Consensus and group decoding

The algorithmic core of a collaborative BCI is how to fuse several users' neural evidence into one decision. Simple schemes average event-related potentials or take a majority vote over per-user classifications; better schemes weight each user by an estimated reliability (confidence, past accuracy, or single-trial signal quality) so that a strong signal from one participant can override weak signals from others. Error-related potentials from observers can also be pooled as a group 'that was wrong' signal.

Fusion helps most when the per-user errors are unbiased and weakly correlated, the regime in which combining independent estimators reduces variance; it can hurt when a shared confound — a common visual artifact, a leader effect — makes errors correlated, so a confident majority is confidently wrong together. Reliability-weighted fusion is therefore preferred over naive voting, and the same trust-weighting idea reappears inside multi-brain interfaces like BrainNet.