Reference

Glossary

Plain-language definitions of the signals, statistical intervals, decoding families, and verification rules used across BCI Report.

Signals & tasks

What the scalp electrodes record and which mental or sensory paradigms the protocols test.

Brain–computer interfaceBCI#

A system that translates recorded brain signals into discrete selections, commands, or state estimates.

Offline decoding accuracy on laboratory recordings is only one part of a working BCI; real deployments also face sensor shifts, movement artifacts, calibration effort, and unwanted triggers during idle periods.

Electrode montage & channel#

The count, scalp locations, and referencing scheme of the electrodes recorded in a protocol.

A single sensor recording is a channel; a referenced pair (such as C3–M2 in sleep polysomnography) is a derivation. Changing from 8 to 4 posterior electrodes or shifting a cap by one spacing alters signal quality directly.

ElectroencephalographyEEG#

Non-invasive recording of microvolt-scale electrical potentials from electrodes placed on the scalp.

Scalp EEG offers millisecond time resolution but mixes neural activity with eye blinks, muscle tension, and line noise. Every core protocol on this site evaluates scalp EEG; non-scalp modalities such as fNIRS and intracortical spiking are kept on their own page.

Mental workload & arithmetic#

Comparing EEG recorded during sustained mental effort (such as serial subtraction) against a resting baseline.

On the core matrix, the arithmetic-and-rest protocol evaluates 36 participants on EEGMAT across 21 channels, where spectral features, Riemannian covariance methods, compact networks, and frozen encoders are all measured under the same split.

Motor imageryMI#

Mentally rehearsing a hand or foot movement without moving, producing measurable changes in sensorimotor rhythms.

Motor imagery alters mu (8–12 Hz) and beta (13–30 Hz) band power over motor cortex, though patterns vary strongly across individuals and recording sessions.

P300 & event-related potentialERP#

Time-locked voltage deflection triggered by a specific stimulus; P300 is a positive peak roughly 300 ms after a rare or attended target.

Unlike steady rhythmic tasks, ERP decoding depends on the exact waveform timing within a short window after stimulus onset, which is why time-domain and xDAWN spatial-filter methods behave very differently from pure band-power features.

Sleep staging#

Classifying consecutive 30-second overnight recording frames into five standard sleep stages: Wake, N1, N2, N3, and REM.

Because stages occur in unequal proportions (N2 is common while N1 is brief), plain accuracy can look reassuring even when a classifier misses an entire stage. Balanced accuracy weights all five stages equally.

Steady-state visual evoked potentialSSVEP#

Periodic occipital brain response locked to the flickering frequency (and harmonics) of an attended visual target.

Because the stimulus frequencies are known in advance, frequency-matching methods such as CCA and FBCCA can decode 40-target SSVEP without any training data. On BETA, uniform guessing across 40 targets is 2.5%.

Evaluation & statistics

How protocols split participants, measure accuracy against chance, and quantify variation.

Balanced accuracy#

The unweighted mean of per-class recall, giving every class equal influence regardless of how often it appears.

In an imbalanced task such as oddball P300 or five-stage sleep, a classifier that always predicts the majority class can score high plain accuracy while achieving only the uniform chance level in balanced accuracy.

Chance level#

The expected balanced accuracy of uniform random guessing: 100% divided by the number of classes.

Chance is 50% on two-class tasks, 20% on five-stage sleep staging, and 2.5% on 40-class SSVEP. A raw percentage cannot be interpreted without its protocol’s chance line.

Descriptive 95% interval#

An interval summarizing variation across held-out participants, explicitly labeled descriptive when training folds overlap.

Because cross-validation folds share training participants, standard independence assumptions do not strictly hold. Printing the interval alongside the point estimate shows whether a result clears chance or overlaps another configuration.

Fixed protocol#

A locked specification of dataset version, cohort, channels, windowing, preprocessing, data split, and metric.

Scores are only directly comparable inside the same fixed protocol. Comparing two methods evaluated on different cohorts, electrode layouts, or split rules mixes method differences with task difficulty.

Idle false activation & abstention#

Firing a command when the user is idle (false activation) versus intentionally declining to output a command (abstention).

A decoder can show 0% false activations simply by abstaining on almost every command. On the idle protocol, command detection and idle false activation must always be read together as a two-axis trade-off.

Macro F1#

The unweighted average of the F1 score (harmonic mean of precision and recall) across all classes.

Reported beside balanced accuracy on every accuracy protocol to surface precision-versus-recall imbalances across classes.

Paired contrast#

The per-participant difference between two methods scored on the exact same people and folds, summarized with its own interval.

Two marginal intervals can overlap even when one method scores consistently above the other on almost every individual, because shared person-to-person variation cancels out in the paired difference. Where multiple planned contrasts are tested together, Bonferroni-adjusted intervals govern the reading.

Participant bootstrap#

Resampling held-out participants with replacement (typically 10,000 draws) to compute the 2.5th–97.5th percentiles of the equal-participant mean.

Resampling trials instead of people would treat correlated windows from the same person as independent and produce misleadingly narrow intervals. Resampling people reflects real person-to-person spread.

Participant-disjoint split#

A cross-validation split in which every recording from a given person stays strictly in either the training fold or the test fold.

Keeping people disjoint prevents a model from recognizing a person’s resting signature or cap impedance rather than the target mental state, directly measuring zero-shot transfer to an unseen person.

Percentage pointspp#

The arithmetic difference between two percentages (for example, moving from 50% to 60% is a change of +10.0 pp).

Using percentage points avoids ambiguity with relative percent change. All paired contrasts and interval widths on this site are stated in pp.

Methods & models

Classical signal-processing pipelines, compact neural networks, and frozen or adapted foundation encoders.

