Information Technology Reference
In-Depth Information
8.3.2 Extension of Mixca to Other Methods
• Extension to sequential methods. The proposed method uses block estimation
of the ICAMM parameters whereas other methods employ sequential esti-
mation of the parameters for each ICA. The extensions developed in this
thesis can be used to generalize ICAMM sequential methods. The estimation
of parameters such as stopping criterion should be improved since they are
usually estimated arbitrarily in current techniques.
• Incorporation of on-line estimation of the parameters. The on-line learning
processing would perform simultaneous structure and parameter identifica-
tion. This kind of learning allows long processes to be monitored on-line.
This assumes a dynamic modelling that can be supported by HMM (as we
propose
in
this
thesis)
for
incorporation
of
sequential
dependencies
in
ICAMM.
• Development of a method for prediction based on ICAMM. This is an
interesting issue that considers the strategy of linear local projections that can
be adapted to partial segments of a data set while maintaining generalization
given the mixture of several ICAs. The resulting algorithm could be applied
to time series prediction, to recover missing data in images, etc.
8.3.3 Other Applications
• There is a myriad of possible applications where a modelling based on
mixtures of ICAs can be valuable. For instance, event-related dynamics of
brain oscillations to study sensorimotor processes, changes during the per-
formance of cognitive tasks, or neurological disorders; development of query
by visual example (QBVE) and query by semantic example (QBSE) systems;
and to investigate in ICAMM-based physical models for NDT (associating
features extracted from the NDT signals with physical properties of the
materials).
Search WWH ::
Custom Search