docs / eeg-emg-foundation-challenge / overview

⌘ K · search Organizers GitHub Register

NeurIPS 2026 · open competition

EEG/EMG Foundation Challenge 2026.

Biosignal foundation models claim to transfer across subjects, sessions, and devices with limited per-task training.

The NeurIPS 2026 EEG/EMG Foundation Challenge tests this against strong task-specific baselines on four controlled shifts: cross-stimulus, cross-session, cross-device, cross-user.

A fifth track, Foundation transfer, tightens the control: one encoder across all four tasks, three EEG and one EMG.

The 4+1 design separates three claims usually conflated in the literature: single-task performance, shift robustness, and cross-task representation reuse.

Top-3 per track are re-run by the organizers on a sealed split. Prize pool: $30.000 across tracks 1–4. Winners present at the NeurIPS workshop in Sydney.

bash · pip · neuralbench

 1# 1. Install the benchmark stack
 2$ pip install neuralbench eegdash braindecode
 3 
 4# 2. Download and prepare data for Track 1
 5$ neuralbench eeg image --download  # THINGS-EEG2
 6$ neuralbench eeg image --prepare
 7 
 8# 3. Train a baseline
 9$ neuralbench eeg image -m eegnet
10  ↳ Track 1 baseline · NeuralBench v1 · facebookresearch/neuroai

4 decoding tasks

14 public datasets

3700 + subjects

Five tracks, one pipeline

Competition tracks

From cognition to clinic to foundation transfer.

One harness, one BIDS-first interface, one reproducibility audit. Holding evaluation fixed lets a per-track score be attributed to the model rather than to the protocol; stacking across tracks then tests reusability of representation.

SubmissionsJul 1 — Sep 1, 2026 (AoE)

FinalsNeurIPS · Sydney

Prize pool$30.000 ($2.500 / top-3 / tracks 1-4)

Track 01 OPEN

EEG-to-IMG

Decodes which image a subject is viewing from a single EEG epoch, evaluated as ranking against held-out candidates. Targets are frozen DINOv2-giant embeddings, so the difficulty is in the EEG side, not in re-learning the image space. The controlled shift is cross-stimulus: test stimuli are unseen during training.

Visual retrieval DINOv2 RSVP Object viewing

Metric: Top-5 acc · Sponsor: Alljoined Track details →

Track 02 OPEN

BCI decoding

Decodes a cued mental command (motor imagery, mental calculation, word association) from short EEG windows. The controlled shift is cross-session: models train on early sessions and are scored on later sessions of the same subject, with no per-session recalibration allowed.

Motor imagery Mental math Word association Session drift

Metric: Balanced acc · Sponsor: Meta FAIR Brain & AI Track details →

Track 03 OPEN

Sleep onset

Predicts the latency in seconds from recording start to the first stable N2 epoch, on consumer-grade wearable EEG. The controlled shift is to home-wearable channels and montages; precise timing replaces full staging because the wearable signal is too sparse to support per-epoch hypnogram reconstruction.

Sleep onset Wearable EEG Muse Home monitoring

Metric: W-bMAE (s) · Sponsor: InteraXon Track details →

Track 04 OPEN

EMG-to-Text

Decodes typed keystrokes from wristband surface EMG as a sequence-transduction problem. The controlled shift is cross-user: 100+ held-out users vary in forearm anatomy, typing strategy, and sensor re-placement, so a model that overfits one user's geometry fails the test.

EMG typing Sequence transduction Wristbands QWERTY

Metric: CER (%) · Sponsor: Meta Reality Labs Track details →

Track 05 OPEN

Foundation transfer

Forces a single set of encoder weights across all four tracks: EEG-to-IMG, BCI, Sleep, and EMG. Heads can be track-specific; the encoder cannot. Sharing weights across EEG and surface EMG makes single-task wins separable from reusable representation learning, which a per-track encoder cannot demonstrate.

Frozen probe Finetune Shared encoder EEG + EMG

Metric: Mean rank across 4 tracks · Host: Yneuro · Single encoder Track details →

Phases · Jul — Dec 2026

Timeline

From warm-up to NeurIPS finals.

One warm-up window, one final submission window, one reproducibility audit, one weekend in Sydney. Dates follow the NeurIPS 2026 competition guidelines: launch July 1, submissions close September 1 (AoE), final rankings November 1, workshop in December.

All timesAnywhere on Earth (AoE)

Daily cap5/day warm-up · 2/day sealed final

AuditTop-3 re-run on sealed split · ±2σ tolerance

InferenceFixed Codabench H100/H200 container · 60 min budget

WorkshopNeurIPS · Dec 11–12, 2026

Phase 01 NEXT
Public starter kit
Jun 1, 2026
Starter kit, baseline code, training data references, and the EEGDash + neuralbench installation path go public. Participants begin local iteration on the warm-up structure.

ReleaseJun 1, 2026

What shipsStart-kit, train data refs, install path

Repogithub.com/neural-interfaces26
Phase 02 FREEZE
Datasets, container, baselines frozen
Jun 15, 2026
Final dataset splits, data-use agreements, scoring container, and baseline weights are frozen. Beta dry-run of baseline submission, scoring, and top-submission reproduction completes.

