- Chief Engineer, PhD
I completed my Masters in electrical engineering from the Department of Automation at the Technical University of Denmark (DTU) in 1984 with the thesis entitled "Adaptive Observere og Parametrisk Identifikation" (in Danish). Then I worked in a research and development group at Danfoss A/S (private company) until 1991, mostly with analysis of dynamic control systems, adaptive control theory and program development (analysis, design and implementation) for microprocessor systems and process control systems.
I was then fortunate to get a Ph.D. scholarship from the Danish Computational Neural Network Center, CONNECT, starting in September 1991, and was enrolled as a Ph.D student at the Electronics Institute, Technical University of Denmark (now Section for Digital Signal Processing Department of Informatics and Mathematical Modelling (IMM) at DTU). My major interests during the study was optimization of Neural Network Models for Time Series Analysis and Classification purposes. I finished my Ph.D study December 31, 1994. The title of the thesis is "Neural Networks for Signal Processing".
January 1994 I moved to the Neurobiology Research Unit at Rigshospitalet where I am working as an Associated Researcher with Signal Processing and Mathematical Modeling (especially, Artificial Neural Network techniques) within the field of brain modeling. Also, I work as system administrator for the computer laboratory in the Neurobiology Research Unit.
My major scientific interests are within signal processing, especially modeling of of brain data. More specifically I have focused on the following issues.
In this area I have mostly worked with estimation of rateconstants in kinetic models of the brain. Estimation of the rateconstants has been based on dynamic FDG PET data. Especially, I have worked with estimation of the rateconstants in Sokoloff's model at a pixel by pixel basis using an Artificial Neural Network technique for improving the speed.
PET and fMRI comparison
My contribution to this area has mostly been in analyzing and comparing some water PET and BOLD fMRI datasets. The analysis techniques that has been used is based on the SPM (Statistical Parametric Model/Generalized Linear Model) and SSM (Scaled Subprofile Model).
Within this area I have mainly worked with different methods for estimating the generalization ability for linear and especially non-linear mathematical models. (The generalization performance is here defined as the model performance at a dataset that hasn't been used for training model parameters.)
Optimization of Artificial Neural Network models
In the design and optimization process for artificial neural network models both the architecture (the number of connections between input, output and neurons) and the parameters (weights) for the connections. I have in this area worked with optimization schemes for especially, feed-forward neural networks, that optimizes the architecture (prunes connections from the structure) so the models are well fitted to a given problems, with respect to the generalization performance.
Optimization of training schemes for Artificial Neural Networks
In a optimization scheme for neural network parameters different ''hyperparameters'' has to be optimized for achieving good generalization performance. Different methods for tuning these parameters by dividing the dataset in training and validation datasets has been evaluated. The different optimization schemes for artificial neural network's can in general also be used for optimization of other types of linear and non-linear models. In the area of modeling medical datasets these are very important while almost always very limited datasets are present.
My full list of scientific publications can be accessed using the following PubMed search string: Svarer-C [Aurthor].
I live in Hørsholm a smaller city 25 km. north of Copenhagen. In the summertime I enjoy going for a sailing trip at Øresund. This is relaxation for body and soul.