Habituation Reflects Optimal Exploration Over Noisy Perceptual Samples

Area of Science:

• Cognitive Science
• Developmental Psychology
• Computational Neuroscience

Background:

• Human visual attention is fundamental from birth, yet the underlying decision-making mechanisms are not fully understood.
• Existing models often lack a comprehensive explanation for dynamic looking time patterns like habituation and dishabituation.

Purpose of the Study:

• To introduce the Rational Action, Noisy Choice for Habituation (RANCH) model, a novel computational framework for visual attention.
• To elucidate the role of rational learning and perceptual noise in guiding visual sampling decisions.
• To validate the RANCH model against empirical data and compare it with alternative theoretical approaches.

Main Methods:

• Developed the RANCH model, a rational learning framework incorporating noisy perceptual sampling and expected information gain (EIG) for decision-making.
• Collected adult looking time data using a paradigm analogous to infant habituation studies.
• Compared RANCH model predictions against baseline models (no learning, no noise) and alternative linking hypotheses (Surprisal, KL divergence).

Main Results:

• Demonstrated that both learning and perceptual noise are critical components for accurately modeling looking time behavior.
• Found that Surprisal and KL divergence serve as effective proxies for EIG within the specific learning context of the study.
• The RANCH model successfully replicated key patterns of habituation and dishabituation observed in empirical data.

Conclusions:

• The RANCH model provides a robust computational account of visual attention and decision-making, integrating learning and perception.
• Perceptual noise and the drive to maximize information gain are essential for understanding habituation and dishabituation.
• The findings support the utility of EIG as a core principle in rational models of attention and suggest practical alternatives for its estimation.