Independent component analysis

cosmos 4th November 2016 at 2:43pm
Unsupervised learning

aka ICA

Similar to Principal component analysis, but to find the independent components of variation of the data. An example is the "cocktail party problem", where there are several speakers speaking simultaneously, and we want to decompose the signal from some microphones, into the speech of each individual speaker.

Intro video, applications to "cocktail party problem"

Cumulative distribution function

Derivation

video

Example the axes after transforming are the observed xx for each of the microphones. The axes before transforming correspond to each of the speakers! If the data are Gaussian, however, ICA is impossible

Algorithm

video –> algo

Applications of ICA

video