Sampling assumptions in inductive generalization.

Area of Science:

• Cognitive Science
• Psychology
• Machine Learning

Background:

• Inductive generalization is a fundamental cognitive ability essential for learning, categorization, and decision-making.
• Existing models of generalization rely on extreme sampling assumptions: strong sampling (data as positive examples) and weak sampling (unrestricted data generation).

Purpose of the Study:

• To propose a more general account of sampling that incorporates a mixture of strong and weak sampling.
• To investigate how people make sampling assumptions in inductive generalization tasks.
• To explore individual differences and the influence of information structure on sampling biases.

Main Methods:

• Developed a generalized sampling model allowing for a mixture of strong and weak sampling.
• Conducted two experiments involving one-dimensional generalization tasks.
• Analyzed participants' generalization patterns to infer their sampling assumptions.

Main Results:

• Participants' generalization behavior aligns with a mixture of strong and weak sampling assumptions.
• Significant individual differences exist in the relative weighting of strong versus weak sampling.
• The structure of presented information experimentally influences the mixture of sampling assumptions.

Conclusions:

• Human inductive generalization is better explained by a flexible sampling model than by extreme assumptions.
• Understanding the psychological basis of mixed sampling is crucial for richer models of generalization.
• This approach offers potential extensions to more complex inductive reasoning problems.