- [[generalized eigendecomposition|GED]], [[principal component analysis|PCA]], [[independent component analysis|ICA]] - [[mutual information]], [[conceptual definitions of entropy]] # Idea BSS methods are used to separate source signals (e.g., brain, noise, non-brain) that are mixed (e.g., by volume conduction in EEG and other electrophysiological recordings). The "blind" in blind source separation indicates that we lack precise information about the channels or the sources. We can only make general statistical assumptions on the sources or the structure of the channels. Given the [[forward model for EEG|forward model]] $x = As + n$, blind source separation methods estimate both $A$ and $s$ jointly. ICA and BSS algorithms differ from PCA because they identify distinct sources of **information** in the data instead of **orthogonal directions of maximal variance in the data**. Thus, BSS algorithms can bypass PCA's spatial orthogonality constraint. By design, blind source separation methods are [[unsupervised machine learning]] methods. They don't make use of class labels but optimize other objectives like statistical independence. # References - [[Dahne 2014 SPoC - relating neural oscillations to behavior parameters]]