Abstracts
[PhD Thesis]
[Lic. Thesis]
[Editorial work]
[Journal articles]
[Conference papers]
[Technical reports]
Learning Multidimensional Signal Processing
PhD thesis
May, 1998
Linköping Studies in Science and Technology. Dissertation No. 531
Abstract
The subject of this dissertation is to show how learning can be used for
multidimensional signal processing, in particular computer vision.
Learning is a wide concept, but it can generally be defined as a
system's change of behaviour in order to improve its performance in
some sense.
Learning systems can be divided into three classes: supervised learning,
reinforcement learning and unsupervised learning. Supervised learning
requires a set of training data with correct answers and can be seen as a
kind of function approximation. A reinforcement learning system does
not require a set of answers. It learns by maximizing a scalar feedback
signal indicating the system's performance. Unsupervised learning can
be seen as a way of finding a good representation of the input signals
according to a given criterion.
In learning and signal processing, the choice of signal representation
is a central issue. For high-dimensional signals, dimensionality
reduction is often necessary. It is then important not to discard
useful information. For this reason, learning methods based on
maximizing mutual information are particularly interesting.
A properly chosen data representation allows local linear models to be
used in learning systems. Such models have the advantage of having
a small number of parameters and can for this reason be estimated by using
relatively few samples. An interesting method that can be used to
estimate local linear models is canonical correlation analysis (CCA).
CCA is strongly related to mutual information. The relation between
CCA and three other linear methods is discussed. These methods are
principal component analysis (PCA), partial least squares (PLS) and
multivariate linear regression (MLR). An iterative method for CCA,
PCA, PLS and MLR, in particular low-rank versions of these methods,
is presented.
A novel method for learning filters for multidimensional signal
processing using CCA is presented. By showing the system signals in
pairs, the filters can be adapted to detect certain
features and to be invariant to others. A new method for local
orientation estimation has been developed using this principle. This
method is significantly less sensitive to noise than previously used
methods.
Finally, a novel stereo algorithm is presented. This algorithm uses
CCA and phase analysis to detect the disparity in stereo images. The
algorithm adapts filters in each local neighbourhood of the image in a
way which maximizes the correlation between the filtered images. The
adapted filters are then analysed to find the disparity. This is done
by a simple phase analysis of the scalar product of the filters. The
algorithm can even handle cases where the images have different
scales. The algorithm can also handle depth discontinuities and give
multiple depth estimates for semi-transparent images.
Click here to get the full thesis in portable data format (pdf) (3108 k)
Reinforcement Learning Using Local Adaptive Models
Licentiate in Technology Thesis
August, 1995
LIU-TEK-LIC-1995:39
Abstract
In this thesis, the theory of reinforcement learning is described and
its relation to learning in biological systems is discussed. Some
basic issues in reinforcement learning, the credit assignment problem
and perceptual aliasing, are considered. The methods of temporal
difference are described. Three important design issues are discussed:
information representation and system architecture, rules for
improving the behaviour and rules for the reward mechanisms. The use
of local adaptive models in reinforcement learning is suggested and
exemplified by some experiments. This idea is behind all the work
presented in this thesis.
A method for learning to predict the
reward called the prediction matrix memory is presented. This
structure is similar to the correlation matrix memory but differs in
that it is not only able to generate responses to given stimuli but
also to predict the rewards in reinforcement learning. The prediction
matrix memory uses the channel representation, which is also
described. A dynamic binary tree structure that uses the prediction
matrix memories as local adaptive models is presented.
The theory
of canonical correlation is described and its relation to the
generalized eigenproblem is discussed. It is argued that the
directions of canonical correlations can be used as linear models in
the input and output spaces respectively in order to represent input
and output signals that are maximally correlated. It is also argued
that this is a better representation in a response generating system
than, for example, principal component analysis since the energy of
the signals has nothing to do with their importance for the response
generation. An iterative method for finding the canonical
correlations is presented. Finally, the possibility of using the
canonical correlation for Learning response generation in a
reinforcement learning system is indicated.
