ICA (Independent Component Analysis)

Independent component analysis (ICA) is a method for separating a signal made of several mixed sources into its original components, assuming they are statistically independent and, except for at most one, non-Gaussian. Its typical goal is to recover the “source” signals from the mixtures that are observed, without knowing beforehand how they were combined; that is why it is called blind source separation.

The cocktail party problem

The canonical example is the cocktail party problem: in a room with several people talking at once and several microphones, each microphone captures a different mixture of all the voices. ICA starts only from those mixed recordings and separates the individual voices, estimating both how they were mixed and what the original sources were.

How it differs from PCA

ICA should not be confused with principal component analysis (PCA). PCA only seeks uncorrelated, orthogonal components—that is, second-order independence; ICA seeks a stronger statistical independence, using higher-order statistics, and can therefore separate sources that PCA cannot. In fact, PCA is often used as a preliminary step (the “whitening”) before applying ICA.

What it is used for

ICA is applied to audio source separation, to biomedical signal processing—for example, to isolate artifacts in recordings of brain activity—and to feature extraction in images and sound.