28.6.2011

Ing. Martin Vejmelka, PhD., from the Institute of Computer Science, Academy of Sciences of the Czech Republic, will present his lecture on July 1, 2011, at the Seminar of the Department of Theoretical Methods, Institute of Measurement Science, Slovak Academy of Sciences.

Abstract

High-dimensional (HD) datasets are being acquired in many areas of science from meteorology through neuroscience to genetics. HD data is difficult to work with due to its volume, visualize do to its size and to comment on due to massive multiple testing difficulties. Also, a shift from "activity studies", where researchers concentrate on describing a single region, to "connectivity studies" where researchers focus on analysing interactions between different regions of a HD system, has occurred. Many techniques have been applied to reduce HD datasets: principal/independent component analysis, factor analysis, various clustering techniques to reduce data data into components. Graph theoretic measures quantifying various interconnection properties of weighted and unweighted graphs have been researched and applied to many social and natural datasets. Our group concentrates on connectivity estimation (causality / synchronization / correlation), data reduction techniques and graph-theoretic analysis. I will present some recent results of the group in this area - causality and synchronization analysis in bivariate data [1,2], average association clustering, detection of nonlinear connectivity in data, effect of nonlinearity on graph-theoretic properties in functional MRI data. Synchronization and causality analysis has been coupled with surrogate testing to determine if a detected coupling is statistically significant. Average association clustering [3] is a clustering algorithm from the spectral graph clustering family which has been developed to cluster fMRI data. Detection of nonlinear connectivity [4] quantifies how much information is being lost when estimating connectivity using linear measures in (possibly nonlinear) data. This technique was applied to atlas-reduced fMRI data to quantify differences in graph structure when the graph edge weights are estimated using correlation coefficient or mutual information from fMRI time series [5]. I will tie these methods together in a framework that we would like to use to systematically examine and analyse new data.