CCA & FBCCAFBCCA#

Training-free frequency-matching methods that correlate scalp EEG against synthetic sine–cosine templates; Filter-Bank CCA weights multiple harmonic sub-bands.

Because CCA and FBCCA rely on the known flicker frequencies rather than labeled calibration trials, they serve as strong training-free references on both 8-channel and 4-channel BETA SSVEP.

Compact neural network#

A small convolutional architecture (typically a few thousand to tens of thousands of parameters) trained from scratch on one protocol’s training folds.

Core-matrix compact networks include EEGNet, ShallowFBCSPNet, and Deep4Net. Each encodes temporal and spatial filtering layers tailored to multi-channel EEG.

CSP + LDACSP+LDA#

Common Spatial Patterns projects multi-channel EEG to maximize variance contrast between two classes, followed by Linear Discriminant Analysis.

A workhorse classical pipeline for motor imagery and oscillatory band-power tasks; extended via one-vs-rest for multiclass protocols and omitted on 40-class SSVEP where it does not apply.

eTRCAeTRCA#

Ensemble Task-Related Component Analysis learns spatial filters from labeled SSVEP calibration trials to maximize reproducibility across repeats.

Unlike training-free CCA, eTRCA requires labeled calibration trials for each target frequency, making it a central baseline in the calibration-budget and dry-vs-wet transfer studies.

Foundation encoder#

A large neural network pretrained via self-supervised objectives on many public EEG corpora before being tested on downstream tasks.

Examples on this site include LaBraM, CBraMod, EEGPT, REVE, LUNA, ST-EEGFormer, BrainOmni, CodeBrain, EEGMamba, ERP-FM, SingLEM, and ZUNA. Each checkpoint is tracked with its parameter count, weights licence, and pretraining corpus record.

Frozen probe#

Evaluating a pretrained encoder with all its weights locked, fitting only a lightweight linear classifier on its extracted features.

A frozen probe tests what the pretrained representation makes linearly accessible under one reproducible recipe. It does not establish what task-specific fine-tuning might reach.

LoRA & model adaptationLoRA#

LoRA inserts small trainable factorized adapter matrices into a pretrained encoder instead of updating every parameter.

The model-adaptation experiments compare head-only updates, last-block updates, and LoRA (r = 4) under a fixed recipe on EEGMAT, reporting paired changes and how many participants improved or declined.

Random-initialization control#

Running the exact same neural encoder architecture with fresh random weights instead of pretrained weights to isolate what pretraining actually adds.

A frozen random convolutional or transformer backbone can still act as a useful nonlinear feature projection. Comparing pretrained weights against three random initializations separates architecture bias from pretraining gains.

Readout head & ridge head#

The small final classifier attached to extracted features; a ridge head is an L2-regularized linear classifier fit on standardized training features.

Standardizing features on the training fold alone prevents test-fold statistics from leaking into scaling. Fixing the ridge penalty keeps every frozen encoder on an identical readout budget.

Riemannian TS+LR & MDMTS+LR#

Geometry-aware classifiers operating directly on symmetric positive-definite channel covariance matrices: Tangent Space + Logistic Regression and Minimum Distance to Mean.

Because EEG channel covariances lie on a curved manifold rather than flat Euclidean space, Riemannian metrics respect matrix geometry and often require no neural-network hyperparameter search.

Spectral ridge#

A fast classical baseline that summarizes EEG windows by their frequency-band power spectrum and fits an L2-regularized linear classifier.

Spectral ridge captures steady rhythmic patterns such as sleep oscillations and alpha/beta shifts, while ignoring exact time-locked phase waveforms such as P300.

xDAWN spatial filteringxDAWN#

A supervised spatial filter designed to enhance the signal-to-noise ratio of time-locked evoked potentials such as P300.

On the core matrix, xDAWN appears in two classical ERP pipelines: xDAWN+LDA and XDAWNCov+TS+LR (which combines xDAWN-filtered template covariances with Riemannian tangent-space logistic regression).

Governance & data

How datasets and models are audited, tracked on the open register, held back, or cited.

Aggregate-only publication#

Publishing only cohort-level summary statistics — never raw EEG recordings, individual participant scores, or redistributed model weights.

This boundary protects participant privacy, respects upstream dataset and model licences, and keeps every downloadable JSON/CSV file lightweight and auditable.

Open register#

Two public catalogues — one for datasets and one for models — listing every EEG resource we track and how far it has been verified.

Anything can be listed on the register regardless of gating or rights status, whereas score tables only publish results that clear rights review and end-to-end file verification.

Pretraining exposure#

A sourced record stating whether a benchmark dataset appears in a foundation model’s published pretraining corpus list, or whether no list is published.

Exposure is documented per cell rather than hidden or treated as an automatic disqualification, so readers can interpret cells flagged “In pretraining list” or “Exposure unknown” with full context.

Provenance & audit hash#

Cryptographic SHA-256 digests pinning each released JSON/CSV file, protocol specification, and archived Zenodo record.

Every published figure on the site traces back to a leaf value in a served data file whose SHA-256 digest is recorded in the release archive, letting anyone verify that a number has not drifted.

Publication hold#

Withholding scores when a dataset’s rights terms, timing integrity, consent record, or experimental sanity check has an unresolved problem.

Held sources are listed openly with the specific reason rather than silently dropped, preventing survivorship bias in the benchmark record.

Verification ladder (Steps 1–5)#

A five-step scale tracking how thoroughly a dataset or model has been checked, from listing its official record (Step 1) to entering the core matrix (Step 5).

Datasets progress through Listed, Read, Traced, Run, and In the matrix; models progress through Identified, Terms recorded, Files checked, Scored here, and In the matrix. Gated data and suspected re-uploads stay capped at Step 1.

New to EEG decoding? Read the five-minute Start here guide →