TRACE: Autoregressive EEG Pre-training Framework
Researchers have introduced TRACE, an autoregressive framework for EEG pre-training that forecasts future EEG segments based on causal context while enabling temporally adaptive and coherent cross-channel calculations. At each time point, TRACE generates a routing decision from the causal history across channels and applies this decision collectively to all channels, maintaining immediate coherence between channels while permitting various temporal regimes to engage distinct computations. This framework tackles the difficulties in acquiring transferable EEG representations due to its multi-channel and non-stationary characteristics, where simultaneous channels yield interrelated neural activity measurements and context-dependent temporal dynamics fluctuate. TRACE aims to address the shortcomings of models that utilize uniform computation over time or treat each channel patch separately.
Key facts
- TRACE is an autoregressive EEG pre-training framework.
- It predicts future EEG patches from causal context.
- It performs temporally adaptive and cross-channel coherent computation.
- At each temporal step, TRACE derives an expert routing decision from the causal cross-channel history.
- The routing decision is applied jointly to all channels at that step.
- This preserves instantaneous cross-channel coherence.
- Different temporal regimes can activate different computation.
- EEG signals are inherently multi-channel and non-stationary.
Entities
—