FreezeJun 15, 2026

LockedSplits, DUAs, container, baselines

Dry-runInternal beta complete
Phase 03 UPCOMING
Warm-up
Jul 1 — Jul 31, 2026
Public validation set, unlimited submissions, baseline models released on Jul 1. Get your pipeline running end-to-end before the sealed test opens.

OpensJul 1, 2026

Validation setPublic release

SubmissionsUnlimited
Phase 04 UPCOMING
Final submissions
Aug 1 — Sep 1, 2026 (AoE)
Sealed test set opens Aug 1. Daily submission caps enforced. Submissions close Sep 1 AoE. Top-ranked teams submit a short method description with their final submission for audit and authorship.

OpensAug 1, 2026

ClosesSep 1, 2026 (AoE)

Daily cap5 / team / day
Phase 05 UPCOMING
Audit & Ranking
Oct 1 — Nov 1, 2026
Reproducibility audit of the top-ranked submissions on Oct 1. Final rankings released Nov 1; competition reports and analysis paper drafted.

AuditOct 1, 2026

Final rankingNov 1, 2026
Phase 06 UPCOMING
NeurIPS finals
Dec 11 — 12, 2026
Track winners present in Sydney at the NeurIPS Competition Track. Awards, lessons-learned keynote, and the v2.0 roadmap. Travel grants for top-3 per track.

VenueSydney · ICC

Prize pool$30.000 ($2.500 / top-3 / tracks 1-4)

New in 2026 · BIDS-first via EEGDash

Datasets

Four new datasets. One per cognitive domain.

Each track ships with a brand-new corpus, alongside curated seed datasets. For foundation-model pretraining, EEGDash exposes 700+ additional EEG corpora under the same BIDS-first streaming interface.

Visual decoding NEW

Alljoined eval cohort

New 11-subject evaluation cohort recorded with the same Emotiv hardware and natural-images paradigm as the public Alljoined-1.6M corpus. Released by Alljoined.

BCI command NEW

Graz / BrainHero

20 subjects × 6 sessions, mixed Graz and BrainHero paradigms.

Sleep onset NEW

Muse wearable

~1,000-subject home-wearable EEG with n2_onset annotations.

EMG-to-text NEW

Wristband EMG cohort

New 100-user evaluation cohort recorded with the same wrist-sEMG hardware, typing task, and collection paradigm as the public emg2qwerty corpus. Released by Meta Reality Labs.

ID Name Mod. Subj Hours Ch. SR

things-eeg1 THINGS-EEG1 · Object viewing EEG 50 42h 64 1000 Hz

things-eeg2 THINGS-EEG2 · Object viewing EEG 10 121h 64 1000 Hz

alljoined-1 Alljoined-1 · Natural images EEG 8 12h 64 512 Hz

alljoined-1.6 Alljoined-1.6M · Natural images EEG 20 100h 32 250 Hz

alljoined-eval Alljoined eval · same setup as Alljoined-1.6M NEW EEG 11 — 32 250 Hz

stieger2021 Stieger 2021 · Continuous MI EEG 62 644h 62 1000 Hz

dreyer2023 Dreyer 2023 · Large MI cohort EEG 87 70h 27 512 Hz

zyma2019 Zyma 2019 · Mental tasks EEG 35 3.5h 19 500 Hz

scherer2015 Scherer 2015 · Individually tuned BCI EEG 9 14h 64 250 Hz

bci-graz Graz/BrainHero · BCI command NEW EEG 20 80h 64 500 Hz

sleepedf-ext Sleep-EDF Extended Sleep 78 3,600h 2 100 Hz

physionet2018 PhysioNet Challenge 2018 Sleep 1,983 14,200h 6 200 Hz

hmc-sleep HMC Sleep Staging Sleep 151 1200h 4 256 Hz

muse-sleep Muse Sleep-Onset · Wearable NEW Sleep 1000+ 1000h+ 4 256 Hz

emg2qwerty emg2qwerty · EMG typing EMG 108 346h 32 2000 Hz

emg2qwerty-eval emg2qwerty eval · same setup as emg2qwerty NEW EMG 100 — 32 2000 Hz

14 seed datasets + 4 new releases for 2026 (Graz/BrainHero and Muse Sleep-Onset for training; Alljoined eval and emg2qwerty eval as hidden test cohorts that mirror their public parents) · all under their original providers' licenses · sleep annotations expert-scored, stimulus/keystroke labels machine-derived · leakage held out by subject, session, stimulus, and keystroke · ground truth confidential, top-3 re-run by organisers · NeuralBench-EEG-Core v1.0 · 700+ additional EEG corpora ready via EEGDash Open the start-kit →

Backed by

Sponsors & organizing institutions

Compute, datasets, prizes, and engineering hours from 14 institutions across 5 countries. Yneuro hosts the platform; AWS provides the GPU cloud. Each track has its own sponsor: Alljoined for EEG-to-IMG, Meta FAIR Brain & AI for BCI decoding, InteraXon for sleep onset, and Meta Reality Labs for EMG-to-Text.

Meet the team →

Compute partner

Track sponsors

Organizing institutions

Start-kit drops Jun 1, 2026 · Warm-up Jul 1, 2026

Register your team for NeurIPS 2026.

Five tracks, sealed evaluation, a $30.000 prize pool ($2.500 per top-3 team for tracks 1-4), and a workshop slot in Sydney. Sign up to the mailing list to receive the start-kit link on June 1, the warm-up split on July 1, and the sealed test on August 1.