This project develops an explainable machine learning framework for cognitive neuroscience. The goal is to learn complex relationships between neural signals and cognitive behavior, then explain model decisions in a way that supports scientific interpretation rather than only predictive performance.

Framework

The current workflow uses random forests to model nonlinear relationships between neural features and behavioral outcomes. We then use Shapley-value-based explanations to quantify how each feature contributes to model predictions at the individual and population levels.

Interpretability

The explanation layer follows the Shapley-value formulation described in Interpretable Machine Learning by Christoph Molnar. In this project, that framework is used to identify which neural features are most informative for prediction and how those contributions vary across behavioral states or task conditions.

Current application

One target application is visual encoding during natural eye movements. In the saccade-related evoked potential work, the framework is intended to connect multichannel neural dynamics to cognitive behavior while preserving an interpretable account of feature importance.

Next steps

• Expand the feature space to include temporal, spectral, and cross-channel interaction terms.
• Compare random forests with other nonlinear models while keeping the explanation layer scientifically interpretable.
• Use Shapley profiles to identify candidate neural markers for closed-loop cognitive neurotechnology.