Can one EEG encoder answer several questions?

A Jev-style interface encourages a useful engineering question: if several decisions concern the same signal, how much computation can they share? In one fixed OpenBMI experiment, a shared EEGNet encoder answered two predefined questions with one encoder pass. However, the simplest shared heads did not establish accuracy within a predeclared two-percentage-point margin of separate models on both questions. Telling the head which question it was answering added nothing measurable: it was equivalent, within the same margin, to a shared hidden layer of similar size that was not told.

The shared-encoder experiment, released on 7 October 2026, separates three design choices: sharing an encoder, choosing the head's capacity, and conditioning that head on a question identifier. The decision-model interface described by ggml-org, which follows TypeSafe's Jev format, motivates this question. An interface that returns several answers does not by itself tell us which choice improves accuracy or reduces cost. Those choices need controls that receive the same data.

Here “Jev-style” means encoding a window once and answering several explicit, typed questions about it. In this study each question reaches the model as an identifier, not in natural language.

This article examines the experiment's primary, three-seed EEGNet arm on OpenBMI. It includes 51 people and 31,653 quality-passing two-second windows. Five outer folds keep each person's sessions together and separate people between training and test. Scores are balanced accuracy, averaged equally over participants. The final epoch is scored, with no validation-based checkpoint selection or early stopping.

There are two primary questions. “Imagery or rest?” is defined over all 31,653 windows. “Which hand?” is defined only on the 10,199 offline motor-imagery trial windows. The latter question has no evaluated answer on a rest window. This distinction matters when turning a multi-question model into an API: not every question is meaningful for every input.

Four configurations make the comparison:

Figure 1. OpenBMI's primary EEGNet comparison, three seeds and 51 participants. Whiskers are published hierarchical paired-bootstrap 95% intervals; dashed lines mark 50% balanced-accuracy chance. Scores compare fixed recipes, not deployment error rates. The imagery/rest target also separates cue display and position in the run; the hand target uses offline trials only. Data: BCI Report, CC BY 4.0, computed from OpenBMI by Lee and colleagues. Figure rendered from published aggregates.

The three shared configurations receive the same windows and labels, with matched initialization, batch order and step count. The independent configuration has a separate encoder for each question, and the ledger counts that extra work. Its total step count differs, so the main comparison should not be described as identical training compute across all four configurations. A separate step-matched sensitivity analysis is available in the full release.

The question-conditioned head adds FiLM modulation from a learned 16-dimensional question embedding. The shared-hidden-layer control is adjusted to a similar parameter budget. Both are more expressive than the small linear heads. Comparing conditioning only against a linear head would combine two changes: extra head capacity and question conditioning.

The experiment fixed a ±2 percentage-point margin before results were known. Its published contrasts carry two independent readings: whether the paired interval excludes zero, and whether it fits the predeclared margin. The relevant EEGNet contrasts are:

For imagery/rest, the conditioned head scores higher than the linear heads under the release's interval rule, but the extra shared hidden layer provides the needed control: that comparison is within the declared equivalence margin. This is evidence about these two predefined questions and this recipe. It does not show that conditioning can never help, or that language-conditioned EEG models are unnecessary.

The cheaper fixed heads also miss the margin against separate models. On the hand question, the paired interval favors the separate model. On imagery/rest, it includes zero but extends beyond the tolerated loss, making the result inconclusive at this sample size. Consequently, “the same accuracy for less” would overstate this experiment. The parameter and encoder-pass savings are real; the all-questions accuracy condition was not established.

Both questions pass the release's prespecified simple-feature check and headroom gate at this EEGNet level. Passing that check means the tested mean-and-variance baseline did not reach its demotion threshold; it does not rule out other recording artifacts or contextual cues. Intervals use 2,000 hierarchical paired-bootstrap draws, resampling people and then seed indices within folds. The full study has 21 predeclared primary entries and no multiplicity correction; at 95%, about one in 20 true-null entries may be flagged by chance. The three seeds add training variation, but the intervals do not remove all dependence arising from shared cross-validation fits.

There are two further limits to the interpretation. First, imagery versus rest also distinguishes cue display and position in the run, so it cannot be read as a pure measure of intention. Second, these are small EEGNet encoders trained from scratch: the shared trunk has 2,096 parameters. The full release separately includes frozen and LoRA-adapted CBraMod and sleep experiments. Their conditions and results are not pooled into this table, and this article establishes no general foundation-model ranking.

For a developer considering a Jev-style EEG service, this suggests a concrete comparison: start with independent models, measure what encoder sharing saves, then compare a conditioned head against an unconditioned head with similar capacity. Define which questions are valid for which windows, and set the tolerated accuracy loss before looking at results. The interface can be designed around typed questions while the evidence determines the model behind it.

The public table can be inspected with a short script:

import json
from urllib.request import urlopen

commit = "21f712d55fc7be7c055fc916bbd53ffe44ddd5d8"
url = (f"https://raw.githubusercontent.com/twu3202/bci-report/{commit}/"
       "site/public/data/shared-representation-update.json")
with urlopen(url, timeout=30) as response:
    result = json.load(response)["results"]["one-representation"]
for row in result["datasets"]["openbmi"]["arms"]:
    if row["level"] == "E1":
        print(row["arm"], row["question"], f'{100 * row["mean"]:.1f}%')

The full result page provides the remaining arms, gates, cost ledger and protocol limits. The Hugging Face dataset and data API make the aggregate export available for reuse; no EEG recordings, participant-level scores or trained weights accompany this article.

OpenBMI was created by Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli and Seong-Whan Lee: EEG dataset and OpenBMI toolbox for three BCI paradigms, with supporting data on GigaDB, released under CC0. The encoder follows EEGNet. BCI Report's aggregate measurements are CC BY 4.0. These evaluations imply no endorsement by the source authors.

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.