Introduction to Independent Components Analysis

                         Barak A. Pearlmutter
                   Hamilton Institute, NUI Maynooth
                   http://www-bcl.cs.may.ie/~barak/

Survival often depends on responding appropriately to potential
threats, food sources and mates (e.g. at a cocktail party), while at
the same time ignoring the many irrelevant sound sources that may
constitute the majority of the acoustic energy received.  Source
separation of this sort is an important problem not only in acoustics,
but across a variety of domains in which multiple sensors record a
mixture of signals.  Doing this without the use of strong a-priori
knowledge of the sources or mixing process is called "Blind Source
Separation," and there is a rapidly growing suite of algorithms for
BSS based on disparate principles ranging from sparseness to maximum
likelihood.  BSS is finding application in a variety of disciplines,
including acoustics and brain imaging.  This talk will explain how BSS
is possible at all, will give some intuition into the major families
of algorithms and their derivations, and will explore the special case
often called "Independent Components Analysis".  We will conclude with
some demonstrations and a discussion of practical considerations.