|
Neural networks are increasingly popular tools for modeling of complex dynamics, noisy
signal processing and pattern recognition problems. Neural
networks are often considered as so-called black box models. They are indeed very well
suited for modeling systems in which the underlying rules are hard to reveal. Neural nets
learn statistical relations from observations rather than relying on algorithmic "solutions".
Most standard neural network architectures posses the property of being universal learners,
i.e., by choosing the architecture carefully it is possible to learn any task (static relation).
The network provides a relation from a set of input to a set of output variables, hence
captures aspects of the distribution of the output variables - conditioned on the
inputs.
The network is trained to implement the desired relation by minimizing a performance measure or
cost function with respect to the network parameters on the set of training data consisting of
input-output examples. Typical cost functions are mean square errors (for regression or
function approximation) and the entropic error measure (for pattern recognition nets).
Costfunctions can be derived applying maximum likelihood methods or using the so-called
Baysian framework.
More background on neural networks:
Neural
Network Links.
CONNECT
The group is a founding member of the
Computational Neural Network Center (CONNECT) which was established
March 1, 1991. The main objective of the research center is to
promote and support the collaboration between Danish researchers in
theory, implementation and application of neural computation as well
as to establish graduate level training in neural networks. Today
CONNECT is centered around two projects: a
theory project at the Niels Bohr
Institute and a signal processing project under the THOR center.
For more information, see CONNECT.
Design and evaluation of adaptive systems
Computer intensive methods for design and evaluation
of nonlinear adaptive systems, e.g., neural networks.
The most salient aspect of an adaptive system is it ability to generalize.
Generalization is the quantitative measure of performance on fresh data.
Robust estimates of generalization are not easily obtained,
data resampling schemes, such as crossvalidation and bootstrap are
promising candidates.
Application specific architectures
Pruning and regularization are used jointly
for overfit control for neural
networks and other adaptive model families. When pruning a neural network
an initial oversized architecture is trained and by successive
elimination of
connections a nested family of models is created.
Applications in both system identification
and pattern recognition have been demonstrated
Neural network ensembles
Ensembles (decision by committee) is a very active research topic.
We coined the term ensemble in Hansen & Salamon: "Neural Network Ensembles,"
IEEE T-PAMI, October 1990.
Using an ensemble we combine several neural network predictions
by voting (for classifiers) or by an averaging scheme (for
function approximation).
When combined with resampling plans - where ensemble members
get different training sets - new schemes like stacking, boosting
and bagging result.
Data Mining
Neural networks have shown their ability to summarize large databases
in many real world applications. In this project we intend to formalize
the use of neural nets in datamining. We will investigate means for
extracting knowledge from trained networks. The objective is to model
the human ability for symbol definition and binding of concepts.
Our primary case for this activity will be functional neuroimaging
where we will explore the binding of many different types of neuroscience
knowledge: summary statistics from past experiments, brain atlases,
published analyses, textbooks etc. Metaexperiments -
computer aided "literature studies" combining information
from large databases in new ways - are of high interest
Mine clearance
In collaboration with Dr. Ole Nymann, Center for Advanced Technology,
Assoc. Prof. Kaj Jackobsen, Assoc. Helge Sørensen Dept. of Applied Electronics, DTU and
Assoc. Prof. Olav Breinbjerg, Dept. of Electromagnetic Systems, DTU
we aim to develop an automated mine clearance system.
The project deals with basic signal and image processing components, statistical detection
device as well as sensor fusion in a mine clearance system. The
key component is the scene model which constitutes the basic objective for data processing
conceived as a hyperdocument.
It conveys a 3D representation of potential mines,
topographical data, physical data and other recognized objects in the field of view. The scene
model interacts with a knowledge database, a statistical detection device, and a user interface.
Basic properties and elements of the statistical detection device as well as sensor fusion are
also discussed.
For more information, see Nordic Demining Research Forum homepage.
Independent component analysis
Independent Component Analysis (ICA) is a new signal processing research field concerned with unsupervised
learning of independent effects in multivariate signals.
Further information.
Gaussian Processes
Gaussian processes is a revived Bayesian nonparametric technique,
well-suited for regression problems.
Gaussian process resources..
|
