Brain-to-Brain Decoder: Leakage-Aware Validation and Interpretable CEBRA Mapping for Dyadic EEG

This proposal is for the ML4SCI GSoC 2026 project “Brain-to-Brain Decoder – Validating Neural Synchrony Patterns in Human Conversation.” The project builds on the 2025 NeuroDyads pilot, which reported 94% accuracy on decoding conversational roles and participant gender with CEBRA on 8 dyads, and extends that work to an expanded dataset of 40+ dyads. My goal is to build a reproducible, leakage-aware analysis pipeline for time-locked dyadic EEG that is not only effective, but scientifically trustworthy. I will focus on four concrete tasks from the project page: implementing stratified cross-validation for dyadic EEG, computing cross-entropy distance metrics between embedding distributions, generating learning curves for stable dyad decoding, and mapping embedding structure back to localization and frequency-band features that contribute to neurotype classification. The final deliverables will include a modular preprocessing and evaluation pipeline, baseline and control experiments, quantitative embedding-comparison tools, sample-size/stability analyses, and interpretable region- and band-level summaries for clinical vs. neurotypical comparison.