A Learning Paradigm for Selecting Few Discriminative Stimuli in Eye-Tracking Research

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Eye-tracking quantifies visual processing and shows potential for group recognition, including autism spectrum disorder (ASD).
• Current eye-tracking research is hindered by stimulus heterogeneity and time-consuming data collection due to numerous stimuli.
• Efficient stimulus selection is crucial for practical applications of eye-tracking in group identification.

Purpose of the Study:

• To develop a computationally efficient method for selecting the most informative stimuli in eye-tracking studies.
• To introduce and quantify 'stimulus discrimination ability' for improved group recognition models.
• To validate a novel approach for reducing the number of stimuli required for accurate ASD identification and other group predictions.

Main Methods:

• Mathematical definition of the stimulus selection problem and introduction of 'stimulus discrimination ability'.
• Development of a scanpath-based recognition model incorporating cross-subject entropy and divergence scores.
• Implementation of an iterative learning mechanism with stimulus-wise attention for refining stimulus selection.

Main Results:

• The proposed method significantly reduces the number of stimuli required (10 vs. 220) while maintaining or improving performance.
• Demonstrated superior performance in identifying autism spectrum disorder (ASD) using the selected stimuli.
• Validated the method's effectiveness on a secondary task, gender prediction, confirming its generalizability.

Conclusions:

• The developed method offers a simple, flexible, and efficient approach to stimulus selection in eye-tracking research.
• This technique has the potential to facilitate large-scale autism spectrum disorder (ASD) screening and advance other eye-tracking applications.
• The focus on discriminative stimuli enhances computational efficiency and model performance in group recognition tasks.