30.11.2017, 10:00 - 10:45 | HS2

Prof. Dr. Christian BÖHM (LMU München, Deutschland)

"Analysis Methods for Uncertain and Heterogeneous Data from Neuroscience Applications."

Währinger Straße 29, 1090 Wien, HS2

Abstract

Analysis Methods for Uncertain and Heterogeneous Data from Neuroscience Applications.

When considering data analysis methods for modern neurosciences like high-resolution imaging (e.g. fMRI, DTI), genome sequencing, questionnaires, observational studies, etc. the major problem is not only the sheer volume of the data (which may reach terabyte scale and beyond) and increasing speed of data generation but also its heterogeneity and varying reliability. Being commonly subsumed by the four "V" (Volume, Velocity, Variety, and Veracity) these issues of "Big Data" play an important role in different sciences and allover society. In the context of neuroscience, these are the major focuses of our research.

The Gauss-Tree is a technology for the efficient storage and analysis of large volumes of uncertain data. The uncertainty of data objects can be explicitly modeled in a probabilistic way as multivariate Gaussians and Gaussian Mixture Models. Specialized query types can be used as building blocks in high-level analysis methods like clustering or classification. Tree-based indexing facilitates a highly time-efficient execution of queries.

With analysis algorithms like INTEGRATE, INCONCO, and others, we have also proposed methods for the joint analysis of heterogeneous data from different modalities having varying scales. We have established information-theoretic principles which enable us to do simultaneous analyses on data of different levels of measurement, simultaneous analyses using different objective functions, and to directly compare these intermediate results and guide the further search. 

Finally, we have a strong background in high-performance machine learning and databases as we have developed various principles to deploy analysis methods on highly parallel and scalable hardware like multi-core vector processors and Graphic Processing Units which are both popular architectures for deep learning and provide orders of magnitude higher computing power compared to classical processors. 

In this talk I discuss the results from exemplary interdisciplinary projects with neuroscientists addressing the challenges of uncertainty, volume and variety: reliable grading of tumors from contrast enhanced MRI; robust identification of fiber bundles from diffusion tensor images; massively parallel exploration of SNP interaction patterns on graphics processing units from genome-wise association data. An outlook on my vision of integrative data analysis to support a systemic understanding of the brain will conclude the talk.

Bio

Christian Böhm is professor of informatics at Ludwig-Maximilians-Universität München since 2003. He is head of the research group for data mining in medicine (dmm-lmu.de). His research focuses on database-related data mining, in particular clustering  and integrative mining of heterogeneous data. His work has been published at SIGMOD, KDD, and ICDM, the top publication venues for data mining. Interdisciplinary projects with neuroscientists and other life sciences have been published in leading research journals like Bioinformatics, Neuroradiology, and Neuroimage. He has received 5 best paper awards. The major part of his education was at Munich's universities (LMU and TUM). He had also an associate professor position at UMIT in Innsbruck and longer research visits at leading research institutions in the USA (AT&T research, Carnegie Mellon University, Florida State University) and Singapore (National University of Singapore).