Overview: EEG/EMG Foundation Challenge 2026

The EEG/EMG Foundation Challenge 2026 is a proposed multi-track competition on shift-robust representation learning for biosignals, and the multi-modal successor to the 2025 EEG Foundation Challenge.

Note

The competition is currently at the proposal stage and has not yet been accepted. This starter kit is shared so the community can experiment with the planned tracks while the proposal is under review.

The competition is organised as 4 + 1 tracks:

1. Track 1 – EEG-to-Image (cross-stimulus): retrieve a viewed image from EEG. Headline metric: Top-5 retrieval accuracy (higher is better).

2. Track 2 – EEG-to-BCI (cross-session, within-user): decode one of three cued mental tasks (motor imagery, mental calculation, word association). Headline metric: balanced accuracy averaged over subject-session-context cells (higher is better).

3. Track 3 – Sleep onset (cross-subject): estimate sleep-onset latency from wearable EEG. Headline metric: recording-level mean absolute error in seconds (lower is better).

4. Track 4 – EMG-to-Text (cross-subject): decode keystroke sequences from surface EMG. Headline metric: character error rate (lower is better).

5. Track 5 – Foundation Transfer (cross-task, EEG only): submit a single EEG encoder; organisers fit linear probes for Tracks 1-3. Headline metric: mean rank across the three EEG leaderboards (lower is better).

Tracks 1-4 accept both task-specific models and foundation models. Track 5 is restricted to a single shared encoder.

This starter kit shows how to reproduce a baseline for each track with NeuralBench, using the closest publicly available dataset when the official competition data is not yet released.

Starter kit pages

• Track 1 – EEG-to-Image

• Track 2 – EEG-to-BCI

• Track 3 – Sleep onset

• Track 4 – EMG-to-Text

• Track 5 – Foundation Transfer

• How to Submit a Model

Each page follows the same shape:

1. What the track measures (data, shift, headline metric).

2. The matching neuralbench task and CLI commands.

3. Where the official competition data diverges from the default.

Representative results

The table below summarises NeuralBench results on publicly available datasets that are similar in paradigm and modality to the ones the competition will use. These numbers are not the competition leaderboard – the official tracks will use distinct datasets, hidden evaluation sets, and rerun protocols. They are useful as a sanity check that your training pipeline behaves like the published baselines on the closest open data.

Model	Image (Top-5 %, higher)	BCI (Bal. acc %, higher)	Sleep (Onset MAE s, lower)	EMG (CER %, lower)
Chance	2.22 +/- 0.31	24.81 +/- 1.03	205.42 +/- 0.01	96.71 +/- 0.00
Dummy	2.50 +/- 0.00	25.00 +/- 0.00	299.90 +/- 0.00	100.00 +/- 0.00
EEGNet	28.13 +/- 0.14	58.58 +/- 0.34	143.30 +/- 0.40	–
REVE (FM)	84.75 +/- 0.38	68.04 +/- 0.73	134.89 +/- 2.02	–
EMG2QwertyNet	–	–	–	25.14 +/- 2.30

REVE is also the reference Track 5 baseline: its three EEG-track numbers come from a single set of encoder weights, so its mean rank on the Tracks 1-3 leaderboards is also its Track 5 score.

Collecting and plotting your results

Every NeuralBench run caches its test-metric dictionary on disk. After the experiments you care about have finished, re-invoke the same CLI command with --plot-cached to aggregate results into comparison plots and CSV tables without retraining:

# 1. Run experiments (cached automatically)
neuralbench eeg image motor_imagery sleep_onset -m eegnet reve

# 2. Aggregate cached results -- no retraining
neuralbench eeg image motor_imagery sleep_onset -m eegnet reve --plot-cached

--plot-cached produces, under <SAVE_DIR>/outputs/:

• core/core_bar_chart.png – bar chart per task and model.

• core/core_results_table.csv – raw per-task metrics.

• core/core_rank_table.csv – ranks per task (the input to the Track 5 mean-rank score).

For programmatic access to the same data, instantiate BenchmarkAggregator directly. The Visualizing Benchmark Results tutorial walks through the full Python API, including how to customise the loss-to-metric mapping and the output directory.

Resources and dependencies

This starter kit relies on the following organiser-maintained open-source libraries:

• NeuralBench (this package): unified benchmark suite.

• NeuralSet: data loading, study registry, event system.

• Braindecode and MOABB: deep-learning EEG architectures and BCI benchmarks.

• MNE-Python and EEG-Dash: signal-processing primitives.

If neuralbench is not yet installed, follow the installation guide and the quickstart before continuing.

Known gaps in this starter kit

Some pieces of the planned competition pipeline are not yet shipped with NeuralBench at the time of writing. The following items are being worked on and will be folded back into the relevant track page once they land:

1. Official Track 2 dataset (MI / Calc / Word, 20 subjects, 6 sessions, Graz + BrainHero). The corresponding study is not public yet. Track 2 currently uses Stieger2021Continuous (4 motor-imagery classes, cross-subject) as the closest analog.

2. Muse sleep-onset training set (~1000 subjects). The Track 3 page currently runs on Kemp2000Analysis (Sleep-EDF) – and the additional Ghassemi2018You / Alvarez2022Haaglanden PSG datasets – with the same SleepOnsetTargetExtractor + bmae metric the competition will use.

3. Hidden evaluation sets. Every track is scored against a hidden test set (Alljoined Emotiv for Track 1, later Graz/BrainHero sessions for Track 2, the Muse cohort for Track 3, 100 new users for Track 4). Only the closest public training data ships here, so the numbers above are sanity checks, not leaderboard scores.

Updates and corrections to this page are tracked on the GitHub repository linked from the (future) competition website.