Click here to get the full thesis in portable data format (pdf) (1139 k)
Editorial work:
Artificial Neural Networks in Medicine and Biology
Helge Malmgren,
Magnus Borga,
Lars Niklasson (Editors)
Proceedings of the ANNIMAB-1 Conference,
Göteborg, Sweden,
May, 2000
In series Perspectives in Neural Computing
ISSN 1431-6854
ISBN 1-85233-289-1
Springer-Verlag London
The book can be ordered from The ANNIMAB Society
Abstract
This book contains the proceedings of ANNIMAB-1, the first international conference on artificial neural networks in medicine and biology. Comprising a selection of papers from leading researchers in the field, it summarises the sate-of-the-art, analyses the relationship between ANN techniques and other available methods and points to possible future biological and medical uses of ANNs. It focuses on three main areas:
-
The medical applications of artificial neural networks;
-
The uses of artificial neural networks outside clinical medicine, for example for data analysis in molecular biology and in simulations of biological systems; and
-
The theoretical aspects of artificial neural networks.
Essential reading for all neural network theorists, this book will be of particular interest in the fields of bioinformatics, statistics and biochemistry, for research and development personnel in biotechnology and medical technology companies, and to students of statistics, neuroscience, clinical medicine and medical biology.
Journal articles:
On Rotational Invariance in Adaptive Spatial Filtering of fMRI Data
NeuroImage, In press, 2005.
Abstract
Canonical correlation analysis (CCA) has previously been shown to
work well for detecting neural activity in fMRI data. The reason is
that CCA enables simultaneous temporal modelling and adaptive
spatial filtering of the data. This article introduces a novel
method for adaptive anisotropic filtering using the CCA framework
and compares it to a previously proposed method. Isotropic adaptive
filtering, which is only able to form isotropic filters of different
sizes, is also presented and evaluated. It is shown that a new
feature of the proposed method is invariance to the orientation of
activated regions, and that the detection performance is superior to
both that of the previous method and to isotropic filtering.
Detection and detrending in fMRI data analysis
NeuroImage, Vol. 22, No 2, June 2004.
Abstract
This article addresses the impact that colored noise, temporal
filtering, and temporal detrending have on the fMRI analysis
situation. Specifically, it is shown why the detection of
event-related designs benefit more from pre-whitening than blocked
designs in a colored noise structure. Both theoretical and empirical
results are provided. Furthermore, a novel exploratory method for
producing drift models that efficiently capture trends and drifts in
the fMRI data is introduced. A comparison to currently employed
detrending approaches is presented. It is shown that the novel
exploratory model is able to remove a major part of the slowly varying
drifts that are abundant in fMRI data. The value of such a model lies
in its ability to remove drift components that otherwise would have
contributed to a colored noise structure in the voxel time series. D
2004 Elsevier Inc. All rights reserved.
Canonical Correlation Analysis of Risk
Factors and Clinical Outcomes in Cardiac Surgery
L Ridderstolpe, H. Gill,
M Borga,
H. Rutberg, H. Åhlfeldt
Journal of Medical Systems,Vol 29, Pages 357-377, 2005.
Abstract
Background: Assessment of the association between risk factors and
outcomes in cardiac surgery is a complex problem. The aim of this
study was to explore the relationship between possible risk factors
and several clinical outcomes in cardiac surgery by using canonical
correlation analysis (CCA).
Methods: This retrospective study of 2605
consecutive adult patients who underwent cardiac surgery, evaluated 74
potential risk factors and up to 12 outcomes by canonical correlation
analysis. For three serious outcomes, sternal wound
complications/mediastinitis, cerebral complications and perioperative
myocardial infarctions, CCA was preceded by univariate analyses and
backward stepwise multivariate logistic regression analyses.
Results:
The CCA suggests that the major risk factors for complications in
these models are intraoperative and postoperative risk factors.
Conclusion: The power of risk prediction models developed with
multivariate regression analysis can be enhanced by application of
canonical correlation analysis, thereby offering new ways of analyzing
and interpreting sets of potential risk factors in relation to sets of
clinical outcomes.
Fibre optic array for curvature assessment - Application in otitis
diagnosis
M Sundberg,
M Borga,
H Knutsson,
A Johansson, T Strömberg, P Å Öberg
Medical and Biological Engineering and Computing, Vol 42, 2004.
Abstract
A contact free sensor consisting of two parallel optical fibre arrays
was designed to assess surface shapes of diffusely scattering media.
By sequentially illuminating objects using one fibre array and
detecting the diffusely back-scattered photons by the other, a
source-detector intensity matrix was formed, where the matrix element
(i,j) was the intensity at detector j when light-source i was excited.
