Humans learn generalizable representations through efficient coding.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning Theory

Background:

• Reinforcement learning (RL) theory posits that behavior is driven by reward maximization.
• Traditional RL models struggle to explain how humans generalize learning to novel situations.
• A gap exists in understanding the representational mechanisms underlying human generalization.

Purpose of the Study:

• To refine the classical reinforcement learning framework by integrating efficient coding principles.
• To develop a computational model that explains human generalization by emphasizing simple representations.
• To test whether efficient coding enhances RL's ability to predict human generalization.

Main Methods:

• Proposed a novel RL framework incorporating efficient coding (maximizing reward with minimal representations).
• The framework predicts distillation of stimuli into abstract states and utilization of rewarding features.
• Two experiments were conducted to assess human generalization performance against model predictions.

Main Results:

• Models integrating efficient coding achieved human-level performance in generalization tasks.
• Conventional RL models demonstrated significant limitations in explaining observed human generalization.
• The refined framework successfully predicted how complex stimuli are mapped to compact representations.

Conclusions:

• Efficient coding, when combined with RL, provides a more robust framework for understanding human learning and generalization.
• The principle of using simple, abstract representations is crucial for intelligent generalization.
• This augmented RL approach offers a more comprehensive computational explanation of human behavior.