Proceedings of the Annual Meeting of the Cognitive Science Society

Electroencephalography (EEG) provides critical insights into brain activity, yet its inherent variability and nonstationary nature pose significant challenges for computational analysis, particularly in cross-subject generalization tasks. We present SIESTA (Spectral Invariant EEG-based Semi-causal Transform Architecture), a novel EEG foundation model that addresses these challenges through three key innovations: (1) VQGAN-based spectral tokenization capturing wavelet representation of EEG; (2) a dual-stream Transformer architecture pre-trained using a semi-causal generative modeling approach; and (3) Contrastive Invariant Fine-Tuning (CIFT), a label-free domain adaptation strategy that aligns feature distributions across subjects by integrating spectral-temporal dynamics. Pre-trained on over 32,900 hours of diverse EEG data, SIESTA achieves state-of-the-art performance in epilepsy monitoring, improves F1-score by $12.4 \%$ on scalp EEG and $8.7 \%$ on intracranial EEG, respectively. Beyond epilepsy, SIESTA demonstrates strong generalizability to non-epilepsy tasks, including motor imagery and sleep stage classification. These results validate that spectrotemporal integration and domain-invariant learning are fundamental for modeling cross-subject EEG variability, establishing new benchmarks for robust brain-computing systems.