Differentiating between Bayesian parameter learning and structure learning based on behavioural and pupil measures.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Predictive processing distinguishes between parameter learning and structure learning.
• Parameter learning updates model parameters, while structure learning alters model causality or parameters.
• Empirical differentiation between these learning types is lacking.

Purpose of the Study:

• To empirically differentiate parameter learning and structure learning.
• To investigate the effects of these learning types on pupil dilation.

Main Methods:

• A within-subject, computer-based learning experiment with two phases.
• Phase 1: Learning cue-target stimulus relationships.
• Phase 2: Learning a conditional change in the established relationship.

Main Results:

• Learning dynamics differed qualitatively between the two experimental phases.
• Learning was more gradual in the second phase compared to the first.
• Results suggest structure learning in phase 1 and parameter learning in phase 2.

Conclusions:

• Pupil dilation can potentially differentiate between parameter and structure learning.
• Structure learning may involve building multiple models, whereas parameter learning updates existing ones.
• The observed learning dynamics support distinct empirical signatures for parameter and structure learning.