Zyphra/zuna

Python package for Zuna, an EEG foundation model for inference.

What it solves

ZUNA1.1 is designed to fix common issues in EEG (electroencephalogram) recordings, such as noisy data, missing channels (dead electrodes), or sparse electrode layouts. It allows researchers to denoise signals, reconstruct dropped channels, and upsample sparse montages to denser ones without needing to retrain the model for different hardware setups.

How it works

ZUNA1.1 is a 380M-parameter transformer encoder-decoder diffusion autoencoder. It processes EEG signals by slicing them into 0.125-second segments and treating them as tokens. Unlike traditional models that rely on a fixed list of channels, ZUNA uses 4D rotary positional encodings (x, y, z, t) based on the electrode's 3D scalp coordinates. This makes the model "channel-agnostic," meaning it can synthesize signals at any location on the scalp, regardless of the number or layout of the electrodes used during recording.

Who it’s for

It is intended for EEG researchers and BCI (Brain-Computer Interface) developers who need to clean and augment their neural data. It is explicitly not a diagnostic tool for clinical medical use.

Highlights

  • Flexible Montage Support: Works with any layout, from 4-channel headbands to 256-channel research caps.
  • Efficient Inference: Requires less than 1 GB of VRAM, allowing it to run on consumer GPUs, Macs (Apple Silicon), or CPUs.
  • Robust Training: Trained on 3.5M channel-hours of public EEG using four different dropout schemes to simulate real-world corruption (whole-channel, full-time, channel-time, and random-uniform).
  • Variable-length Inputs: Supports signal segments ranging from 0.5 to 30 seconds.
  • Easy Integration: Provides a Python API and a browser-based playground for quick denoising and upsampling.

Related

  • Project
    wjq-learning/CBraMod

    A foundation model for EEG decoding designed for clinical and Brain-Computer Interface (BCI) applications through a pretraining and fine-tuning pipeline.

  • Project
    rwth-i6/returnn

    An extensible training framework for universal recurrent neural networks optimized for fast multi-GPU training and decoding.

  • Project
    braindecode/braindecode

    Braindecode is a Python toolbox that brings deep‑learning models to raw brain‑signal data (EEG/ECoG/MEG). It provides dataset loaders, preprocessing, several ready‑made neural‑network architectures, and visualisation utilities, all built on top of PyTorch and MNE‑Python. Installable via `pip install braindecode`, it targets neuroscientists and ML researchers working with neurophysiological data.

  • Project
    MrGiovanni/ModelsGenesis

    A self-supervised learning framework that pre-trains 3D models on unlabeled CT and MRI scans to create foundation models for medical image segmentation and classification.

  • Project
    InternLM/xtuner

    A next-generation LLM training engine optimized for ultra-large-scale MoE models up to 1T parameters, supporting multimodal training and advanced reinforcement learning.