Tests of the ratio rule in categorization

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Machine Learning

Background:

• Learning and memory theories often output psychological magnitude terms.
• The ratio rule is a common assumption for translating these magnitudes into choice probabilities.
• Existing models often rely on the constant-ratio rule or choice axiom.

Purpose of the Study:

• To test the validity of the ratio rule in categorization tasks.
• To evaluate a connectionist model's ability to account for categorization data.
• To investigate the underlying choice process in categorical decisions.

Main Methods:

• Conducted two categorization experiments using artificial, visual stimuli.
• Stimuli were prototype-structured, constructed from symbols on a grid.
• Assessed if magnitude terms were univariate functions of category-appropriate symbols.

Main Results:

• The ratio rule was found to be incorrect under the study's assumptions.
• A connectionist winner-take-all model successfully accounted for the experimental data.
• The successful model assumes a Thurstonian choice process with non-double exponential noise.

Conclusions:

• The ratio rule is not universally applicable for predicting choice probabilities in categorization.
• Connectionist models incorporating Thurstonian choice processes offer a viable alternative.
• The nature of the noise distribution in choice processes is critical for accurate modeling.