Research jobs in neural nets and pattern recognition

[w.penny@ic.ac.uk]

THREE POST-DOCTORAL RESEARCH POSITIONS
IN PATTERN RECOGNITION  / NEURAL NETWORKS RESEARCH

Three post-doctoral research positions are available within the Neural
Systems Section of the Department of Electrical & Electronic Engineering to
work on the theory and application of advanced pattern recognition
techniques, in particular the use of Bayesian methods and neural networks.
Two positions are funded for two years and the third nominally for three
years with a yearly evaluation.  All projects involve research in
statistical pattern recognition with applications in the biomedical field.
Experience in pattern recognition and Bayesian statistics would be an
advantage.  A good understanding of data processing (especially signal
processing) techniques is desired as is experience of UNIX, C and Matlab.
The positions are funded by the Jefferiss Research Trust, the European
Commission and British Aerospace plc respectively. The salary scale will be
RA1A , GBP 14,732 - 22,143 per annum (exclusive of London Allowance of GBP
2,134) depending on age and experience.  Further information may be
obtained from http://www.ee.ic.ac.uk/research/neural/positions.html or via
e-mail to Dr Stephen Roberts (s.j.roberts at ic.ac.uk).  The closing date for
applications is April 11th 1997.

In recent years, great interest has developed in the use of non-classical
methods for statistical analysis of data as part of a general increase
towards the use of artificial intelligence methods.  One genre which has
shown itself to be particularly suitable is that of connectionist models, a
subset of which are referred to as a artificial neural networks (ANNs).
Classical statistical methods rely upon the use of simple models, such as
linear or logistic regression, in order to 'learn' relationships between
variables and outcomes. ANNs offer a far more flexible model set, indeed it
has been shown that they have the property of universal approximation so
they are able, in principle, to estimate any set of arbitrary relationships
between variables.  Furthermore, they may model non-linear coupling between
sets of variables.  Part of the momentum of the recent development of ANNs
for pattern recognition, regression and estimation problems must be
attributed to the manner in which ANNs conform to many of the traditional
statistical approaches, i.e. they may estimate Bayesian probablilities in
the case of classification and conditional averages in the case of
regression.

1) The use of Neural Networks to Predict the Development and Progression of
Kaposi's Sarcoma (KS).

This is a joint project funded by the Jefferiss Research Trust between the
Department of Electrical and Electronic Engineering, Imperial College of
Science, Technology & Medicine and the Department of Genito-urinary
Medicine, St. Mary's Hospital.

Kaposi's sarcoma (KS) is a vascular tumour, which is more common and often
aggressive in patients with underlying immunosuppression (post-transplant
KS and AIDS-associated KS).  KS was first described by the Hungarian
pathologist Moritz Kaposi in 1872, yet still remains something of a
clinical enigma, being an unusual tumour of unknown origin.

The aim of this research is to determine factors that influence the
variable progression rate of KS in HIV infected individuals. There is
currently no means of predicting which patients will develop KS and no
understanding of the relationship between the forms of the disease. The aim
of the project is to carry out multi-variable analyses in order to define
clinical end-points and provide guidelines for better patient management.

A number of variables will be available to the system.  The reliability and
utility of each with regard to the prediction of patient outcome, however,
is generally unknown.  Classical regression analysis offers some powerful
methods of selection and ranking within a subset of features or variables.
Whilst such methods should be used for completeness and comparison, it is
noted that recent developments in Bayesian learning theory have offered the
possibility to assess the utility of variables from within the ANN
structure.  Each input variable has a separate weighting factor, or in
Bayesian terminology, a hyper-prior, associated with it.  This technique
has become known as automatic relevance determination or ARD.  Such an
assessment is devoid of the strong assumptions of independence and
linearity of most of the classical regression methods.

It is feasible for an ANN to produce not only a set of output variables
(predictions or classifications, for example) but also an associated set of
confidence or validation measures (describing the probable error on each
output).  This enables the tracking of predictions of future events in a
more robust framework and furthermore allows for the accurate fusion of
information from more than one source and the incorporation of temporal
information, i.e. poor quality information from the present time may be
suppressed in favour of more reliable information from past or future as it
becomes available.

If we may regard the system as aiming to produce a probability distribution
in some 'outcome space', then several possible approaches to analysis are
made available.  As temporal information is retained (i.e. outcomes are
based upon the entire course of the patient's history, not just present
information) we may seek information regarding the effect of each piece of
information (test result or partial diagnosis) on the probability
distribution in the 'outcome space'.  Two pieces of information may be
obtained from this approach. How important a partial decision or test
result is to the probability of certain outcomes and how important it is to
changing the uncertainty we have in the outcome results.  Clearly, the goal
will be to indicate tests and/or procedures which not only increase the
survivability probabilities but also make the estimated outcomes less
variant, so we have more confidence in the predictions (this means not only
increasing the height of a favourable node in the posterior probability
space, but also attempting to reduce the variance of the distribution).  In
order to accommodate for multiple output hypotheses we propose to utilise a
procedure similar to that detailed in (Bishop 1995) whereby the output
distribution is modelled multi- modally.  This has the added benefit that
individual modes (possible outcomes) may be tracked separately.  This
representation is also similar to that taken in a mixture of experts
approach.

REFERENCES 

1. Bishop CM. Neural Networks for Pattern Recognition. Oxford University
Press, Oxford, 1995.

2. Ripley BD. Pattern Recognition and Neural Networks. Cambridge University
Press, Cambridge, 1996.

3. Roberts SJ and Penny W. Novelty, Confidence and Errors in Connectionist
Systems.  Proceedings of IEE colloquium on fault detection and intelligent
sensors, IEE, September 1996.

4. Penny W and Roberts SJ. Neural Networks with Error Bars. Departmental
report, also submitted to IEEE transactions on neural networks, February
1997, available from http://www.ee.ic.ac.uk/staff/hp/sroberts.html


2)	SIESTA (EU funded project)

Siesta is a an EU funded project which involves Imperial College and 10
other European partners.  The aim of the project is to define and produce a
system which is capable of continuous evaluation of the state of the brain
during the sleep-wake cycle.  Such an automated system is of enormous value
in the clinical field and the research into multi-channel signal
processing, fusion and pattern recognition form a challenge to the most
modern techniques.  The state of the brain will, primarily, be monitored
via its electrical activity (the EEG).

One of the most well-known approaches from literature to achieve a
continuous description of EEG state is the system developed by Roberts &
Tarassenko (1992a, 1992b). This approach will be used as a general basis
for the research in SIESTA.  Roberts & Tarassenko (henceforth, R&T') used
a self-organizing feature map (SOM) to perform unsupervised topographic
mapping of feature vectors consisting of 10 coefficients of a Kalman filter
algorithm applied to the raw EEG. This self- organizing network discovered
eight distinct clusters in which the brain state remained preferentially.