ICA is the name for a family of techniques which in recent years have been producing interesting results on multi- or single-source audio data, natural images, and biomedical data such as brain recordings (EEG) and brain-scans
(fMRI). It is a simple, general-purpose method for transforming ensembles of multivariate data into 'natural' coordinate systems. New applications are being found all the time.

What it does is very simple. Like PCA (Principal Component Analysis), it finds a (linear) co-ordinate system for the data, transforming it by a matrix, W, into a new basis set. However, unlike PCA, the transformation is free to be non-orthogonal (whoever said co-ordinates should be at right-angles to each other?). Furthermore, the exact method for choosing
these axes is sensitive to statistics of all orders, while in PCA, only the second-order statistics of the covariance matrix are used.

When the data are transformed by W, the resulting output variables are supposed to be as *statistically independent* as possible. Examples I will give are:

1. separate people speaking in a mixed-audio environment
2. eyeblink and noise artifacts in EEG recordings
3. localised and oriented 'edge-patches' as the independent generators of natural images.

I will discuss relations to projection pursuit, related to the issue of sparse coding.