Multi-timescale adaptive attention with trainable timescale aggregation for EEG foundation models
Use a foundation-model design for EEG that (1) replaces fixed temporal-scale attention with a multi-timescale attention mechanism that can learn representations at multiple temporal resolutions, and (2) adds a trainable aggregation module that selects and adapts to the most informative timescale(s) for each input signal.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
EEG contains neural dynamics that naturally evolve across multiple timescales. A fixed-scale design can miss time-varying characteristics; a multi-timescale attention plus adaptive aggregation approach targets better capture and reuse of cross-timescale information for downstream tasks.
What changed
The approach moves from a fixed temporal scale to dynamic, input-dependent representation learning via multi-timescale attention and a trainable module that identifies/adapts to informative timescales.
Where it helps
- EEG classification or regression tasks that rely on temporal dynamics
- Downstream tasks expected to benefit from different temporal scales
- Foundation-model representations for physiological time-series
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The described method and claims are specific to EEG foundation-model settings; applicability to other signal types or modalities is not established in the provided text, and no quantitative evidence details are included here.
Sources
- MAGE: Towards Generalizable Multi-timescale EEG Representations ↗
arXiv · primary · Checked 2026-11-10
