Montage-Free Riemannian Calibration: Learning a Scalp-Coordinate Projection for Few-Trial EEG Transfer

Brain-computer interfaces and clinical EEG studies usually have very few labeled trials per person, and in that regime the most reliable decoders do not use deep networks. They summarize each trial by its spatial covariance matrix (how the voltages at every pair of electrodes co-vary) and classify these matrices using the curved geometry of positive-definite matrices. The best way to calibrate a new user from a handful of trials is to borrow class averages from other people and blend them with the few target trials along this geometry. That trick has a hard precondition. Every recording must use the same electrodes in the same order, because entry $(i,j)$ of the covariance must mean the same electrode pair for everyone. In practice, public datasets and hospitals use 22, 32, 62, 64 or 118 electrodes, and electrodes fail or shift between sessions. The pool of "other people" a new user can borrow from is therefore much smaller than it could be.

We propose to remove that precondition. We learn one smooth spatial field over the scalp that turns any electrode set, given its 3D coordinates, into the same small number $p$ of "virtual channels." Each subject's covariance is then projected into a shared $p \times p$ space where geodesic calibration can pool subjects across datasets. This is the EEG version of the antenna-grouping and subspace ideas from wireless arrays. Nearby, correlated sensors carry few real degrees of freedom, so they can be represented by a low-dimensional summary. The difference is that the grouping here is a continuous function of sensor position, so it transfers to arrays it has never seen. The claim to test is specific: with 1 to 10 labeled trials per class, a covariance classifier calibrated on a new montage should beat fixed interpolation onto virtual electrodes, restriction to shared channels, and 2D-grid imputation, when it borrows from subjects recorded with different montages through this learned projection.

Core Claim and What Is New

The ingredients exist separately. Rodrigues et al. learn an orthonormal projection $W$ that maps $n \times n$ EEG covariances to $p \times p$ while preserving Riemannian distances to the dataset mean. They show that going from 118 to 32, and often to 12, dimensions keeps classification AUC intact. Their $W$ is a free matrix indexed by channel number, fit per dataset, so it cannot be applied to a different montage. The Minimum Distance to Weighted Mean method, MDWM, blends a target user's few-trial class means with source class means along the geodesic, controlled by a weight $\lambda$ (best near 0.7). It beats CSP+LDA with as few as 5 target trials, but it assumes identical electrode layouts. IMAC and the sensor-configuration CNN do handle heterogeneous layouts. They interpolate onto a flat 2D grid and train image-style models, which gives up the covariance geometry that the MOABB benchmark found decisive below roughly 150 trials per class.

The new element is to parameterize the dimensionality-reducing projection as a function of electrode coordinates, fit it jointly across many montages, and feed the result into geodesic few-shot calibration. Concretely, let $\Phi_s \in \mathbb{R}^{n_s \times B}$ be $B$ real spherical-harmonic basis functions evaluated at the $n_s$ electrode positions of subject $s$. Let $A \in \mathbb{R}^{B \times p}$ be a coefficient matrix shared by all subjects. The subject-specific projection is

$$ W_s = \Phi_s A \left(A^\top \Phi_s^\top \Phi_s A\right)^{-1/2}, $$

which is orthonormal by construction and full rank whenever $n_s \ge p$. We fit $A$ with the unsupervised distance-preservation objective of Rodrigues et al., pooled over subjects and trials:

$$ \min_A \sum_{s}\sum_{k} \Big(\delta(C_{sk}, M_s) - \delta\big(W_s^\top C_{sk} W_s,\; W_s^\top M_s W_s\big)\Big)^2 . $$

Here $C_{sk}$ is trial $k$'s covariance, $M_s$ is subject $s$'s Riemannian mean, and $\delta$ is the affine-invariant Riemannian distance.

We use the unsupervised form on purpose. Rodrigues et al. found that the supervised version gave no gain and scales quadratically in trial count. Being unsupervised also means a new user's unlabeled trials can join the fit. The projected covariances are recentered per subject, in the spirit of the domain-specific centering in TSMNet. They are then classified by MDWM in the shared $p$-dimensional space. Optimization uses the trivialization parameterizations and stable spectral gradients from SPD Learn, which address the log-map instability that appears when eigenvalues approach zero.

Research Program

The first stage builds the multi-montage source pool from MOABB motor-imagery datasets that share a left-hand versus right-hand contrast and publish electrode coordinates. These are BNCI2014001 (22 channels), PhysioNet MI (64), the KU dataset (62), BCI Competition III IV (118), BNCI2015001 and Zhou2016. The 3-channel BCI IV 2b is excluded because it cannot support $p > 3$. The second stage fits $A$ for a grid of $p \in \{6, 8, 12\}$ and harmonic degree $L$. It also inspects the learned virtual channels as scalp maps, to check whether they concentrate on sensorimotor cortex or simply reproduce uniform interpolation. The third stage is few-shot calibration. For each held-out subject we draw $k \in \{1, 2, 5, 10, 20\}$ labeled trials per class, with 10 random draws, and test on the remaining trials. The fourth stage stresses the spatial assumption. It randomly deletes 10 to 40 percent of the target's electrodes and jitters electrode coordinates by 1 to 2 cm, mimicking cap misplacement. This probes the "rigidity under array perturbation" limitation of static spatial projections.

