Surprise-minimization as a solution to the structural credit assignment problem.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• The structural credit assignment problem challenges agents to infer hidden causal relationships between actions and outcomes.
• Understanding this problem is crucial for explaining goal-directed behavior and learning in humans.

Purpose of the Study:

• To investigate how the human brain solves the structural credit assignment problem.
• To test a computational model of action selection based on surprise minimization.

Main Methods:

• Recorded behavioral and electrophysiological data from human participants performing a novel bandit task with hidden action-outcome mappings.
• Developed and applied a computational model formalizing action selection via competition between structural representations.
• Utilized single-trial latent-variable analysis to link neural patterns to model predictions.

Main Results:

• Human behavior demonstrated clear evidence of credit assignment and learning despite uninstructed, hidden task structures.
• The computational model successfully accounted for participant data by minimizing surprise.
• Neural activity patterns quantitatively supported the surprise minimization mechanism and credit assignment predictions.

Conclusions:

• The human brain employs surprise minimization to arbitrate between competing structural representations for credit assignment.
• Neural activity reflects not only reinforcement learning but also the core mechanisms of credit assignment and behavioral control.
• Findings support computational models of decision-making and provide insights into the neural basis of learning complex tasks.