Blind separation of sources (Independent Component Analysis)

• The Problem

  The task is to separate an unknown number N of unknown sources S from M measured linear mixtures of signals.

• Complete source case (M=N)

  For solving the blind source separation problem we have developed a multi-layer learning algorithm with a biologically justified local learning rule (Cichocki:Kasprzak:Amari:NOLTA95).

  

  Fig. 6 Multi-layer feed-forward (top) and feedback (recurrent) (bottom drawing) neural networks with global and local learning algorithms (LA) for blind separation of sources.

  For comparison a very robust, recently published global learning rule (Amari:Cichocki:Yang:NIPS95) has also been tested on image sources.

  An example of blind separation of image sources if M=N=5:
  

  These are the original images which are unknown to the neural net:

  MONO JPEG (121 kB). or COLOR JPEG (149 kB).
  

  Here are five mixed images from the input of the neural net:

  MONO JPEG (168 kB). or COLOR JPEG (215 kB).

  A blind separation process for above five mixtures of four natural images and additional noise image is demonstrated by following movies:

  1. MONOMPEG (1.97 MB).
  Its longer version (MPEG, 3.56 MB !) (with double sampling frequency ) is also available.
  2. COLORMPEG (2.59 MB).
  Be careful with its longer version (MPEG, 4.59 MB !) (double sampling frequency).

  ---

  [go back to top]

  ---