A probabilistic model of eye movements in concept formation

Area of Science:

• Cognitive Science
• Machine Learning
• Psychology

Background:

• Understanding how learners acquire concepts, especially when their beliefs change significantly, is crucial.
• Previous research has not fully clarified the role of optimal experimental design in data selection for such dynamic learning tasks.

Purpose of the Study:

• To develop a principled probabilistic model for concept learning that captures belief development.
• To investigate whether optimal experimental design principles can predict learner behavior, specifically eye movements, in an active concept learning task.

Main Methods:

• Utilized data from an eye-movement version of the Shepard, Horland, and Jenkins concept learning task.
• Introduced a probabilistic concept-learning model to describe belief evolution.
• Employed a sampling function derived from optimal experimental design theory to predict eye movements.

Main Results:

• The developed concept-learning model successfully described belief changes over time.
• A single rational sampling function accurately predicted eye movements at both early (high uncertainty) and late (certainty) stages of learning.
• This suggests a unified mechanism for information seeking across different levels of learner certainty.

Conclusions:

• Optimal experimental design principles can inform models of active learning and evidence acquisition.
• The findings provide a parsimonious explanation for information-seeking behavior in concept learning, applicable across the entire learning trajectory.
• This research bridges computational modeling with empirical data on human learning and decision-making.