Thesis Defense - Computer Science: Shreyas Ajit Keelary

Please join the Computer Science Department for the upcoming thesis defense:

Presenter: Shreyas Ajit Keelary

Thesis title: Beyond Signal Noise: A Framework for Assessing and Correcting Label Noise in EEG Datasets

Abstract: The reliability of machine learning models in electroencephalography (EEG) research is frequently undermined by label noise. In many cases, trial annotations do not accurately reflect the subject’s true cognitive state due to attentional drift or task switching. This thesis presents a comprehensive, end-to-end framework for identifying and correcting this issue. First, it establishes a robust diagnostic methodology that quantifies the nature and extent of label noise. This is done by integrating an ensemble of outlier detection algorithms with model-based data valuation using Data Shapley. The analyses reveal distinct noise profiles in public datasets. In cognitive tasks, noise is systematic and subjectdriven. In motor imagery paradigms, noise is more randomly distributed, but it remains detrimental. Second, to address these findings, this thesis proposes a novel framework that is universal (agnostic to channel and recording length), subject-adaptive, and extensible. The Mixture-of-Experts (MoE) architecture enables automated label correction and includes a formal task hierarchy for further model expansion. The hierarchical system routes EEG signals through an Activity Detector and a Domain Router. It successfully classifies unseen EEG segments with high accuracy for motor imagery and cognitive tasks. These lead to subject-specific specialized fusion experts that combine geometric, spectral, and temporal features. The MoE architecture provides reliable classification performance, making the system a powerful tool for data auditing. Segment-wise relabeling showed that 95% of cognitive EEG trials contained multiple shifting cognitive states, indicating a significant attentional drift. In contrast, motor imagery trials had a consistent cognitive state, with label noise concentrated at trial onset due to carryover effects. By bridging the gap between label noise identification and correction, this work presents a practical methodology to improve the quality, reliability, and validity of EEG-based research.


Committee Members:
Dr. Garima Bajwa (supervisor, committee chair), Dr. Thiago E Alves de Oliveira, Dr. Vijay Mago (York University)

Please contact grad.compsci@lakeheadu.ca for the Zoom link. Everyone is welcome.