Experimental data from convex and concave polyacetal plastic surfaces
were recorded. A mathematical model was used for simulating
source-detector intensity matrices for the surfaces analysed in the
experiments. Experimental results from the system were compared to the
theoretically expected results provided by the mathematical model. The
shape and relative amplitude showed similar behaviour in the
experiments and simulations. A convex/concave discriminator index, D,
representing the detected intensity difference between two
sourcedetector separations was defined. The relative dynamic range of
D, defined as the difference of the max and the min divided by the
mean of the index, was 1.37 for convex surfaces and 0.68 for concave
surfaces, at a measuring distance of 4.5 mm,. The index D was positive
for convex surfaces and negative for concave surfaces, which shows
that the system can distinguish between convex and concave surfaces,
an important result for the diagnosis of otitis media.
Adaptive Analysis of fMRI Data
NeuroImage, Volume 19, Number 3, 2003
Abstract
This article introduces novel and fundamental improvements of fMRI
data analysis . Central is a technique termed constrained Canonical
Correlation Analysis, which can be viewed as a natural extension and
generalization of the popular General Linear M odel method. The
concept of spatial basis filters is presented and shown to be a very
success ful way of adaptively filtering the fMRI data. A general
method for designing suitable hemodynamic res ponse models is also
proposed and incorporated into the constrained canonical correlation
approach. Results that demonstrate how each of these parts
significantly improves the dete ction of brain activity, with a
computation time well within limits for practical use, are prov ided.
Exploratory fMRI Analysis by
Autocorrelation Maximization
NeuroImage, Vol. 16, Number 2, June, 2002.
Abstract
A novel and computationally efficient method for exploratory analysis
of functional MRI data is
presented. The basic idea is to reveal underlying
components in the fMRI data that have maximum autocorrelation. The
tool for accomplishing this task is Canonical Correlation Analysis.
The relation to
Principal Component Analysis and Independent Component Analysis
is discussed and the performance of the methods is compared
using both simulated and real data.
Detecting Neural Activity in fMRI using Maximum Correlation Modelling
NeuroImage, Vol. 15, Number 2, February 1, 2002
Abstract
A technique for detecting neural activity in functional MRI data is
introduced. It is based on a novel framework termed Maximum
Correlation Modelling. The method employs a spatial filtering
approach that adapts to the local activity patterns, which results in
an improved detection sensitivity combined with good specificit y. A
spatially varying hemodynamic response is simultaneously modelled by a
sum of two Gamma functions. Comparisons to traditional analysis
methods are made using both synthetic and real data. The results
indicate that the Maximum Correlation Modelling approach is a strong
alternative for analyzing fMRI data.
Detection of Neural Activity in Functional MRI
using Canonical Correlation Analysis
Magnetic Resonance in Medicine, Volume 45, Issue 2, 2001.
Abstract
A novel method for
detecting neural activity in functional magnetic resonance imaging
(fMRI) data is introduced. It is based on Canonical Correlation
Analysis (CCA), which is a multivariate extension of the univariate
correlation analysis widely used in fMRI. To detect homogeneous
regions of activity, the method combines a subspace modeling of the
hemodynamic response and the use of spatial relationships. The spatial
correlation that undoubtedly exists in fMR images is completely
ignored when univariate methods as t-tests, F-tests, ordinary
correlation analysis etc. are used. Such methods are for this reason
very sensitive to noise, leading to difficulties in detecting
activation and significant contributions of false activations. In
addition, the proposed CCA-method also makes it possible to detect
activated brain regions based not only on thresholding a correlation
coefficient, but also on physiological parameters such as temporal
shape and delay of the hemodynamic response. Excellent performance on
real fMRI data is demonstrated.
Conference papers:
Correlation Controlled Bilateral Filtering of fMRI Data
ISMRM05, Miami beach, USA,
May 2005
Correlation Controlled Adaptive Filtering for fMRI Data Analysis
NBC05, Umeå, Sweden,
June 2005
Abstract
In analysis of fMRI data, it is common to average neighboring voxels
in order to obtain robust estimates of the correlations between voxel
timeseries and the model of the signal expected to be present in
activated regions. This paper presents a novel method for analysis of
fMRI data, which extends this approach by averaging only neighboring
voxels whose time-series have similar correlation coefficients. A
comparison between the new method and two other filtering strategies
is also presented, and the novel method is shown to have superior
ability to discriminate between active and inactive voxels.
