EEG decoders stop depending on the cap: hTSMNet aligns any montage from electrode coordinates alone and leads every dataset after brief calibration

An EEG decoder can now ingest recordings from arbitrary electrode layouts and place them in one shared coordinate space, without first selecting a common electrode subset or interpolating onto a fixed template. Earlier routes either forced every recording into a common input shape before training or treated each subject as a domain within one channel space for statistical alignment, and neither addressed montage incompatibility itself. Under a leave-one-dataset-out protocol spanning three motor-imagery datasets with 172 subjects and eight comparators, hTSMNet matches the strongest comparator at strict zero-shot, with a three-fold mean of 67.3%.

A handful of labelled calibration trials turns that shared space into the highest accuracy on every dataset, and the edge also holds subject by subject. At zero-shot it does not lead on every dataset; on PhysionetMotorImagery it trails the channel-matched NeuroGPT, 68.5% against 71.4%. At 20 calibration trials per class hTSMNet reaches 80.7%, 70.1% and 76.3% on the three datasets, 3.0, 4.1 and 2.8 percentage points above Deformer, the strongest comparator on every fold; paired tests over 171 subjects are significant against all eight comparators at p < 10⁻³.

The shared space works because it makes the left-versus-right discriminative direction point the same way across datasets, so a linear rule learned on source data is already correctly oriented on a held-out one. With no learning, projecting through geometry alone and then pooling covariance leaves the class axes of the three datasets barely agreeing, at a consistency cosine of 0.20. hTSMNet attains a cross-dataset class-direction consistency of 0.71±0.02, about 3.4 times the no-learning baseline and above EA–TSLR at 0.17 and the three deep comparators at 0.58–0.64.

The same frozen motor-imagery model also transfers to other paradigms, but only for oscillatory signals: mental workload and steady-state visual evoked potentials decode, while stimulus-locked ERP stays at chance. That split follows from the architecture, since covariance pooling reads band power and largely discards the phase-locked waveform ERP tasks depend on. Refitting only the linear classifier, mental workload reaches 84.3% at 50 trials per class and SSVEP reaches 68.0% and 72.2% at 20 trials per class, while both P300 datasets stay between 49.1% and 50.9% from zero-shot to 50 trials per class.

The demonstrated scope is leave-one-dataset-out left-versus-right-hand discrimination on three motor-imagery datasets, plus transfer from a frozen motor-imagery model to five target datasets covering mental workload, SSVEP and P300; all data pass through one MOABB preprocessing pipeline and each subject keeps its native montage. The authors propose next steps including a phase-preserving evoked-template branch alongside the shared-anchor machinery and a comparison against models fully trained on the target paradigm.

The class-direction analysis is correlational, and the authors note that a direct ablation randomising the sensor-to-anchor assignment has not yet been run. When the anchor count exceeds a dataset's channel count the aligned covariance becomes rank-deficient, and the 8-channel targets fall in that regime. All real-data results are binary sub-tasks, and preprocessing covers only one fixed bandpass, time window and reference scheme.