Abstraction and model evaluation in category learning.

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Machine Learning

Background:

• Category learning research often assumes strict prototype or exemplar models.
• Existing models may overlook nuanced abstraction levels, potentially misinterpreting learning processes.
• Previous analyses focused on limited datasets, restricting generalizability.

Purpose of the Study:

• To reanalyze 30 category learning datasets using the Varying Abstraction Model (VAM).
• To investigate the prevalence and evidence of partial abstraction in category learning.
• To compare model evaluation methods, considering both fit and complexity.

Main Methods:

• Reanalysis of 30 published category learning datasets.
• Application of the Varying Abstraction Model (VAM) to assess abstraction levels.
• Model evaluation using both maximal likelihood (fit) and marginal likelihood (complexity).

Main Results:

• The VAM identified evidence for partial abstraction across multiple datasets where extreme models (prototype/exemplar) showed none.
• Results generalize and extend previous findings on partial abstraction to a larger set of studies.
• Complexity differences among VAM family models were found to be small across datasets.

Conclusions:

• The Varying Abstraction Model (VAM) provides a more comprehensive framework for understanding category learning abstraction.
• Evidence for partial abstraction is more widespread than previously acknowledged.
• Computationally straightforward, complexity-insensitive model evaluation is justified for VAM analyses in this context.