Choosing how to adapt an EEG foundation model

Before deciding how much of an EEG foundation model to update, check how strong its frozen readout already is. In a fixed LaBraM experiment on mental arithmetic, updating the last transformer block or adding LoRA improved on a short head-only training run. A separately published frozen ridge readout, however, reached a similar mean score. The choice of readout changes what the apparent adaptation gain means.

The adaptation experiment asks about new people performing a known task in the same recording setup. It uses EEGMAT, a public dataset of rest and serial-subtraction mental arithmetic, with 36 participants and 2,160 two-second windows. Five folds keep participants separate between training and test. No labels from a test participant are used to adapt the model. This is not a test of transfer to another day, device or dataset.

Three update rules start from LaBraM Base: train a linear head while keeping the encoder frozen; train the last transformer block and the same head; or train rank-4 LoRA adapters and the same head. Within each seed and fold, the three arms start from the same head, receive the same training batches in the same order, and train for five epochs. The recipe uses AdamW, learning rate 0.0001, weight decay 0.01 and batch size 32. There is no hyperparameter search or checkpoint selection; the final epoch is evaluated.

Each rule is run with three declared seeds. Scores are averaged across seeds within each participant and then equally across the 36 participants. This averages scores, not model predictions into an ensemble.

Figure 1. The upper three rows are the matched five-epoch experiment. The lower row is the separate published frozen-encoder ridge reference, not a fourth matched arm. Whiskers show descriptive participant-bootstrap intervals; the dashed line marks 50% chance. Data: BCI Report, CC BY 4.0, computed from EEGMAT by Zyma and colleagues. Figure rendered from published aggregates.

Against head-only training, the paired mean changes are +9.1 percentage points for updating the last block and +7.5 points for LoRA. For the LoRA comparison, 30 participants have a higher seed-averaged balanced accuracy and six have a lower one. The mean benefit does not describe every participant.

The direct LoRA-minus-last-block contrast is −1.6 percentage points, with a descriptive paired interval of −4.4 to +1.0 points. That interval does not establish an ordering, and it does not demonstrate equivalence. The published intervals resample people after averaging their seeds. They ignore dependence from shared cross-validation models and are not confirmatory generalization intervals.

The separate frozen ridge result deserves equal attention: 64.6% balanced accuracy, with a descriptive interval of 60.5–68.6%. It uses the same people and folds, but a closed-form ridge head on standardized features. The matched head-only arm above is a five-epoch gradient-trained head on unscaled features. The release explicitly marks the ridge reference as unpaired with the adaptation arms. It cannot be used to calculate a paired LoRA benefit, but it shows why “LoRA improves over this head-only recipe” is a narrower claim than “LoRA improves over frozen representations.”

The experiment also illustrates a practical trap in parameter-efficient adaptation. Here LoRA updates 38,802 parameters, far fewer than the last-block arm's 482,882. Yet the recorded training totals across 15 fits per arm were about 128.2 seconds for LoRA and 65.3 seconds for the last-block arm on one Apple-silicon Mac. This implementation forms effective weights rather than using an optimized low-rank kernel. The timings describe that implementation and device; trainable-parameter count alone does not establish a speed or memory advantage.

For a new project, this evidence suggests a sequence worth testing: establish a frozen ridge reference, run a matched adaptation comparison, and measure the actual resource cost on the intended hardware. Keep model-selection decisions inside the training data. Treat the five-epoch result as a starting recipe for investigation; it is not the best attainable score of each method. EEGMAT is also a known development cohort. It is not in LaBraM's published pretraining list (checked 4 October 2026), which is not proof that its recordings were never seen.

The following example retrieves the exact released three-arm experiment. It prints the table and the separately computed paired contrasts. It does not train a model or expose participant-level records.

import json
from urllib.request import urlopen

commit = "21f712d55fc7be7c055fc916bbd53ffe44ddd5d8"
base = f"https://raw.githubusercontent.com/twu3202/bci-report/{commit}/site/public/data/"
with urlopen(base + "adaptation-update.json", timeout=30) as response:
    result = json.load(response)["results"]["eegmat-labram-adaptation"]

for arm in result["arms"]:
    print(arm["label"], arm["trainable_parameters"],
          f'{100 * arm["balanced_accuracy"]["mean"]:.1f}%')
for contrast in result["paired_contrasts"]["balanced_accuracy"]:
    print(contrast["id"], f'{100 * contrast["mean_change"]:+.1f} pp')

The full result page includes per-seed cohort means, the paired changes and implementation limits. BCI Report also provides a data API and a Hugging Face mirror for reusing the aggregates. A later multi-encoder extension is linked on the result page; the numeric comparisons in this article are specifically the LaBraM release of 1 October 2026 and its earlier ridge reference.

EEGMAT was released by Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh and Oleksii Shpenkov as EEG During Mental Arithmetic Tasks 1.0.0, hosted on PhysioNet. See Zyma et al. 2019 for the study, Pollard et al. 2026 for the PhysioNet platform, and the LaBraM project (MIT) for the model. The source dataset and model authors do not endorse this evaluation. BCI Report's aggregate measurements are CC BY 4.0; EEGMAT retains its Open Data Commons Attribution License 1.0. No EEG recordings or adapted weights accompany this article.

To cite these results: BCI Report (2026), doi:10.5281/zenodo.23123296, together with the upstream dataset.

BCI Report, which produced these evaluations, is a personal, noncommercial project; this article uses only its published aggregate results.