What is Independent Component Analysis?
Independent component analysis, or ICA, attempts to break multivariate signals down into subcomponents, specifically independent non-Gaussian signals. If N sources are present, then to recover the original signals, N observations must be made. ICA is incredibly accurate as long as the source of the signals are actually independent of each other, and the values in each source have non-Gaussian distributions. ICA is often represented in a matrix:
Where A is an unknown, square, invertible mixing matrix.
A Practical Example
Since one of the most popular applications of ICA is when it comes to noise and sound, the most well-known and simple examples is something called “the cocktail party effect”. The cocktail party effect is when the brain can focus on a single voice in a room, even if the room is noisy and makes it difficult to hear. Our brain can separate the noise into different streams and then decide which streams are the most important. The purpose of ICA is to copy this phenomenon and seperate all signals from different sources - in this problem, individual voices, music, and other noises - from each other, and maximize independence among features.
Application in AI
ICA is, at its very basic level, a component of pattern recognition. When we apply ICA to the idea of identifying faces, it will find specific selectors (since it maximizes independence of features). These selectors could be a hair, mouth, or nose selector.