Predicting transfer performance: a comparison of competing function learning models.

Area of Science:

• Cognitive Psychology
• Machine Learning Theory
• Human Learning and Memory

Background:

• Function learning and transfer are key cognitive processes.
• Existing models like POLE and EXAM offer different explanations for these phenomena.
• POLE posits learning via prestored linear functions, while EXAM uses exemplar-based association with rule-based transfer.

Purpose of the Study:

• To compare the predictive accuracy of the POLE and EXAM models.
• To investigate how training data characteristics influence model fit and transfer predictions.
• To determine which model better explains human generalization of learned functional relationships.

Main Methods:

• Participants were trained on a function composed of two linear segments with mirror slopes.
• Experiments varied training density (Experiment 1) and segment separation (Experiment 2).
• Individual participant data were used to fit both POLE and EXAM models, and their predictions for transfer to new inputs were evaluated.

Main Results:

• POLE generally provided a better fit to the training data compared to EXAM.
• EXAM demonstrated superior accuracy in predicting and fitting transfer behaviors across both experiments.
• Experiment 2's transfer patterns strongly supported EXAM's predictions over POLE's, despite training conditions favoring POLE.

Conclusions:

• The extrapolation-association (EXAM) model offers a more robust account of human function transfer than the population of linear experts (POLE) model.
• EXAM's ability to capture transfer behavior, particularly in conditions with distinct training segments, highlights the importance of associative learning.
• Findings suggest that while linear rules may be involved, associative mechanisms are crucial for generalizing learned functions.