NEURIPS 2026(NEURREPS · PROCEEDINGS TRACK)EEG FOUNDATION MODEL

LONG EEG. ANY HEADSET. ONE MODEL.

SAMBA

Learn from long EEG across different headsets.

Jiazhen Hong · Geoffrey Mackellar · Soheila Ghane

Play with SAMBA ↓

Will release soon

SLINEAR TIME

EEG SIGNALS

SHARED SPACE

O(n)linear with sequence length

3Done space for every headset

TSR + DMlearn signals, reduce shared noise

01 / PLAYGROUND

Try the model.
Watch it change.

Move a slider. The visual updates beside it.

LIVE VIEWLong EEG

VISIBLE50%

SHARED SPACE64 ch

SEQUENCE LENGTH

15,360 time steps

Compared with the tested 120 s setting.

Transformer O(n²)

×1.0

SAMBA O(n)

×1.0

SHARED LOG SCALE

RAW EEGMASKEDRECONSTRUCTED

DIFFERENTIAL MAMBA

Remove noise. Keep the EEG.

4 PAIRED HEADS · 8 MAMBA BRANCHES70% illustrative noise amplitude

CURRENT OUTPUTUNDERLYING EEG

Illustration only. Move “Differential heads” to change noise amplitude; the underlying EEG stays fixed.

RESOURCE TREND

Measured → estimated

ESTIMATEDPast reported resource points

GPU memory2,695 MB

Measured: 207 MB at 200; 1,438 MB at 2,000 steps.

Inference / sample0.48 ms

Measured: 0.28 ms at 200; 0.38 ms at 2,000 steps.

SOLID: REPORTED DASHED: ESTIMATED LAST REPORTED POINT

Solid lines use reported measurements. Dashed lines extend the trend; they are not new benchmarks. Paper ↗

02 / THE MODEL

Four steps.
One model.

Tap a step.

SELF-SUPERVISION

Hide EEG spans

Hide part of the signal, then rebuild it.

WHY

Learn the larger pattern.

ΔMAMBA

hidden EEGcontextrebuilt EEG

EPOC X · 14 ch

PAIRWISE 3D DIFFERENCEΔ(x, y, z)MLP → softmax over input channelsX_target = W X_input

PhysioNet · 64 ch

FULL MODELOpen model code↗

03 / SPATIAL MAPPING

Any headset.
One pretraining space.

3D coordinates map each headset into the same space.

INPUT HEADSETEPOC X · 14 chChoose a layout.

INPUT SPACE14 channels

SIMULATED MAPPINGAF3

Peak: Fp1 · 36.1%

Coordinate-distance proxy

ILLUSTRATIVE W COLUMN

SHARED PRETRAINING SPACE64-channel target

LOWHIGH CONTRIBUTION

SIMULATION ONLY · NO CHECKPOINT WEIGHTS · REPOSITORY COORDINATES; MN8 IS SCHEMATIC.

REAL COORDINATESOpen montage files↗

04 / RESULTS

What SAMBA
learns.

Pick a task to see where it focuses.

Animated spatial weights for eyes open and closed

FRONTAL REGION

Eyes open / closed

Focuses on frontal signals linked to eye closure.

EXPLORE

Go deeper.

01CodeGitHub ↗02PaperarXiv ↗03Electrode layoutsData ↗

CITE SAMBABIBTEX + PLAIN TEXT

Hong, J., Mackellar, G., & Ghane, S. (2026). SAMBA: Toward an EEG Foundation Model via Differential Mamba and Coordinate-Driven Spatial Embedding. NeurReps 2026 Proceedings Track.