Transferring structural knowledge across cognitive maps in humans and models.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Task performance often relies on recognizing underlying structures like hierarchies or periodicities.
• Transferring this structural knowledge to new situations requires flexible, generalizable representations.
• Current methods struggle with adapting structural knowledge across different environmental specifics.

Purpose of the Study:

• To propose and validate a model where humans represent structural forms as abstract basis sets.
• To investigate how transferring these basis sets facilitates learning in novel tasks.
• To demonstrate that this representation enables inference of hidden task states and policies.

Main Methods:

• Development of a computational model simulating abstract basis set representation.
• Experimental design involving participants learning abstract graph structures over two days.
• Analysis of how Day-1 structural knowledge impacts Day-2 task performance and inference.

Main Results:

• The computational model successfully inferred underlying structures, task states, and policies.
• Participants with correct structural priors on Day 1 showed improved performance on Day 2.
• Participants demonstrated an ability to infer unobserved state-trajectories and effective policies.

Conclusions:

• Abstract basis sets provide a flexible framework for representing and transferring structural knowledge.
• This representation is crucial for generalizing learning to novel environments and inferring hidden task dynamics.
• The findings support a cognitive mechanism for abstract structure learning and transfer in humans.