TRANSFORMER MODELS FOR ROBUST EEG SOURCE LOCALIZATION

Abstract

Electroencephalography (EEG) measures brain activity using electrodes placed on the scalp. Estimating the underlying source locations from these measurements is an ill-posed inverse problem, and traditional methods often struggle in the presence of noise and complex source configurations. Recent approaches have applied neural networks to this task, but many rely on assumptions that limit their ability to generalize across different data settings. This work presents a transformer-based model for EEG source localization that supports both single-source and multi-source prediction. The model is designed to handle challenges commonly found in EEG data, including measurement noise, varying source configurations, and missing electrodes. Training is performed on large-scale simulated EEG data generated from known source locations. A structured evaluation framework is introduced to assess performance under different conditions, including varying noise levels, source region densities, and electrode dropout scenarios. Across these settings, the transformer model consistently achieves lower localization error compared to both classical methods and existing neural network approaches. In addition, a set of ablation studies is conducted to analyze the impact of architectural and training design choices. The code and data used in this work will be made publicly available.