help us understand cognitive processes by providing rich oculomotor information. Thereby, it has become one of the most popular research methods to understand the cognitive capacities of infants, who have limited verbal and motor abilities. However, the cognitive nature of infants’ movements is still unclear. For example, researchers have not reached a consensus on what type of eye movement represents infants’ face perception abilities. To this end, the current study explored the relations between infants’ eye movements and their underlying cognitive processes by using data-driven machine learning approaches.

Machine learning is a powerful method that seeks out complex relations within multi-dimensional data. In particular, these methods can analyze the intrinsic structures between a large number of features (e.g., spatial and temporal characteristics of eye movements). Machine learning has recently been used to decode adults’ and children’s high-level perceptual processing from their eye movement patterns. Unlike adults’ and children’s eye-tracking quality, infants’ eye movement recordings are much noisier and are often interfered with by irrelevant factors, such as body movements. Thus, it is unclear whether machine learning can be successfully applied to understanding the nature of infant’ eye movements. To explore this possibility, we tried to decode infants’ visual perception of faces vs. vases.

Forty infants (224-375 days, 19 females) participated in this study, in which they watched images of faces and vases while their eye movements were recorded. The features used to classify face vs. vase perception were: mean fixation duration, number of fixations, percent of time fixating on face or vase regions, mean saccade amplitude, mean saccade distance, and the percent of the fovea covered by the image, face regions, and vase regions. The eye movement features were averaged within each trial and pooled across participants. To build the machine learning models, we iteratively selected eye movement data from 80% of the trials as the training set. On each iteration, we used the remaining 20% to assess model performance in classifying infants’ percept of faces vs. vase (i.e., 5-fold cross-validation).

We used two machine learning algorithms: Random Forests (RF) and k-Nearest Neighbors (k-NN) to classify each trial and assess the contribution of the eye movement features. RF measured the contribution of each feature in the classification. k-NN assessed which combinations of features achieve the highest decoding accuracy.

The results showed that the RF model achieved 60.00% accuracy and the k-NN model achieved 63.06% accuracy. Both models’ performance was above chance (50%, ps < .001). Moreover, there was a high degree of overlap between important features chosen by the two models (Figures 1 & 2). These results indicate that machine learning can be used to reliably predict the type of stimuli from infants’ eye movements. Furthermore, the consensus of the features selected by the two models suggests a reliable relation between infants’ visual perception and eye movement patterns, which can be uncovered by machine learning methods. By establishing the methodological feasibility, this study paves the foundation to better understand the cognitive nature of infants’ eye movements.

Authors

• Anna Herbolzheimer, Princeton University
  Presenting Author

• Lauren L. Emberson, Princeton University
  Non-Presenting Author

• Naiqi G Xiao, McMaster University
  Non-Presenting Author