CaMBRAIN: Real-time EEG Inference with Causal State Space Models
Researchers propose CaMBRAIN, the first causal Mamba-based state space model for real-time EEG inference. Existing deep learning methods struggle with EEG's variable-length signals due to attention mechanisms' quadratic scaling and fixed-length input requirements. CaMBRAIN leverages the causal, unidirectional nature of EEG to avoid expensive bidirectional approaches, enabling continuous processing of signals spanning seconds to hours. Training is challenging because crucial EEG events can be extremely brief yet separated by long intervals. The model aims to provide global understanding of entire EEG signals without sliding-window constraints.
Key facts
- CaMBRAIN is a causal Mamba-based state space model for EEG inference.
- It addresses quadratic scaling of attention mechanisms in long sequences.
- Raw EEG signals are processed without fixed-length input requirements.
- The model exploits the causal, unidirectional nature of EEG.
- Training is difficult due to brief but separated EEG events.
- It enables real-time inference of continuous EEG signals.
- Existing models use sliding-window approaches that prevent global understanding.
- The approach is presented as the first of its kind for EEG.
Entities
—