Researchers working with EEG often face variable recording lengths, missing channels, noisy segments, and diverse electrode layouts. Traditional models require fixed-length inputs and specific montages, forcing costly preprocessing or data discard. ZUNA1.1 addresses these pain points directly. It is a 380M-parameter masked diffusion auto-encoder that accepts any length from half a second to thirty seconds, works with arbitrary channel counts, and can reconstruct or predict signals at unseen scalp positions. The model uses 4D rotary positional encoding tied to electrode coordinates, so it is truly channel-agnostic. Trained on four realistic dropout patterns—whole-channel loss, time-stretch blocks, clustered gaps, and scattered point loss—it learns to denoise, fill missing data, and upsample across layouts without retraining. On held-out tests ZUNA1.1 matches or improves the reconstruction error of its predecessor while outperforming classic spline interpolation, especially when entire brain regions are missing. Weights are freely available under Apache 2.0, inference runs on a consumer GPU and tolerates CPU for many workloads. Install via pip, run the provided preprocessing scripts, and plug the model into your pipeline to handle real-world EEG variability out of the box. No need to re-segment, re-reference, or discard usable trials.
#AI #Product #EEG #MachineLearning #OpenSource #Neurotech