A new framework for brain signal models

ScaleBrain uses multi-scale representations to improve how electroencephalography signals are processed.

Researchers developed ScaleBrain, a scale-consistent pre-training framework for electroencephalography (EEG) foundation models. These models are pre-trained on large unlabeled EEG corpora. They are used for Brain-Computer Interface (BCI) applications.

ScaleBrain uses multi-scale EEG tokenization. This represents the same EEG segment through analysis windows with different temporal spans and shared temporal anchors. It also uses scale-specific reconstruction to align with the multi-scale input structure.

The framework includes an axial spatio-temporal attention mechanism. This mechanism contextualizes the fused multi-scale representation along spatial and temporal dimensions. Multi-scale representations are derived from fixed-length patches.

ScaleBrain was tested on six public BCI benchmarks against 12 representative baselines. It achieved the highest mean Balanced Accuracy on five datasets. It ranks third on the remaining one with only 5.0M parameters.

Key facts

  • ScaleBrain uses 5.0M parameters.
  • The framework achieves the highest mean Balanced Accuracy on five out of six public BCI benchmarks.
  • ScaleBrain uses multi-scale EEG tokenization and scale-specific reconstruction.

All sections

Edition