Generation of Patient Specific Bone Models From Volume Data Using Morphons
NBC05, Umeå, Sweden,
June 2005
fMRI Data Analysis Using Correlation Controlled Adaptive Filtering
SSBA05, Malmö, Sweden,
March 2005
Adaptive bilateral filters for denoising of digital mammography images
SSBA05, Malmö, Sweden,
March 2005
Some Issues on the Segmentation of the Femur in CT Data
SSBA05, Malmö, Sweden,
March 2005
Motion Artifact Reduction in MRI
through Generalized DFT
ISBI04, Arlington, USA,
April 2004
Abstract
This paper presents a method that dramatically reduces artifacts
caused by respiratory (and similar types of) patient motion in
magnetic resonance imaging (MRI). The basis for the method is the
observation that affine deformations of an object will correspond to a
different but unique affine coordinate transform (plus phase shift) of
the Fourier representation of the object. The resulting sample points
will be irregularly distributed prohibiting the use of standard IFFT
to reconstruct the object. The object can however be reconstructed
through the use of a weighted regularized pseudo inverse. A standard
pseudo inverse is, however, not possible due to excessive
computational demands. For this reason a a novel fast sequential
pseudo inverse algorithm is also presented. Significantly improved
results are obtained on both synthetic and clinical data.
Dimensionality and degrees of freedom in fMRI data analysis
- A comparative study
ISBI04, Arlington, USA,
April 2004
Abstract
Two- and three-dimensional isotropic and anisotropic spatial filters
for adaptive fMRI data analysis are compared in terms of activation
detection sensitivity and specificity. Evaluations using both real
and artificial data are presented. It is shown that
three-dimensional anisotropic filters provide superior activation
detection performance.
A Bayesian Approach to Image Restoration
ISBI04, Arlington, USA,
April 2004
Abstract
A method for reducing additive noise in images by explicit analysis of
local image statistics is introduced and compared to other noise
reduction methods. The proposed method, which makes use of an a priori
noise model, has been evaluated on artificial and real (MRI)
image data.
Some Issues on the Segmentation of the Femur in CT Data
SSBA 2004, Uppsala, Sweden
March 2004
Abstract
This paper presents a recently started project which goal is to
automatically generate patient specific models for visual and haptic
simulation of hip fracture surgery. It includes a preliminary study
of a computed tomography (CT) dataset of the pelvic region. The
paper emphasizes some issues encountered when segmenting bones in
this region, especially in the area around the proximal femur.
Click here to get full paper in pdf
Fully Automatic Segmentation of the Hippocampus in MR Images
G Starck,
M Borga,
M Friberg, E Olsson, S Ribbelin
H Knutsson,
S Ekholm, H Malmgren
ESMRMB 2002, Cannes, France
August 2002
Click here to get full paper in pdf (580 k)
Recognizing Emphysema - A Neural Network Approach
ICPR 2002, Québec City, Canada
August 2002
Abstract
An accurate and fully automatic method for detecting and quantifying emphysema in CT-images is presented. The method is based on an image preprocessing step followed by a neural network classifier trained to separate true emphysema from artifacts. The proposed approach is shown to be superior to an established method when applied on real patient data.
A Canonical Correlation Approach to Exploratory Data Analysis in fMRI
ISMRM 2002, Honolulu, Hawaii, USA
May 2002
Abstract
A computationally efficient data-driven method for exploratory analysis
of functional MRI data is presented. The basic idea is to reveal
underlying components in the fMRI data that have maximum
autocorrelation. The tool for accomplishing this task is Canonical
Correlation Analysis. The proposed method is more robust and much more
computationally efficient than independent component analysis, which
previously has been applied in fMRI.
Click here to get full paper in pdf (50 k)
Hierarchical Temporal Blind Source Separation of fMRI Data
ISMRM 2002, Honolulu, Hawaii, USA
May 2002
Abstract
Blind Source Separation (BSS) of fMRI data can be done both temporally
and spati ally. Temporal BSS of fMRI data has one fundamental problem
not encountered in the spatial BSS approach. There are thousands of
observed timecourses in an f MRI data set while the number of samples
of each timecourse typically is less than two hu ndred. This relation
makes the problem of recovering the underlying temporal sources
ill-posed. This contributi on eliminates this problem by introducing a
hierarchical approach for performing temporal BSS of fMRI data.
Click here to get full paper in pdf (24 k)
Non-linear Gaussian Filtering for Image Resampling
SSAB 2002, Lund, Sweden
March 2002
Abstract
This paper presents an alternative to standard interpolation and
resampling of images. The method is based on a non-linear low-pass
filter with an intensity dependent filter kernel. The main advantage
is that interesting image features such as edges and lines are
better preserved than in traditional resampling.
