EEG Foundation Models Show Spectral Bias Against Oscillatory Signals
A new study published on arXiv (2605.26434) reveals that EEG foundation models, pre-trained on large-scale unlabelled EEG data, exhibit a fundamental spectral bias. Despite promising results in data-rich scenarios, these models often underperform smaller supervised models in low-resource settings. The research attributes this to a mismatch between reconstruction-based pretext tasks and the spectral structure of EEG signals, which consist of high-power aperiodic and low-power oscillatory components. Using synthetic EEG inputs, the authors demonstrate that embeddings are biased toward aperiodic components, under-representing oscillatory activity, especially at higher frequencies. Linear probe evaluations on real-world BCI datasets confirm this bias.
Key facts
- Study published on arXiv with ID 2605.26434.
- EEG foundation models pre-trained on large-scale unlabelled EEG data.
- Models fail to outperform smaller supervised models in low-resource settings.
- Attributed to mismatch between reconstruction tasks and EEG spectral structure.
- EEG signals decompose into high-power aperiodic and low-power oscillatory components.
- Embeddings biased to capture aperiodic components, under-representing oscillatory ones.
- Bias particularly affects higher frequency oscillatory components.
- Findings confirmed via linear probe evaluations on real-world BCI datasets.
Entities
Institutions
- arXiv