Decoding Language from the Cerebellum: An MVPA Approach

Project Overview

This project computationally validates the "Cerebellar Language Network" hypothesis (as discussed in recent literature, e.g., Neuron 2025). Using functional MRI data from the Human Connectome Project (via Nilearn), I implemented a Machine Learning pipeline to determine if cerebellar activity alone is sufficient to decode language processing.

Hypothesis

If the cerebellum contains a dedicated language network mirroring the cortex, a Support Vector Machine (SVM) should be able to classify "Language" vs. "Control" tasks with high accuracy, relying solely on voxels within the cerebellar mask.

Methodology

• Data: Brainomics Localizer (Language vs. Control contrasts).
• Preprocessing: Explicit masking of the Cerebellum (excluding all cortical activity).
• Model: Linear Support Vector Machine (SVM) with Leave-One-Group-Out cross-validation.
• Analysis: Multi-Voxel Pattern Analysis (MVPA) to extract feature weights.

Key Results

• Decoding Accuracy: Mean Decoding Accuracy: 97.50% and Standard Deviation: 10.90%
• Results: Successful decoding! The cerebellum contains distinct information distinguishing language from control.
• Localization: The SVM weight maps reveal distinct clusters in the Right Posterior Cerebellum, consistent with contralateral connectivity to the left-hemisphere language cortex.

Dependencies

• Python 3.x
• Nilearn
• Scikit-learn
• Matplotlib

Results Images