EEG-based brain-computer interfaces let a person interact with the world without moving, which matters for assistive technology and rehabilitation. Decoding methods built on symmetric positive definite matrices, which capture how signals across electrodes covary, have become the strong performers in this space.

Their weakness, as the authors put it, is architectural. These methods tend to use fairly basic networks and never explicitly model local relationships between EEG signals. Electrodes are not an unordered bag: neighbouring sites on the scalp are related in ways a method that ignores structure has to rediscover from data every time.

The proposal brings self-attention into that Riemannian setting, so the model can weigh which relationships between signals matter rather than treating the covariance structure as one undifferentiated block. Attention is well suited to the job, since deciding what to look at is exactly what it does.