- I received my M.Sc. degree in engineering from the Technical University
of Denmark in the fall of 1998. My thesis concerned independent component
analysis (ICA), where I studied the principle idea and limits. Also I used
the algorithm for analysis of functional brain data in PET and fMRI. I have
resently recived my Ph.D in Modelling tools for shared virtual worlds.
The idea is to apply different statistical modelling techniques to make
it easier to navigate and search virtual worlds like the Internet, and combine
different multimedia techniques.
+45 4525 3920
+45 7742 5506
- Independent Component Analysis (ICA)
The activities around ICA has grown fast the last few years and has
been applied to many different applications and research areas. As
part of my Ph.D. project one objective is to develop tools that can
work in different multimedia environments. At the time being I have
used ICA in separating sound, pictures and text. A small demonstration
file can be downloaded to see ICA preformed on sound and text, click
here, or use the DTU:Toolbox.
Introducing ICA in a Bell MME framework and exploring the limits
and properties. Problems concerning choosing the unlinear function
are looked at experimental by observing the parameter space of the
costfunction, and showing that it only results in a scaling of the
output. Comparing gradient assent, Natural gradient and the Newton
method, different optimization techniques are compared. Especially
by looking into the natural gradient the reason for its good performens
is explained, partially by the algorithm itself and partially by
the parameter space only dependent on the curtosis of the source
signals. Finally ICA is used for decomposing fMRI and PET images
and showing how well the independents criteria works in comparison
to the variance criteria used in PCA. Reducing the number of outputs
using PCA the ICA algorithm shows evidence of clustering the components
most dependent. For more details se publications.
- Text analysis
In the recent years the need to search and maintain and overview
in large amounts of data has grown rapidly. This is not only because
of the Internet but also because of the growing use of computeres
Latent semantic analysis (LSA):
Without having any knowledge of text grammatic and structure in
general the LSA method represents the data in such a way that regular
statistical methods can be applied directly.
Using LSA first, we have used Gaussian-mixture models to cluster
the data both using labeled and unlabeled data. One application
for this method, is to use this structure directly in the Web SHERPA.
For more details se publications.
ICA applied to text:
As already suspected in my master thesis the ICA algoritme can be
used as an clustering method. As such, using LSA and applying ICA
on top of that, can separate groups within the data that results
in most independence between the groups. For more details se publications.
- Virtual worlds
The two main applications for virtual world modelling are presentation
for seismic data and developing tool to help navigate and search.
The later is part of a greater ongoing multimedia project by which
my Ph.D. grant is given.
The project develops a WEB-Site HelpER Personalized Agent (WEB-SHERPA)
which assists navigation in complex virtual environments. We envision
an adaptive agent that can assist in search for specific objects/behaviors
in rich 3D environments. Further, as virtual environments proliferate
we will need help in negotiating these unknown worlds. WEB-SHERPA
shall assist in spatial indexing: by adaptive exploration of a virtual
environment, it provides an annotated, view based, overview of a 3D
environment (a 3D "street map"). For this we will develop adaptive
systems that can resolve spatial and behavioral (incl. emotional)
ambiguities, relate 3D structures to object databases, and communicate
these in terms of a hyperscript street map. In a later stage we envision
WEB-SHERPA to learn from other agents under development, and in this
process be enhanced with enhanced communication skills and visualisation
Try out a demonstration mock-up, click
In order to visualise huge amount of seismic data a VRML model was build
with sliding walls, based on the principle used by Finn Nielsen in
data. Doing so makes it easier to find compartments with possible
oil or gas resources. Also the model serves as a demonstration and
a verification tool. To view screen dumps press 
When developing applications different widgets come to life that other
VRML 2.0 developer might find useful and are free to use. Click
- Software development
In order to supply my income during my engineering studies, I have
been part of a small consortium that made custom software solutions.
The consortium is called CodeGroup and is basically made up of engineers
in the fields of information, electronic and construction. Our current
project is a program called Sigma for calculating overhead and construction
cost in engineering applications. The program is currently available
in Danish, German and English and a free trial version can be downloaded
from the CodeGroup web page.
I have been playing paintball since 1989 and it is one of my biggest
passions. Paintball started out in Denmark in 1988 and has ever
since been growing steady. Im currently involved in the Danish
Paintball Federation (DPF) that helps to develop and structure paintball
in Denmark. Since my studies has been taking up most of my time
the resent years, I had to quit my team the Ugly Ducklings in 1997,
but hope to take up the more serious part of paintball soon again.