Blind Source Separation of Functional MRI Data
SSAB 2002, Lund, Sweden
March 2002
Abstract
A computationally efficient method for exploratory analysis
of functional MRI data is presented. The basic idea is to reveal
underlying components in functional Magnetic Resonance Imaging
data that have maximum autocorrelation. The tool for accomplishing this task is
Canonical
Correlation Analysis. The proposed approach is more robust and much more
computationally efficient than Independent Component Analysis, which
is an established method for the type of problem considered.
Emphysema Detection in CT Images
SSAB 2002, Lund, Sweden
March 2002
Abstract
This paper describes a fully automatic approach for detecting
emphysema in CT im ages of the lungs. The method combines an image
processing step, where potential emphysematous area s are extracted,
and a neural network step trained to recognize true emphysema. Results
demonstra ting the viability of the approach are shown.
Respiratory Artifact Reduction in MRI using Dynamic Deformation Modeling
SSAB 2002, Lund, Sweden
March 2002
Abstract
This paper presents a novel magnetic resonance imaging (MRI)
reconstruction method that will reconstruct an object correctly
despite the presence of respiratory-type motions. The basis for the
method is the observation that a ne deformations of an object will
correspond to a different but unique a ne coordinate transform of the
Fourier representation (k-space) of the object. The resulting sample
points will be irregularly distributed prohibiting the use of standard
IFFT to reconstruct the object. The object can however be
reconstructed through the use of a weighted regularized pseudo
inverse. Short computing times are obtained using a novel fast
sequential pseudo inverse algorithm.
Generation of Representations for Supervised Learning
- A Velocity Estimation Example
SCIA 2001, Bergen, Norway
June 2001
Abstract
A two-step learning method for velocity estimation is presented.
First, an efficient representation of velocity is found using a
learning technique based on canonical correlation analysis. This
results in a spherical representation. Then, given this new
representation, the mapping from input data to velocity estimates are
learned by minimizing the mean square error between the output and the
desired output on a training set. The non-linear mapping on the
'velocity-sphere' representation improves the performance of the
linear method in the supervised learning step.
Click here to get full paper in pdf (500 k)
A correlation framework for functional MRI data analysis
SCIA 2001, Bergen, Norway
June 2001
Abstract
A correlation framework for detecting brain activity in functional MRI
data is p resented. In this framework, a novel method based on
canonical correlation analysis follow s as a natural extension of
established analysis methods. The new method shows very good detect
ion performance. This is demonstrated by localizing brain areas which
control finger movements an d areas which are involved in numerical
mental calculation.
Click here to get full paper in pdf (1.5 M)
Increased detection sensitivity in fMRI by
adaptive filtering
Int. Soc. Mag. Reson. Med (ISMRM)
9, Glasgow, Scotland
April 2001
Abstract
Previously, a new method for detecting neural activity in fMRI
was introduced. In this contribution, the new method which is based on
canonical correlation analysis is shown to be a natural extension of
established detection methods when the latter are seen from a
correlation perspective. In addition, the increased sensitivity of the
new method is demonstrated.
Click here to get full paper in pdf (31 k)
or here to get the poster in pdf (48 k)
Canonical correlation analysis in early vision processing
ESANN 2001, Brugges, Belgium
April 2001
Abstract
This paper illustrates how canonical correlation analysis can be
used for designing efficient visual operators by learning. The
approach is highly task oriented and what constitutes the relevant
information is defined by a set of examples. The examples are pairs
of images displaying a strong dependence in the chosen feature but
are otherwise independent. Experimental results are presented
illustrating the learning of local shift invariant orientation
operators, representation of velocity, and image content invariant
disparity operators.
Click here to get full paper in pdf (300 k)
Canonical Correlation as a Tool in Functional MRI analysis
SSAB 2001, Norrköping, Sweden
April 2001
Abstract
In this contribution an application of canonical correlation analysis
(CCA) in image processing is presented. Functional MRI is a technique
used for exploring the functionality of the brain. From an image
analysis viewpoint, the problem is to segment or detect active brain areas
in MR images. In contrast to traditional segmentation problems
where the spatial shape of objects are of vital interest, the main
feature in the present problem is the temporal behavior of the pixels.
The use of CCA enables us to make a combined spatio-temporal
analysis of the data with an improved detection sensitivity as result.
Click here to get full paper in pdf (1.5 M)
Learning Corner Orientation using Canonical Correlation
SSAB 2001, Norrköping, Sweden
April 2001
Abstract
This paper shows how canonical correlation can be used to learn a
detector for corner orientation invariant to corner an-gle and
intensity. Pairs of images with the same corner ori-entation but
different angle and intensity are used as training samples. Three
different image representations; intensity values, products between
intensity values, and local orien-tation are examined. The last
representation gives a well behaved result that is easy to decode into
the corner orien-tation. To reduce dimensionality, parameters from a
poly-nomial model fitted on the different representations is also
considered. This reduction did not affect the performance of the
system.
