EEG Foundation Model for Heterogeneous Sensor Layouts Trained with Task-Switch Contrastive Learning

Project Overview and Role

Currently, I am part of a team working on improving the existing BENDR framework for performing self-supervised learning on EEG data. The key limitations of the BENDR approach are that BENDR assumes a fixed electrode configuration and training objectives that ignore the cognitive structure of brain signals. To address this, my team and I proposed incorporating a graph neural network encoder for each dataset to compress multi-channel EEG data, improving robustness across varying channel layouts, and introducing a neuroscience-grounded contrastive loss that groups brain recordings by mental state rather than by time alone. My role in this project has provided me with a more profound knowledge of neural networks and their applications. However, more importantly, I have strengthened my skills in literature review. I was required to rely on research skills and prior knowledge to understand technical papers in a field I was unfamiliar with. Through this, I’ve learned that many scientific breakthroughs are a result of finding new approaches to existing ideas, requiring the ability to interpret current literature. My team and I won the spotlight paper award for best paper and presentation overall at the Canadian Undergraduate Conference of AI (CUCAI) 2026!

View Code (GitHub) Paper