Every day, the brain sounds a little different. Yesterday, the thought of moving your left hand produced one pattern of electrical waves; today, it's slightly different. For a brain-computer interface (a device that turns mental commands into actions), this is a big problem: yesterday's settings don't work today.
The new method, called MRieHy, doesn't learn from a single recording but from a whole collection of brain activity "fingerprints" gathered over many days. First, it finds the common center of these fingerprints, like creating an averaged portrait from dozens of photos of the same person, making it easier to match new snapshots. For this, it uses Riemannian geometry, a way to measure distances between the complex matrices that describe the brain's electrical activity. Then it builds a hypergraph: a network where connections can link not just two points but several at once, like a group chat rather than a phone call. One hypergraph is built from geometric distances between fingerprints, another from the similarity of deep features extracted from the signals. The two networks are combined with weights that are tuned automatically.
During real-time use, the system keeps a buffer of the most recent brain fingerprints, aligns them with the averaged portrait, and instantly outputs a command. Tests on data from implanted electrodes and standard EEG caps showed that recognition accuracy for four types of imagined movements is higher than with the best previous methods.
This approach makes the neural interface like an experienced interpreter who has heard the same voice in many moods and therefore understands it even with a slight rasp.
The brain signals of the same person can differ more than those of two different people on the same day, which is why interfaces need daily adaptation.