feature extraction

Raw brain signal is a torrent of numbers — thousands of tiny voltage readings every second, most of it noise and detail a decoder does not need. Feature extraction is the step that boils that flood down to a handful of informative numbers, called features, that actually capture what matters. It is like describing a song not by listing every sound wave but by saying "fast tempo, lots of bass" — a few summaries that hold the useful gist.

Which features you choose depends on the BCI. For rhythm-based systems you might compute the band power in a couple of frequency bands; for systems that watch for a brain response to a flash or a beep, you might measure the height and timing of a particular peak in the waveform. The aim is always the same: keep the parts of the signal that distinguish the user's intentions and throw away the rest.

Good features make the decoder's job dramatically easier, because the hard pattern-finding has already been done. In classic BCIs a human designs the features by hand; some modern systems instead let a neural network learn its own features straight from the signal.