Click here to get full paper in pdf (266 k)
An improved algorithm for computerized detection and quantification of pulmonary emphysema at high resolution computed tomography (HRCT)
Medical Imaging (SPIE),
San Diego, California USA,
February 2001
Abstract
Emphysema is characterized by destruction of lung tissue with
development of small or large holes within the lung. These areas will
have Hounsfield values (HU) approaching -1000. It is possible to
detect and quantificate such areas using simple density mask
technique. The edge enhancement reconstruction algorithm, gravity and
motion of the heart and vessels during scanning causes artefacts,
however. The purpose of our work was to construct an algorithm that
detects such image artefacts and corrects them. The first step is to
apply inverse filtering to the image removing much of the effect of
the edge enhancement reconstruction algorithm. The next step implies
computation of the antero-posterior density gradient caused by gravity
and correction for that. Motion artefacts are in a third step
corrected for by use of normalized averaging, thresholding and region
growing. Twenty healthy volunteers were investigated, 10 with slight
emphysema and 10 without. Using simple density mask technique it was
not possible to separate persons with disease from those without. Our
algorithm improved separation of the two groups considerably. Our
algorithm needs further refinement, but may form a basis for further
development of methods for computerized diagnosis and quantification
of emphysema by HRCT.
Click here to get full paper in pdf (674 k)
Novel Canonical Correlation Method for Processing fMRI data
Annual meeting of the Swedish Biophysical Society, Linköping, Sweden
September 2000
Novel Applications of Canonical Correlation Analysis in True Spatio-Temporal fMRI Analysis
ESMRMB 2000, Paris, France
September 2000
Automated Generation of Representations in Vision
ICPR 2000, Barcelona, Spain
September, 2000
Invited talk
Abstract
This paper presents a general strategy for automated generation
of efficient representations in vision. The approach is highly task
oriented and what constitutes the relevant information is defined by
a set of examples. The examples are pairs of situations that are
dependent through the chosen feature but are otherwise independent.
Particularly important concepts in the work are mutual information
and canonical correlation. How visual operators and representations
can be generated from examples are presented for a number of
features, e.g. local orientation, disparity and motion. Interesting
similarities to biological vision functions are observed. The
results clearly demonstrates the potential of combining advanced
filtering techniques and learning strategies based on canonical
correlation analysis (CCA).
Click here to get the full paper in portable data format (pdf) (664 k)
FSED - Feature Selective Edge Detection
ICPR 2000, Barcelona, Spain
September, 2000
Abstract
We present a novel method that finds edges between certain image
features, e.g. gray-levels, and disregards edges between other
features. The method uses a channel representation of the features
and performs normalized convolution using the channel values as
certainties. This means that areas with certain features can be
disregarded by the edge filter.
The method provides an important new tool for finding tissue
specific edges in medical images, as demonstrated by an MR-image example.
Click here to get the full paper in portable data format (pdf) (144 k)
Finding Efficient Nonlinear Visual Operators using Canonical
Correlation Analysis
SSAB 2000, Halmstad, Sweden
March, 2000
Abstract
This paper presents a general strategy for designing efficient
visual operators. The approach is highly task oriented and what
constitutes the relevant information is defined by a set of
examples. The examples are pairs of images displaying a strong
dependence in the chosen feature but are otherwise independent.
Particularly important concepts in the work are mutual information
and canonical correlation. Visual operators learned from examples are
presented, e.g. local shift invariant orientation operators and
image content invariant disparity operators. Interesting
similarities to biological vision functions are observed.
Click here to get the full paper in portable data format (pdf) (432 k)
Learning Visual Operators from Examples: A New Paradigm in Image Processing
ICIAP '99, Venice, Italy
September, 1999
Invited talk
Abstract
This paper presents a general strategy for designing efficient
visual operators. The approach is highly task oriented and what
constitutes the relevant information is defined by a set of
examples. The examples are pairs of images displaying a strong
dependence in the chosen feature but are otherwise independent.
Particularly important concepts in the work are mutual information
and canonical correlation. Visual operators learned from examples are
presented, e.g. local shift invariant orientation operators and
image content invariant disparity operators. Interesting
similarities to biological vision functions are observed.
