Tracking human skill learning with a hierarchical Bayesian sequence model.

Area of Science:

• Cognitive Science
• Neuroscience
• Machine Learning

Background:

• Implicit learning allows humans to acquire complex perceptuo-motor skills through extensive practice.
• Skill acquisition likely involves leveraging increasingly complex predictive relationships in the environment.

Purpose of the Study:

• To develop a novel computational model for characterizing implicit sequence learning.
• To investigate how humans adapt their internal predictions based on practice and environmental dependencies.

Main Methods:

• Fit a hierarchical Bayesian sequence model to reaction time data from a serial reaction time task.
• The model uses a forgetful, trial-by-trial update mechanism, integrating information from variable-length past event windows.
• Analyzed how model parameters, reflecting prediction strategies, changed across ten sessions of practice.

Main Results:

• Participants initially relied on trigrams (two preceding elements) to predict sequences, adapting to higher-order structures.
• Reliance on local statistical fluctuations decreased with practice, indicating reduced forgetting and improved skill.
• Some participants later utilized longer context windows (more than two elements), and demonstrated resistance to sequence changes.

Conclusions:

• The Bayesian model provides a principled account of adaptive complexity in human sequence representations during skill learning.
• Learning involves dynamic adjustments in prediction window depth, influenced by practice and individual differences.
• The findings offer insights into the mechanisms underlying long-term implicit skill acquisition and memory.