Conclusions - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 170

Chapter 8
Conclusions
The overall objective of this thesis was to research pattern recognition based on the
modelling of the M-dimensional probability density function of the data using
mixtures of independent component analyzers. The proposed methods define a
general framework that is suitable for application to problems that involve com-
plex probability densities. In order to evaluate the potential of the proposed
methods, novel applications in several fields were explored. Therefore, the capa-
bilities of the methods to solve real-world problems has been demonstrated.
This chapter summarizes the research findings, revisiting the specific objectives
given in the Introduction chapter. Section 8.1 reviews the contents of this work,
drawing out the main conclusions that were derived from each chapter. The
contributions of this dissertation are included in Sect. 8.2 . Recommendations for
future research lines are listed in Sect. 8.3 .
8.1 Summary
The first two chapters presented the motivation, problems, and techniques sought in
the thesis. The outlined problems focused on signal classification and blind source
separation (BSS). Thus, the fundamental area of research to deal with these problems
was independent component analysis (ICA) and its extension to mixtures of
ICA models (ICAMM). Two principal methods for classification and hierarchical
clustering which incorporate unsupervised, semi-supervised, and supervised learn-
ing were proposed. These methods were evaluated in diverse applications in order to
solve real-world problems.
ICAMM has established a framework for non-linear processing of data with
complex non-gaussian distributions. Data complexity is captured by a combination
of local linear ICA projections and, thus, the resulting probability density function
of the final model can be used to model class-conditional densities. In addition,