Click here to get the full paper in portable data format (pdf) (316 k)
Estimating Multiple Depths in Semi-transparent Stereo Images
SCIA '99, Kangerlussuaq, Greenland
June 1999
Abstract
A stereo algorithm that can estimate multiple depths in
semi-transparent images is presented. The algorithm is based on a
combination of phase analysis and canonical correlation analysis.
The algorithm adapts filters in each local neighbourhood of the
image in a way which maximizes the correlation between the filtered
images. The adapted filters are then analysed to find the disparity.
This is done by a simple phase analysis of the scalar product of the
filters. For images with different but constant depths, a simple
reconstruction procedure is suggested.
Click here to get the full paper in portable data format (pdf) (486 k)
Learning Multidimensional Signal Processing
ICPR '98, Brisbane, Australia
August, 1998
Invited talk
Abstract
This paper presents our general strategy for designing learning
machines as well as a number of particular designs. The search for
methods allowing a sufficient level of adaptivity are based on two
main principles: 1. Simple adaptive local models and 2. Adaptive model distribution.
Particularly important concepts
in our work is mutual information and canonical correlation. Examples
are given on learning feature descriptors, modelling disparity,
synthesis of a global 3-mode model and a setup for reinforcement
learning of online video coder parameter control.
An Adaptive Stereo Algorithm Based on Canonical Correlation Analysis
ICIPS '98, Gold Coast, Australia
August, 1998
Abstract
This paper presents a novel algorithm that uses CCA and phase
analysis to detect the disparity in stereo images. The algorithm
adapts filters in each local neighbourhood of the image in a way
which maximizes the correlation between the filtered images. The
adapted filters are then analysed to find the disparity. This is
done by a simple phase analysis of the scalar product of the
filters. The algorithm can even handle cases where the images have
different scales. The algorithm can also handle depth
discontinuities and give multiple depth estimates for
semi-transparent images.
Click here to get the full paper in portable data format (pdf) (394 k)
Learning Canonical Correlations
In
Proceedings of the Scandinavian Conference on Image Analysis, Lappeenranta,
Finland, 9-11 June 1997
Abstract
This paper presents a novel learning algorithm that finds the
linear combination of one set of multi-dimensional variates that is the
best predictor, and at the same time finds the linear combination of
another set which is the most predictable. This relation is known as
the canonical correlation and has the property of being invariant with
respect to affine transformations of the two sets of variates. The
algorithm successively finds all the canonical correlations beginning
with the largest one. It is shown that canonical correlations can be
used in computer vision to find feature detectors by giving examples of
the desired features. When used on the pixel level, the method finds
quadrature filters and when used on a higher level, the method finds
combinations of filter output that are less sensitive to noise compared
to vector averaging.
Click here to get the full paper in portable data format (pdf) (480 k)
Hierarchical Reinforcement Learning
In
Proceedings of the International Conference on Neural Networks, Amsterdam,
The Netherlands, 13-16 September 1993
S. Gielen and B. Kappen (editors)
Springer-Verlag
London 1993
Abstract
A hierarchical representation of the input-output transition function
in a learning system is suggested. The choice of either representing
the knowledge in a learning system as a discrete set of input-output pairs
or as a continuous input-output transition function is discussed. The
conclusion that both representations could be efficient, but at
different levels of abstraction is made. The difference between
strategies and actions is defined. An algorithm for using
adaptive critic methods in a two-level reinforcement
learning system is presented. Simulations of a one dimensional hierarchical
reinforcement learning system is presented.
See also the
report
with the same name.
Reports:
A Unified Approach to PCA, PLS, MLR and CCA
November 25, 1997
LiTH-ISY-R-1992
Abstract
This paper presents a novel algorithm for analysis of stochastic
processes. The algorithm can be used to find the required solutions
in the cases of principal component analysis (PCA), partial least
squares (PLS), canonical correlation analysis (CCA) or multiple
linear regression (MLR). The algorithm is iterative and sequential
in its structure and uses on-line stochastic approximation to reach
an equilibrium point. A quotient between two quadratic forms is used
as an energy function and it is shown that the equilibrium points
constitute solutions to the generalized eigenproblem.
Click here to get the full paper in portable data format (pdf) (1540 k)
Generalized Eigenproblem for Stochastic Process Covariances
December 3, 1996
LiTH-ISY-R-1916
Abstract
This paper presents a novel algorithm for finding the solution of the
generalized eigenproblem where the matrices involved contain
expectation values from stochastic processes. The algorithm is
iterative and sequential to its structure and uses on-line stochastic
approximation to reach an equilibrium point. A quotient between two
quadratic forms is suggested as an energy function for this problem
and is shown to have zero gradient only at the points solving the
eigenproblem. Furthermore it is shown that the algorithm for the
generalized eigenproblem can be used to solve three important problems
as special cases. For a stochastic process the algorithm can be used
to find the directions for maximal variance, covariance, and canonical
correlation as well as their magnitudes.
