Learning Representations from EEG Brain Signals

Brain-computer interfaces (BCIs) promise direct communication between neural activity and external devices. Electroencephalography (EEG) offers a non-invasive approach through lightweight, portable hardware. Despite decades of research, however, EEG-based BCIs remain largely confined to laboratories. The core challenge lies not only in hardware limitations but also in extracting robust and meaningful representations from noisy and variable neural signals. Traditional pipelines extract task-specific features that work in controlled settings but not in new situations involving different users, sessions, or tasks. Performance drops by 19% when models trained on one subject are tested on another. This prevents the practical deployment of BCIs in assistive technologies, healthcare, and human-computer interaction. This dissertation demonstrates how deep representation learning can overcome limitations by discovering invariant features directly from data and aligning neural patterns with human-interpretable semantics. Instead of engineering features for each application, we learn general-purpose representations that transfer across subjects, sessions, and tasks. We make three primary contributions: (1) Novel neural architectures including frequency-aware multi-band encoders and subject-conditional transformers that capture both neural patterns and individual differences; (2) One of the first end-to-end systems for open-vocabulary semantic decoding from EEG, achieving BLEU-1 of 42.75% and BERTScore-F of 53.86% by aligning brain representations with language model spaces across 30 subjects; (3) Neural tokenization via vector quantization that converts continuous signals into discrete "brain tokens," enabling efficient downstream models suitable for real-time, edge-deployed applications. The methods have been validated across diverse experimental paradigms and hardware configurations, including consumer-grade dry-electrode devices and high-density clinical systems. This demonstrates the methods' practical viability across different acquisition settings. Comprehensive evaluations of motor imagery, emotion recognition, and semantic decoding tasks show that learned representations outperform handcrafted features and generalize across the variability that has historically limited BCI deployment. This work establishes representation learning as a viable path toward generalizable, interpretable, and deployable BCIs.