Home      Log In      Contacts      FAQs      INSTICC Portal
 

Keynote Lectures

Machine Learning with Limited Size Datasets
M. Verleysen, Machine Learning Group, Université Catholique de Louvain, Belgium

Indoor Localization - Solved, Finally?
Kay Römer, Graz University of Technology, Austria

 

Machine Learning with Limited Size Datasets

M. Verleysen
Machine Learning Group, Université Catholique de Louvain
Belgium
 

Brief Bio
Michel Verleysen is a Professor of Machine Learning at the UCLouvain, Belgium. He has been an invited professor at EPFL (Switzerland), Université d'Evry Val d'Essonne, Université ParisI-Panthéon-Sorbonne and Université Paris Est (France). He is an Honorary Research Director of the Belgian F.N.R.S. (National Fund for Scientific Research), and the Dean of the Louvain School of Engineering. He is editor-in-chief of the Neural Processing Letters journal (Springer), chairman of the annual ESANN conference (European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning), past associate editor of the IEEE Trans. on Neural Networks journal, and member of the editorial board and program committee of several journals and conferences on neural networks and learning. He is author or co-author of more than 250 scientific papers in international journals and books or communications to conferences with reviewing committee. He is co-author of the "Nonlinear Dimensionality Reduction" book (Springer). His research interests include machine learning, feature selection, nonlinear dimensionality reduction, visualization, high-dimensional data analysis, self-organization, time-series forecasting and biomedical signal processing.


Abstract
Big data are now ubiquitous in many domains of science and technology research, but also in many application areas.  Data available in large amounts enable the development of new paradigms for model design, such as deep learning.  However there exist countless application contexts where the limited number of data is a concern.  An obvious example is patient-based data in healthcare: databases measuring the same information in the same settings for more than a few hundreds or thousands of patients are rare.

Most machine learning methods rely in some way to the approximation of a distribution of data.  While such approximation is reasonable when many data are available in a small-dimensional space (small p, large n), it is not in other small data, large-dimensional space contexts (large p, small n); this is the “curse of dimensionality”.  Machine learning algorithms may fail in these situations.

This talk will introduce some areas of machine learning that are useful to answer these questions.  It will cover fundamental aspects of feature selection, dimensionality reduction, missing data imputation and introduce challenges related to combining data from different sources.  Feature selection and dimensionality reduction can lower the dimensionality of the data space, hence enhancing the performances of machine learning methods; missing data imputation and combining data from different sources are seen as ways to take the most of existing data.  The talk will be accessible to participants with minimal knowledge of machine learning.

 



 

 

Indoor Localization - Solved, Finally?

Kay Römer
Graz University of Technology
Austria
 

Brief Bio
Kay Römer is professor at and director of the Institute for Technical Informatics and head of the Field of Expertise "Information, Communication & Computing" at TU Graz. He obtained his doctorate in computer science from ETH Zurich in 2005 with a thesis on wireless sensor networks. Kay Römer is an internationally recognized expert on networked embedded systems, with research focus on wireless networking, fundamental services, operating systems, programming models, dependability, testbeds, and deployment methodology. He has co-chaired the program committees of leading conferences in the field such as SenSys or IPSN, he is also chairing the steering committee of the EWSN conference series. He is coordinator of the TU Graz Research Center "Dependable Internet of Things" and leads the research area "Cognitive Products" in the research center Pro2Future - Products and Production of the Future.


Abstract
Accurate and reliable indoor localization of smart objects is a key services in many applications domains of the Internet of Things such as smart homes, smart factories, or smart healthcare. While the related problem of outdoor localization has been (mostly) solved by global navigation satellite systems (GNSS) such as GPS, GLONASS, or GALILEIO - indoor localization has been an active research topic over several decades without finding an ultimate solution that is as mature as GNSS.

A very promising technology in this regard are Ultra-Wide-Band (UWB) radio transceivers. While UWB has a long research history, only recently low-power and low-cost UWB transceivers, for example ones produced by DecaWave, have appeared on the market and are being included in the latest smartphone generation, such that UWB will likely become a ubiquitous technology.

Due to the ultra-wide bandwidth, UWB radios transmit very short pulses that allow accurate measurement of time-of-flight even in multi-path environments, which in turn allows for distance measurements with an accuracy of a few centimeters. While UWB radios thereby provide a very promising technology foundation, the indoor localization problem isn't automatically solved by the availability of UWB.

Instead, serveral important research questions have to be addressed to arrive at a global indoor localization system. Firstly, there is the challenge to minimize the infrastructure required to cover large indoor areas. With GPS some tens of satellites are sufficient to cover the globe, but due to the limited communication range of UWB, this is not possible with indoor localization. Secondly, how can we support localization of an arbitrarily large number of densely deployed tags with high update rates? Traditional UWB-based distance measurement requires pair-wise sequential measurements between each tag and each reference station, which does not scale. Thridly, how can robust and accurate localization be achieved in typical indoor enviroments with many obstacles, where line-of-sight between reference and tag is obstructured or even blocked? In this keynote, we present present latest results we obtained in the "Dependable Internet of Things" research center at TU Graz towards addressing these challenges.



footer