A Context-Dependent Bayesian Account for Causal-Based Categorization.

Area of Science:

• Cognitive Science
• Psychology
• Artificial Intelligence

Background:

• The causal view of categories posits that categories are defined by features and their causal relationships.
• Bayesian causal models are frequently employed to investigate how causal knowledge influences categorization.
• Categorization within this framework can be approached as either a likelihood or a posterior computation.

Purpose of the Study:

• To investigate whether human categorization relies on likelihood or posterior computations within a Bayesian causal model framework.
• To determine how people utilize causal probabilistic information in category judgments.
• To explore the flexibility of human probability computation in different judgment tasks.

Main Methods:

• Conducted three experiments involving human participants.
• Utilized computational modeling to analyze experimental data.
• Compared participants' judgments against predictions from likelihood and posterior computation models.

Main Results:

• Evidence suggests that human categorization is better explained by posterior computations rather than likelihood computations.
• Participants demonstrated the ability to compute likelihoods in a related task assessing consistency, not category membership.
• Findings indicate that individuals flexibly compute likelihoods or posteriors contingent on the specific task demands.

Conclusions:

• Human categorization decisions are primarily driven by posterior probability calculations.
• Despite previous assumptions, likelihood computations do not fully approximate human category judgments.
• Individuals adapt their probabilistic reasoning strategies based on task context, utilizing causal information effectively.