Graphical Models Project

1 Introduction
Signal processing in general, and digital signal processing in particular is a wide
and developped study eld. We have powerful tools to process, analyse and
interpret signals. However, these tools work only when we extract a particular
signal. What would happen if the only thing we could observe is a mix of
dierent signals ? This problem is known as the cocktail − party problem : we
are in a room full of people and we hear dierent conversations occur, and we
want to extract separately the dierent conversations. It seems infeasible at
rst sight, but under some assumptions it can be done :
• The number of signals and observations are equal.
• The observations are linear combinations of the signals.
• The signals are independent.

• The signals are nongaussian.

2 ICA and FastICA

The ICA is a method that can separate n independent signals sj (t) from n
observations xj (t) where x(t) = As(t). The method is based on the fact that
the signals are nongaussian, so the goal is to maximize the nongaussianity or
independence (it is equivalent) of wjT x(t) with wj ∈ Rn .
The FastICA is an algorithm that take as inputs x(t) and s(t) and outputs
W s.t W x(t) = s(t). It's based on the maximization of an approximation of
negentropy and a Gram-Schmidt decorrelation of the output signals.

3 Data
The data used in this project are ECG (electrocardiogram) records from dier-
ent patients. The ECG is done with 9 electrodes which generate 12 temporal
series that we can view as a mix of dierent independent signals from the body.
Our goal is to separate the dierent signals from the body and estimate their
importance in the ECG.
We started by preproccessing the data (centering and whitening).