EEG artifact rejection

a self-supervised approach to cleaning neural signals — specifically removing artifacts (noise from eye movements, muscle activity, electrode movement, power line interference) from EEG data without requiring labeled training data. artifact rejection is one of the biggest pain points in EEG research and BCI applications; bad cleaning loses real signal, and manual cleaning is extremely time-consuming.

the self-supervised framing is the interesting technical angle. instead of training a model to classify "artifact" vs "clean" with labeled examples (which require expert annotation), you use the structure of EEG itself — temporal consistency, spatial correlations between electrodes, known frequency properties of real neural signals — as a self-supervision signal. the model learns what "clean" looks like without being told, then can detect and remove deviations. similar approaches have worked well in other time-series domains.

this sits at the intersection of several interests: ML for biosignals, hardware for sensing (→ sensor capturer), and BCI more broadly (→ EMG bracelet, pupilometry glasses). the research angle would fit well as a paper or an open-source tool — there's a clear gap in the literature for robust self-supervised artifact rejection that generalizes across datasets and electrode configurations.

related: symbolic regression