A collection of 216 symmetric positive-definite 32-by-32 matrices with four event-label groups. The matrices represent test epochs from one participant in the MNE sample audiovisual experiment, after supervised xDAWN spatial filtering and prototype augmentation. The 32 coordinates comprise 16 class-prototype features and 16 filtered trial features, rather than 32 original EEG or MEG sensors.
Usage
data(ERP)Format
a named list containing
- covariance
an \((32\times 32\times 216)\) array of covariance matrices.
- label
a length-\(216\) factor with levels LA, LV, RA, RV and counts 57, 57, 53, 49.
Source
Historical pyRiemann ERP recipe, by Pedro Rodrigues (the historical filename says EEG, but the recipe explicitly selects MEG and excludes EEG).
MNE sample data on OpenNeuro, version 1.2.4;
pinned source metadata and CC0 declaration.
The MNE sample description
identifies the audiovisual experiment. The separate writing workspace
records the reconstruction in
maintenance/reports/erp-provenance-followup-2026-09-21.md.
Details
A September 21, 2026 reconstruction used the historical
pyRiemann 0.2.7 example and numerical routines at commit
176e766f540bd4c7846f38573165fc3d27fc69ca. All 216 stored matrices
matched with maximum relative Frobenius error \(5.56\times10^{-12}\),
and the complete label sequence matched exactly. No rescaling,
reordering, or eigenvector-sign correction was needed. The reconstruction
used recorded contemporary dependencies; it does not claim byte identity
or recovery of the original 2021 runtime.
The recipe reads MNE's prefiltered 0–40 Hz sample recording, sampled
at approximately 150.15 Hz after fourfold decimation, applies a 2 Hz
IIR high-pass filter, excludes MEG 2443, and selects 305 MEG
channels, with EEG excluded. Epochs cover 0–1 second without baseline
correction or projection, giving 288 epochs with 151 samples each.
A random 25 percent training split (seed 42, without stratification)
supplies 72 epochs for supervised xDAWN fitting with four filters per
class. The remaining 216 epochs supply the packaged matrices.
Covariance estimation uses centered empirical covariance with denominator
151, on the combined prototype and filtered-trial features.
These are repeated epochs from one participant, not independent participants. The representation was learned with labels from the separate 72-epoch training set, so it must not be described as an unsupervised transformation of the original sensors. A later split of the packaged matrices does not recreate raw-data preprocessing or establish performance for new participants.
The matrices retain their original numerical scale; small absolute
eigenvalues alone do not imply poor conditioning. No unit conversion,
regularization, data values, or labels were changed in this revision.
The MNE sample data's public OpenNeuro record, ds000248 version
1.2.4, declares CC0. See the source record for attribution and terms.
References
Gramfort A et al. (2013). "MEG and EEG data analysis with MNE-Python." Frontiers in Neuroscience, 7, 267. doi:10.3389/fnins.2013.00267 .
Gramfort A et al. (2014). "MNE software for processing MEG and EEG data." NeuroImage, 86, 446–460. doi:10.1016/j.neuroimage.2013.10.027 .