Date
Thursday May 23, 2019 from 12:30 PM to 2:00 PMLocation
TU/eOrganizer
Data SciencePrice
FreeBuilding
MetaForum building - MF 11/12About the event
Many TU/e researchers are advancing or using Artificial Intelligence in their research projects. To support cross disciplinary learning and to strengthen the TU/e AI network, we organize a series of (internal) lunch meetings where various researchers talk about their projects.
PROGRAM
Start | Speaker | Title |
12:30 | Johan Lukkien (Dean M&CS) | Introduction |
12:45 | Wouter Duivesteijn, Data Mining (M&CS) | Mining Exceptional Models with the police, a bank, and on social media |
13:05 | Mitko Veta, Medical Image Analysis (BME) | Medical image analysis challenges |
13:25 | Stella Kapodistria, Stochastics (M&CS) | Integrated learning and decision making |
13:45 |
| Wrap up |
ABSTRACTS
Wouter Duivesteijn
Exceptional Model Mining strives to find subgroups of the dataset that display unusual behavior. It is up to us to define what kind of "behavior" we want to look at, and when such behavior becomes "unusual" as opposed to normal behavior. In this talk, we will have a look at several kinds of unusual behavior in several kinds of practical settings: we will look at unusual ways in which a bank makes its decisions about customers, unusual streams of tourists and commuters moving through the Tokyo transportation system as measured via social media, and unusual spatio-temporal behavior of cars driving through the road network. The latter application allowed the police to apprehend an actual criminal, so we can show real-life value to this work.
Mitko Veta
One of the developments that brought about the success of deep learning methods in computer vision was the availability of large, public annotated image datasets. In medical image analysis, this was mirrored in the form of medical image analysis challenges – friendly competitions in which researchers worldwide evaluate their solutions on the same data with the same criteria, in a blinded manner. Organising a challenge constitutes collecting and publicly distributing a dataset that can be used to address a specific clinical task (e.g. assessment of breast tumor proliferation). Medical image analysis challenges have been a significant driver for improving the state of the art for a variety of such tasks in medical imaging. This talk will give an overview of several challenges that I have organised on histopathology image analysis and the achieved results.
Stella Kapodistria
Inspired by the practice of Philips, we considered an illustrative case (maintenance of screens) for which a machine learning approach is used to predict imminent failures. For this case: 1. We tested different approaches to gain trust and interpretability in the model and to simplify the underlying machine learning mechanism. 2. We built a surrogate simple stochastic model, reducing the dimension of the problem. 3. From the intuition built from the surrogate model, we showed how to enhance the machine learning approach accounting for the maintenance decision and its underlying costs. This is joint work with Collin Drent.