An electroencephalogram (EEG) is like having a bunch of microphones placed all over your head. Each of these “microphones”—a sensor — listens in on the electrical chatter of neurons. But sometimes one of them peels off, gets noisy, or records interference instead of a signal. In the past, we'd patch up the gaps with a simple mathematical brush, the spherical interpolation method, which smears neighboring signals into the empty space. You'd get sound, but the details were lost.
The new model, ZUNA1.1, works differently. It's a neural-network restorer trained on thousands of samples of brain activity. Instead of just blending adjacent channels, it reconstructs a plausible wave pattern based on the overall picture. If the old method is like smudging a scratch on a photo with your finger, ZUNA1.1 is like an artist who fills in the missing piece in the style of the whole canvas.
The main advantage is flexibility. ZUNA1.1 isn't tied to a fixed sensor grid. You can feed it a recording from 16 electrodes, then one from 64, and it figures out what to do. You can ask it to restore just one second within a long recording, and it will. It's like a restorer who can work with any canvas and any brush, without asking for instructions.
Such a tool isn't just for cleaning up scientific data. It could make EEG-based seizure detectors or brain-computer interfaces more reliable, where losing even a single sensor breaks control. By restoring lost signals, ZUNA1.1 helps us hear the brain where there was only noise before.
The standard interpolation method that ZUNA1.1 was compared against is built into the popular EEG analysis package MNE, the one used by nearly all neurophysiologists worldwide.
In William Gibson's novel Neuromancer, characters jack into cyberspace through electrodes on their heads, and losing a signal would mean digital death.