Nonlinear projections
|
Tools for Non-linear Data Analysis
|
Objective
The study and development of methods for non-linear data analysis resulting
in a software toolbox that may be used in industry. Applications generally
include data compression and feature extraction, both can be applied in
many practical applications. Our interest is both in methods to acquire
one latent variable model for the total original space, as well as in models
that involve a combination of several 'local' latent variable models (possibly
related to each other).
Research group members
dr. Dick de Ridder (Delft University)
dr. N. Vlassis
drs. J.J. Verbeek
dr. ir. B. Kröse
Funding
The project is funded by Stichting Technische
Wetenschappen
Motivation
Current computerized measurement systems and data acquisition systems
deliver a huge amount of data. For example, in the petrophysical industry
more and more advanced measuring devices are used to determine the characteristics
of the borehole. Because the sensors are often measuring on the same physical
phenomenon (in the above application for example the porosity), the intrinsic
dimensionality of the data will in many cases be lower than the dimensionality
of the data itself and only depend on the degrees of freedom of the observed
phenomenon. If the dimensionality of the measurement space is not reduced
correspondingly by some mapping, the outcomes of any analysis of the measurements
may suffer from an increased noise resulting from more sensor signals,
instead of taking advantage of the increased information or resolution.
Feature extraction and feature reduction thereby become more and more important
in relation with increasing sensor capabilities. However, standard analysis
packages are often limited to linear projections, while the data not necessarily
resides on a linear manifold.
Recently, a number of novel promising techniques for nonlinear projections
were proposed by the involved groups (Intelligent Autonomous Systems at
University of Amsterdam and Pattern Recognition at Delft University). The
techniques will be further studied and elaborated, and novel methods will
emerge per case. Depending on the application (visualisation, compression
or classification) we will define criteria to assess the performance. All
methods will be tested on these criteria and on speed.
There is an existing collaboration with the following users: Shell with
applications in the analysis of petrophysical and seismic data, TNO-FEL
with applications in the classification of radar profiles, Noldus Information
Technology with applications in the analysis of behavioural data, KiQ and
Cap Gemini with applications in the analysis of time series, and Unilever
with applications in visualization of physical processes. To enable an
easy and broad utilisation we will implement the developed methods in a
toolbox, compatible with a standard data analysis software package (for
example Matlab or SPSS). For the exploitation of such a toolbox we could
use another user in the users group.
| 