Eye pupils mirror information divergence in approximate inference.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Information Theory

Background:

• Pupil dilation is a physiological measure linked to cognitive arousal and various mental processes.
• Existing research suggests unifying pupillometry findings using information theory and Bayesian methods.
• Bayesian methods are often computationally intractable for complex, real-world scenarios.

Purpose of the Study:

• To investigate if pupil responses reflect information quantified via approximate inference.
• To explore the relationship between pupil dilation and information-theoretic measures in a dynamic prediction task.
• To assess the utility of approximate inference as a practical framework for pupillometry.

Main Methods:

• Measured pupil diameters in 27 healthy adults performing a number prediction task with changing criteria.
• Employed a variational Bayes method to model individual differences in performance and pupil response.
• Quantified prediction uncertainty using Kullback-Leibler divergence (D_KL) and change point probability using Shannon's surprise (SS).

Main Results:

• Found a covariance between trial-wise pupil dilation and trial-wise D_KL that varied with prediction accuracy.
• Observed that Shannon's surprise (SS) was sensitive to discrete change points in the task.
• Demonstrated a model-based link between pupil responses and information divergence during approximate inference.

Conclusions:

• The pupil-linked arousal system appears to reflect information divergence during approximate inference.
• Pupillometry can serve as a non-invasive measure of cognitive information processing in dynamic environments.
• Approximate inference provides a viable computational framework for interpreting pupillometry data.