The main protocol is leave-one-dataset-out. The target dataset's montage never appears when $A$ is fit, so any gain must come from the field generalizing across layouts, not from memorizing a montage. The secondary protocol is within-dataset leave-one-subject-out, where the montage is shared. It separates the benefit of dimensionality reduction from the benefit of cross-montage pooling.

Baselines

The naive reference point is target-only tangent-space logistic regression on the subject's full native montage, trained on the same $k$ trials per class. It shows how much any borrowing helps at all. It will be regularized with the same cross-validated shrinkage used everywhere.

The strongest existing method on this question is MDWM (Rodrigues/Chevallier). We run it in the only way its design allows across montages: sources are restricted to the channels shared with the target, and $\lambda$ is tuned on the same grid and the same source-only validation split as ours. The file reports no absolute MDWM score. It reports only significance ($p<0.001$ over CSP+LDA) and $\lambda \approx 0.7$, so we have no reported result on Lacuna for MDWM's AUC and will produce it ourselves. For orientation on BNCI2014001, the file reports FBCSP at mean kappa 0.569 (4-class, full training data; Chin et al.), DOT at 78.83% accuracy in unsupervised cross-session adaptation, and FusionGen raising cross-subject accuracy from 53.62% to 57.78%. These use different protocols and class counts, so they are context rather than targets. Beating MDWM means a higher subject-averaged AUC at $k \le 10$, significant under a paired Wilcoxon test over subjects, in at least four of six held-out datasets.

The ablation that isolates the key ingredient is the fixed-field variant. It projects each montage onto $p$ virtual electrodes at fixed scalp locations using spherical-spline interpolation, followed by the same orthonormalization, recentering and MDWM. This is the proposed pipeline with $A$ frozen to a hand-chosen interpolation kernel, so any gap is attributable to learning the field. A second ablation replaces MDWM with target-only classification in the shared space. It separates what the projection contributes from what pooling across subjects contributes.

The heterogeneous-layout competitor is an IMAC-style pipeline (IMAC). It interpolates every montage to a common 64-channel grid by radial basis functions, then runs both its Transformer classifier and tangent-space MDWM on the gridded signals. IMAC reports 90.92% cross-center motor-imagery accuracy, but under its own pretraining protocol. We rerun it under our $k$-shot splits with the same pretraining data and the same GPU-hour budget as our fit of $A$.

The deep-learning reference is EEGNet trained with Euclidean Alignment on the shared-channel intersection, with the same number of hyperparameter trials. Chevallier et al. report that deep pipelines often sit at 0.60 to 0.70 AUC where Riemannian pipelines exceed 0.80 in motor imagery.

Why not just…?

Why not just interpolate every montage onto a common dense grid and run a Riemannian classifier there? Interpolation is a linear map, so a 22-channel recording spread onto 64 grid points produces a 64-by-64 covariance of rank at most 22. That matrix is singular, which is exactly the log-map failure mode listed among Direction 1's open problems. Shrinkage patches this but adds a nuisance parameter that interacts with $k$. IMAC sidesteps the issue by abandoning covariance geometry for a Transformer, and that trade is what the MOABB results warn against at small sample sizes. The grid baseline above measures this directly.

Why not just interpolate onto a small fixed set of virtual electrodes, with $p$ below the smallest montage? This is the real threat. It avoids rank deficiency, needs no training, and could capture most of the benefit. It is also exactly our frozen-field ablation. If the learned field does not beat it by a significant margin at $k \le 10$, we will report that the contribution is the cross-montage geodesic calibration protocol and that the fixed virtual electrodes are sufficient. We expect learning to matter most under electrode dropout and displacement. A fixed kernel's quality depends on having sensors near each virtual location, whereas the learned field can shift weight toward whichever informative regions remain.

Why not just use the channels shared by all montages? Across 22-, 62-, 64- and 118-channel caps, the intersection shrinks to a handful of standard 10-20 positions. It also shrinks further when any electrode drops out. The shared-channel MDWM baseline tests this directly. Rodrigues et al. show that choosing the subspace well, rather than truncating arbitrarily, is what keeps performance at $p = 12$.

Why not just use Euclidean Alignment with a linear classifier or EEGNet? Kobler et al. found that recentering on the positive-definite manifold outperformed Euclidean alignment and Euclidean batch normalization. MOABB places deep models well below Riemannian pipelines under 150 trials per class. Both results point toward our design, but we still include EA+EEGNet on the shared channels so the conclusion does not rest on other papers' protocols.

Decisive Checks

The idea stands or falls on four results. The first is that in leave-one-dataset-out evaluation with $k \le 10$, the learned field beats the frozen spherical-spline field at the same $p$. If it does not, the "learned spatial correlation" part of the claim is refuted. The second is that the learned field beats shared-channel MDWM, which shows that pooling across montages is worth more than the channels lost by projection. The third is that the gap over target-only classification shrinks as $k$ grows toward 20. The fourth is that under 40 percent electrode dropout and 2 cm jitter, the learned field loses less AUC than both the frozen field and grid imputation. The second and third together would show that the gain comes from borrowing, as it should in a small-sample regime. A negative outcome on the fourth would mean the field has merely memorized standard cap positions. Scalp maps of the learned virtual channels, and their stability across random seeds and held-out datasets, give a qualitative check that the projection has found physiologically plausible structure and not dataset-specific artifacts.