Learning How to Generalize.

Area of Science:

• Cognitive Science
• Machine Learning
• Psychology

Background:

• Generalization is crucial for cognitive systems across all domains.
• Understanding how humans learn to generalize in new contexts remains an open question.
• Existing research highlights context-dependent variations in human generalization.

Purpose of the Study:

• To investigate the mechanisms by which humans learn appropriate generalization strategies.
• To propose a mathematical framework for learning how to generalize by learning inductive biases.
• To test the framework's predictions against human generalization behavior.

Main Methods:

• Developed a normative mathematical framework for learning generalization strategies.
• The framework learns inductive biases from the statistical structure of observed features and concepts.
• Compared framework predictions to human behavior in three experiments (perceptual and conceptual domains).

Main Results:

• The proposed framework accurately predicts human generalization behavior across different domains.
• Experimental results support the framework's prediction that ideal learners average or select hypothesis spaces.
• Individual-level data provided specific support for the averaging strategy in one study.

Conclusions:

• The framework offers a principled account of learning how to generalize.
• Human generalization learning involves acquiring domain-specific inductive biases.
• Future research can further explore the interplay between hypothesis spaces and learning strategies.