Data-driven experimental design and model development using Gaussian process with active learning.

Area of Science:

• Cognitive Science
• Machine Learning
• Computational Neuroscience

Background:

• Parametric models in cognitive science rely on strong assumptions, risking model misspecification and biased inference.
• Efficient data collection and robust inference are crucial for advancing computational modeling of cognition and behavior.

Purpose of the Study:

• To propose a data-driven, nonparametric framework for cognitive model development that integrates optimal experimental design.
• To introduce Gaussian Processes with Active Learning (GPAL) as an extension of existing adaptive design optimization (ADO) frameworks.

Main Methods:

• GPAL combines Gaussian Processes (GP) for flexible modeling with active learning for iterative model fitting and experimental design optimization.
• The framework was applied to a delay discounting task and compared against the parametric ADO framework in two experiments.

Main Results:

• GPAL demonstrated high sensitivity to individual differences, identifying novel patterns overlooked by the model-constrained ADO.
• The results validate GPAL as a viable cognitive modeling framework capable of capturing nuanced data patterns.

Conclusions:

• GPAL offers a data-driven approach that balances raw data interpretability with the assumptions of parametric models.
• This framework represents a significant step towards more flexible and sensitive cognitive modeling.