Structure learning and the Occam's razor principle: a new view of human function acquisition

Area of Science:

• Cognitive Science
• Psychology
• Machine Learning

Background:

• Humans frequently encounter and learn functional relationships between variables.
• Human responses often align with these learned functional relationships after training.

Purpose of the Study:

• To investigate how humans predict unobserved function segments after training.
• To compare human function prediction with various computational models.
• To understand the cognitive mechanisms underlying human function abstraction.

Main Methods:

• Participants predicted function segments based on observed data.
• Human predictions were compared against predictions from computational function-learning models.
• A hierarchical Bayesian framework modeled human learning at function and parameter levels.
• Generating functions were surreptitiously changed over time to analyze learning dynamics.

Main Results:

• Human performance was best predicted by polynomial functions.
• Participants could often explicitly report the generating function.
• The hierarchical Bayesian model successfully analyzed learning time courses.
• Humans acquired new functions effectively, even with parameter shifts.
• Simpler functions were selected more frequently and learned faster.

Conclusions:

• Human function learning demonstrates abstraction capabilities.
• Function learning aligns with Occam's razor principle, favoring simplicity.
• Humans efficiently learn and adapt to changing functional relationships.