Click here to get the full paper in portable data format (pdf) (492 k)
Learning Canonical Correlations
June 5, 1995
LiTH-ISY-R-1761
Abstract
This paper presents a novel learning algorithm that finds the linear
combination of one set of multi-dimensional variates that is the best
predictor, and at the same time finds the linear combination of
another set which is the most predictable. This relation is known as
the canonical correlation and has the property of being
invariant with respect to affine transformations of the two sets of
variates. The algorithm successively finds all the canonical
correlations beginning with the largest one.
Click here to get the full paper in portable data format (pdf) (222 k)
Reinforcement Learning Trees
1996
LiTH-ISY-R-1828
Abstract
Two new reinforcement learning algorithms are presented. Both use a
binary tree to store simple local models in the leaf nodes and coarser
global models towards the root. It is demonstrated that a meaningful
partitioning into local models can only be accomplished in a fused
space consisting of both input and output. The first algorithm uses a
batch like statistic procedure to estimate the reward functions in the
fused space. The second one uses channel coding to represent the
output- and input vectors allowing a simple iterative algorithm based
on competing subsystems. The behaviors of both algorithms are
illustrated in a preliminary experiment.
Click here to get the full paper in portable data format (pdf) (142 k)
On-Line Singular Value Decomposition of Stochastic Process Covariances
June 5, 1995
LiTH-ISY-R-1762
Abstract
This paper presents novel algorithms for finding the singular value
decomposition (SVD) of a general covariance matrix by stochastic
approximation. General in the sense that also non-square, between
sets, covariance matrices are dealt with. For one of the algorithms,
convergence is shown using results from stochastic approximation
theory. Proofs of this sort, establishing both the point of
equilibrium and its domain of attraction, have been reported very
rarely for stochastic, iterative feature extraction algorithms.
Click here to get the full paper in portable data format (pdf) (306 k)
A binary Competition Tree for Reinforcement Learning
August 26, 1994
LiTH-ISY-R-1623
Abstract
A robust, general and computationally simple reinforcement learning
system is presented. It uses a channel representation which is robust
and continuous. The accumulated knowledge is represented as a reward
prediction function in the outer product space
of the input- and output channel vectors. Each computational
unit generates an output simply by a vector-matrix multiplication and
the response can therefore be calculated fast. The
response and a prediction of the reward are calculated simultaneously
by the same system, which makes TD-methods easy to implement if
needed. Several units can cooperate to solve more complicated
problems. A dynamic tree structure of linear
units is grown in order to divide the knowledge space into a
sufficiently number of regions in which the reward function can be
properly described. The tree continuously tests split- and prune
criteria in order to adapt its size to the complexity of the problem.
Click here to get the full paper in portable data format (pdf) (4246 k)
Hierarchical Reinforcement Learning
January 27, 1993
LiTH-ISY-R-1766
A hierarchical representation of the input-output transition function
in a learning system is suggested. The choice of either representing
the knowledge in a learning system as a discrete set of input-output
pairs or as a continuous input-output transition function is
discussed. The conclusion that both representations could be
efficient, but at different levels is made. The difference between
strategies and actions is defined. An algorithm for using adaptive
critic methods in a two-level reinforcement learning system is
presented. Two problems that are faced, the hierarchical credit
assignment problem and the equalized state problem are
described. Simulations of a one dimensional hierarchical reinforcement
learning system is presented.
Click here to get the full paper in portable data format (pdf) (610 k)
A Survey of Current Techniques for Reinforcement Learning
June 4, 1992
LiTH-ISY-I-1391
Abstract
This survey considers response generating systems that improve their
behaviour using reinforcement learning. The difference between
unsupervised learning, supervised learning, and reinforcement learning is described. Two general problems concerning
learning systems are presented; the credit assignment problem
and the problem of perceptual aliasing. Notations and
some general issues concerning reinforcement learning systems are
presented. Reinforcement learning systems are further divided
into two main classes; memory mapping and projective
mapping systems. Each of these classes is described and some examples
are presented. Some other approaches are mentioned that do not fit
into the two main classes. Finally some issues not covered by the
surveyed articles are discussed, and some comments on the subject are
made.
Click here to get the full paper in portable data format (pdf) (239 k)