Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 27·8 min read
Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models
TL;DR
EEG foundation models trained to reconstruct brain signals exhibit spectral bias—they learn low-frequency and aperiodic (non-rhythmic) components better than high-frequency oscillations. Researchers identified and characterized this bias in reconstruction-based EEG models.
✦ Why It Matters
Engineers can adjust EEG model architectures or training objectives to balance frequency learning and improve downstream clinical or neuroscience applications.
Key Takeaways
- Evaluate your EEG models for spectral bias towards low-frequency components.
- Incorporate auxiliary losses targeting high-frequency oscillatory structures.
- Consider using smaller supervised models in low-resource environments.
- Analyze subject identity encoding in your EEG model embeddings.
How It Works
EEG foundation models are trained using reconstruction-based tasks, which focus on predicting EEG signals from their representations. This approach inadvertently emphasizes aperiodic components, leading to a lack of representation for oscillatory components that are essential for understanding brain activity.