File size: 3,704 Bytes
601b8e0 0965a1f 601b8e0 0965a1f 601b8e0 0965a1f 601b8e0 0965a1f 601b8e0 0965a1f 601b8e0 0965a1f 601b8e0 0965a1f 601b8e0 0965a1f e4af7e0 0965a1f e553287 0965a1f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 | ---
pipeline_tag: feature-extraction
tags:
- eeg
- brain-computer-interface
- self-supervised-learning
- mamba
- pytorch
---
# SAMBA-2s
[Project website](https://samba.dcode.ai/) 路 [Code](https://github.com/Jiazhen-Hong/SAMBA) 路 [Paper](https://arxiv.org/abs/2511.18571) 路 [SAMBA-100s](https://huggingface.co/Dcode-AI/SAMBA-100s)
SAMBA-2s was pretrained on 2-second EEG windows from Emotiv (Alpha, Attention, Crowdsource, DriverDistraction, and STEW), HBN resting state, Muse sleep onset, and DICE eyes closed. Pretraining used 256 samples per window at 128 Hz and 4, 14, 19, and 129 EEG channels. The encoder supports other window lengths and electrode layouts.
## Download
[Download config](https://huggingface.co/Dcode-AI/SAMBA-2s/resolve/main/config.json?download=true) 路 [Download weights](https://huggingface.co/Dcode-AI/SAMBA-2s/resolve/main/model.safetensors?download=true)
Keep both files in the same directory. To download them together from a terminal:
```bash
pip install -U huggingface_hub
hf download Dcode-AI/SAMBA-2s config.json model.safetensors --local-dir ./SAMBA-2s
```
## Usage
Install SAMBA in a CUDA environment:
```bash
git clone https://github.com/Jiazhen-Hong/SAMBA.git
cd SAMBA
pip install -e '.[mamba,hub]'
```
Load the downloaded directory. If you skipped the download step, the following example fetches both files automatically:
```python
import torch
from huggingface_hub import snapshot_download
from samba_eeg import SambaModel
folder = snapshot_download(
"Dcode-AI/SAMBA-2s",
allow_patterns=["config.json", "model.safetensors"],
)
model = SambaModel.from_pretrained(folder).cuda().eval()
```
For files you already downloaded, pass their local directory to `SambaModel.from_pretrained` in place of `folder`.
Prepare EEG at 128 Hz. Use `signals` shaped `[batch, channels, samples]` and electrode `positions` shaped `[channels, 3]` or `[batch, channels, 3]`. Match the signal channel order and use the same coordinate frame and units as the target positions in `config.json`. The decoder reconstructs EEG at the input electrode coordinates.
```python
signals = signals.to(device="cuda", dtype=torch.float32)
positions = positions.to(device="cuda", dtype=torch.float32)
# Skip these two lines if the signals are already normalized this way.
scale = torch.quantile(signals.abs(), 0.95, dim=-1, keepdim=True)
signals = signals / scale.clamp_min(1e-8)
with torch.inference_mode():
features = model.encode(signals, positions)
pooled_features = features.mean(dim=1)
```
`features` has shape `[batch, floor(samples / 8), 128]`. Use `pooled_features` for linear probing.
## Learned EEG spatial weights
SAMBA uses electrode coordinates to learn how to combine EEG channels. These maps show how the spatial weights change during training.
| Eyes open vs. closed | Driver distraction | Motor imagery |
|:---:|:---:|:---:|
| <img src="https://raw.githubusercontent.com/Jiazhen-Hong/SAMBA/main/Figures/Topo_gif/EyesState_GIF.gif" width="250" alt="Eyes open vs. closed"> | <img src="https://raw.githubusercontent.com/Jiazhen-Hong/SAMBA/main/Figures/Topo_gif/DriverDistraction_GIF.gif" width="238" alt="Driver distraction"> | <img src="https://raw.githubusercontent.com/Jiazhen-Hong/SAMBA/main/Figures/Topo_gif/MI_GIF.gif" width="245" alt="Motor imagery"> |
## Citation
```bibtex
@inproceedings{hong2026samba,
title = {{SAMBA}: Toward an {EEG} Foundation Model via Differential Mamba and Coordinate-Driven Spatial Embedding},
author = {Hong, Jiazhen and Mackellar, Geoffrey and Ghane, Soheila},
booktitle = {NeurReps 2026: Symmetry and Geometry in Neural Representations (Proceedings Track)},
year = {2026}
}
```
|