Thomas Fabricius
 |
Master of Science, Ph.D. student. Tel: (+45) 45 25 38 94. E-mail:
tf@eivind.imm.dtu.dk
I am currently working on my Ph.D.
Bayesian Signal Detection and Processing with Application in Wireless
Communication which is a continuation of my Master Thesis
Bayesian Model Selection by Markov Chain Monte Carlo Methods. Model selection impl
ies modelling, for that reason am I very interested in Machine Learning
which deal with the problem of infering parameters. I am at most devoted to the Bayesian learning paradigm of probabilistic learning. Solving Bayesian model problems quickly gets out o
f the closed form solution scope, for that reason is nummerical methods and approximations the only solutions. The
sampling approach is the one which has taken most of my time, but recently the variational approximations has taken
some of my time and I have
proposed, in my master thesis, a variational approximation to the Artificial Neural Network (ANN).
The model selection problem have I tried to solve by "path-sampling", unfortunately models can often implement different intepretations of data. As an example, which I have examined in my master thesis, regression outputs can be implemeted as func
tions over inputs with additive Gaussian noise. One intpretation could then be to implement all the signal as noise, which would be the most plausible explanation when the effective number of data examples are less than the number of parameters in the mod
el. Another intepretation of data is, the one we expect exists, to implement data as the functions over the inputs added by noise with some level. These different ways of intepreting data have a large tendency to bias the Model Selection towards the start
ing model i.e. the model intepretaion that is valid in the initial
samples.
All the theoretical aspect is very well applied to communication theory. In communication theory we often make an
entropy decoding with a binary alphabet as output, eventually we also
perform some redundancy encoding. The problem is
then to infer the send symbols after transmission
over a noisy channel. This implies solving Bayesian integrals (or rather sums over the possible states). This is in
genral a complex task. For that reason is it necessary to use
approximations to get it run realtime, we call the
sollutions for nonoptimal decoders. The
approximations are borrowed from statistical physics of disordered
spin-systems, where the naive mean-field and
Thouhless-Anderson-Palmer (TAP) approximations are very well applyable. This field also gives some nice ways to
analyse the nonoptimal decoders, which in general are good at low signal
to noise ratios and at weak dependencies
between the decoded bits. Often the system of bits behave differently in different regimes of the estimated or
postulated signal to noise ratio or what would be identical in physics,
the temperature. Physicists talk about the
magnetic or ferromagnetic or even spin-glass phase of the system, where we
go from one global sollution to many equal probable solutions.
The work on nonoptimal decoders is conducted together with Nokia Digital Signal Processing Group, where we look at
more specific detector systems.
Curriculum Vitae
links
ICASSP2001 accepted :
Dynamic Components of Linear Stable Mixtures From Fractional Low Order Moments
Not submittet :
Phantom Belief Propagation for Fast Approximative Inference
Globecom2002 submission :
Approximations to Joint-ML and ML
Symbol-Channel Estimators in MUD CDMA
ISPACS2002 submission:
Improved Multistage Detector by Mean-Field Annealing
in Multi-User CDMA
Last modified: Wed June 5 10:50:33 CEST 2002