ICA

ICA separates a multi-channel Signal into independent components using FastICA. Where PCA finds the directions of greatest variance, ICA looks for the directions that are statistically independent — which is what you want when several sources have been mixed together on the way to your sensors.

Sources are the estimated independent signals at unit variance; Mixing is the matrix that recombines them (rows are components, columns are channels).

It is warm-started: each update runs Iterations of the fixed-point algorithm from the previous unmixing matrix, so components stay put from frame to frame instead of being re-derived and re-ordered. Converged says the step was tiny and Change reports how much it moved.

This node is marked experimental. It works for linear, instantaneous mixtures — blinks in EEG, two tones on two microphones — and handles up to 16 channels.

• Remove blink artefacts — Separate an EEG window, identify the blink component, and subtract it.
• Unmix two sources — Two microphones hearing two things gives, in the ideal case, two clean signals.
• Find hidden drivers — Components that no single channel shows on its own.
• Compare with PCA — Run both on the same window; where they disagree is informative.

• ICA cannot separate more sources than you have channels, and it assumes the mixing is linear and instantaneous.
• Component order and sign are inherently arbitrary in ICA; the warm start keeps them stable, not canonical.
• Watch Converged. An un-converged ICA is not a result.
• Band-pass and detrend first — ICA is sensitive to drift.
• Reset when the source changes, or the warm start will drag the old solution into the new data.