Zyphra Releases ZUNA1.1 EEG Model Under Apache 2.0

Zyphra released ZUNA1.1 this week, an open-source electroencephalography (EEG) foundation model available under the Apache 2.0 license. This updated model builds upon ZUNA1, offering enhanced flexibility by accepting variable-length inputs ranging from 0.5 to 30 seconds, a significant improvement over ZUNA1's fixed five-second segments. ZUNA1.1 is designed to reconstruct, denoise, and upsample EEG data across arbitrary channel layouts, addressing the inherent messiness and variability of real-world EEG recordings.

The ZUNA1.1 model features 380 million parameters and functions as a masked diffusion autoencoder for scalp-EEG signals. It can denoise existing EEG segments and channels, reconstruct missing ones, and predict novel channel signals based on physical coordinates on the scalp. The model's parameter count remains unchanged from ZUNA1, and it is capable of running on consumer GPUs, with acceptable performance on CPUs for many tasks. Zyphra has made the model weights available on Hugging Face and the inference and preprocessing code on GitHub, with installation facilitated by `pip install zuna`.

Zyphra also provides a free browser-based EEG Playground for users. The architectural innovation enabling ZUNA1.1's flexibility lies in its tokenization and positional encoding. The model, a transformer encoder-decoder diffusion autoencoder, segments each channel into 0.125-second intervals (32 samples at 256 Hz), with each segment treated as a continuous-valued token. These tokens are serialized in channel-by-time order. Crucially, each token incorporates a 4D rotary positional encoding that includes the electrode's 3D scalp coordinate and a coarse-time index. This approach, which relies on spatial position rather than array index, makes ZUNA channel-agnostic, allowing it to process any electrode layout and generate signals at unrecorded positions, thereby enabling arbitrary channel upsampling by location. All ZUNA1.1 components are released for research use only.