MoBI: when the body is part of the experiment
The classic lab pins the participant still precisely to keep the body out of the recording. Mobile brain-body imaging (MoBI) does the opposite: it records brain, body and behaviour together in people who walk, reach and interact, because that is where cognition actually happens. The cost is that every movement injects artifacts, so MoBI forces you to treat artifact handling not as pre-processing hygiene but as central to the measurement.
The artifact zoo and how to clean it
Real-world EEG carries a menagerie of non-brain signals: ocular potentials from blinks and saccades, EMG from jaw and neck muscles, cardiac and pulse artifacts, and broadband motion artifacts from electrodes shifting on the scalp. Some are larger than the neural signal. Two workhorses dominate cleaning. Independent component analysis unmixes the recording into components and lets you subtract those identified as eye or muscle. Artifact subspace reconstruction (ASR) learns the covariance of clean calibration data and, sample by sample, detects and reconstructs bursts that exceed it — well suited to the transient, high-amplitude artifacts of movement.
Non-stationarity is a different problem
Even after cleaning, the signal's statistics change. Neural non-stationarity and everyday drift mean the feature distribution the decoder sees on Tuesday is not the one it was trained on Monday. Framed formally, this is covariate shift: the input distribution changes while the relationship between input and intent is assumed stable. A decoder that was optimal for the training distribution is now biased, and accuracy sags without anything being 'broken'.
Covariate shift, the formal skeleton of everyday non-stationarity: only the input distribution p(x) drifts. This is exactly the assumption that licenses input-side fixes — re-aligning the features — rather than relabelling data.
Alignment: making sessions look alike
The most robust real-world fix is to align distributions rather than retrain from scratch. Riemannian methods treat each trial as a covariance matrix on a curved manifold of symmetric positive-definite matrices; Riemannian recentering whitens each session by its own reference mean so that different days are moved to a common origin before decoding. This simple, unsupervised step absorbs much of the between-session shift and is a large part of why Riemannian alignment and transfer learning are now standard for cross-session BCI.
Left: recentering whitens each session's covariances C_i by that session's reference mean C̄, cancelling a shared shift. Right: the affine-invariant Riemannian distance used to compare covariances; its invariance to linear mixing is what makes the alignment robust to changing sensor conditions.
When drift continues within a session, test-time adaptation updates the decoder (or its normalisation statistics) online from unlabelled incoming data, and covariate-shift-aware training deliberately exposes the model to varied conditions so it generalises. The theme is co-adaptation with the world: assume the signal will move, and build the moving-target assumption into the method rather than fighting it with daily recalibration.