Naïve and Robust: Class-Conditional Independence in Human Classification Learning.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Human categorization abilities are sophisticated but computationally demanding.
• Learning novel probabilistic classification tasks presents significant computational challenges.
• The assumption of class-conditional independence simplifies inference and enhances robustness.

Purpose of the Study:

• To investigate the role of class-conditional independence in human category learning.
• To develop and test a Bayesian model (DISC-LM) that learns and adapts to feature independence assumptions.
• To compare model predictions with human classification behavior in experiments.

Main Methods:

• Developed the Dependence-Independence Structure and Category Learning Model (DISC-LM).
• Incorporated varying prior beliefs in class-conditional independence within the model.
• Conducted simulation studies to assess theoretical properties of the model.
• Designed and executed two human experiments using optimal experimental design principles.

Main Results:

• Simulation studies showed DISC-LM adapts effectively to unexpected task structures.
• Human participants' classification decisions were best explained by a model with strong initial belief in class-conditional independence.
• Learners adapted their behavior to the true environmental structure over time.
• The class-conditional independence assumption proved robust across different statistical environments.

Conclusions:

• Class-conditional independence may serve as a powerful default assumption in category learning.
• Human category learning appears to leverage this simplifying assumption initially.
• Learners can adapt their reliance on this assumption based on environmental feedback.
• The DISC-LM provides a computational framework for understanding adaptive category learning.